<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: JANCTION Render</title>
    <description>The latest articles on DEV Community by JANCTION Render (@_903a3934b362f0afe2a8).</description>
    <link>https://dev.to/_903a3934b362f0afe2a8</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3202149%2Feb704b46-d350-4ee9-ad89-c8596a49e46e.png</url>
      <title>DEV Community: JANCTION Render</title>
      <link>https://dev.to/_903a3934b362f0afe2a8</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/_903a3934b362f0afe2a8"/>
    <language>en</language>
    <item>
      <title>Blender render farm GPUs and VRAM in 2026: who gives a Cycles job 96 GB?</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Mon, 16 Nov 2026 11:01:50 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/blender-render-farm-gpus-and-vram-in-2026-who-gives-a-cycles-job-96-gb-1bdl</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/blender-render-farm-gpus-and-vram-in-2026-who-gives-a-cycles-job-96-gb-1bdl</guid>
      <description>&lt;p&gt;Disclosure: I work on JANCTION Render, one of the services below.&lt;/p&gt;

&lt;p&gt;Earlier today I compared how long render farms keep your renders. The other question I keep getting is simpler: which GPU will my scene actually run on, and will it fit? So I went through the official pricing, spec and FAQ pages again, this time writing down the GPU model and the memory per card.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fv0gd5026gj12xxnnddr7.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fv0gd5026gj12xxnnddr7.jpg" alt="GPU memory a Blender job gets: JANCTION Render 96 GB, GarageFarm 96 GB on Medium and High, BlendSwap and RenderStreet 48 GB, Renderjuice, Ranch Computing and iRender 32 GB" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The numbers (official pages, checked 2026-11-16)
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Service&lt;/th&gt;
&lt;th&gt;GPU for Cycles&lt;/th&gt;
&lt;th&gt;GPU memory per card&lt;/th&gt;
&lt;th&gt;Lowest published price unit&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;JANCTION Render&lt;/td&gt;
&lt;td&gt;NVIDIA RTX PRO 6000 Blackwell, every job&lt;/td&gt;
&lt;td&gt;96 GB&lt;/td&gt;
&lt;td&gt;0.1 JPY per GPU-second (360 JPY per GPU-hour)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GarageFarm.NET&lt;/td&gt;
&lt;td&gt;&lt;a href="https://garagefarm.net/pricing" rel="noopener noreferrer"&gt;RTX 6000 Pro Blackwell 96GB; also RTX 4000 Ada, L40s, RTX A5000&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;96 GB on Medium and High, up to 48 GB on Low&lt;/td&gt;
&lt;td&gt;from $0.004 per OB-hour&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;BlendSwap Render Farm&lt;/td&gt;
&lt;td&gt;&lt;a href="https://www.blendswap.com/render-farm" rel="noopener noreferrer"&gt;NVIDIA L40S&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;48 GB per worker&lt;/td&gt;
&lt;td&gt;$4.30 per server-hour&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;RenderStreet&lt;/td&gt;
&lt;td&gt;&lt;a href="https://render.st/features/" rel="noopener noreferrer"&gt;NVIDIA L40s&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;48 GB per server&lt;/td&gt;
&lt;td&gt;$4.49 per server-hour&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Renderjuice&lt;/td&gt;
&lt;td&gt;&lt;a href="https://www.renderjuice.com/pricing" rel="noopener noreferrer"&gt;primarily RTX 4090s and 5090s&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;"up to 48 GB" (24 GB / 32 GB per card)&lt;/td&gt;
&lt;td&gt;from $49 a month for 2,000 GPU minutes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Ranch Computing&lt;/td&gt;
&lt;td&gt;&lt;a href="https://www.ranchcomputing.com/en/documentation/farm-choice/" rel="noopener noreferrer"&gt;RTX 2080 Ti, 3090, 4090, 5090&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;11 to 32 GB&lt;/td&gt;
&lt;td&gt;from 0.005 EUR per OB-hour&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;iRender (machine rental)&lt;/td&gt;
&lt;td&gt;&lt;a href="https://irendering.net/blender-render-farm-service/" rel="noopener noreferrer"&gt;RTX 4090 or 5090 servers&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;24 or 32 GB; larger scenes spill to system RAM&lt;/td&gt;
&lt;td&gt;from $8.20 per hour&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;RebusFarm&lt;/td&gt;
&lt;td&gt;&lt;a href="https://rebusfarm.net/" rel="noopener noreferrer"&gt;Quadro RTX 6000&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;not stated&lt;/td&gt;
&lt;td&gt;0.45 cent per OB-hour&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;SheepIt&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://www.sheepit-renderfarm.com/faq" rel="noopener noreferrer"&gt;volunteers' GPUs&lt;/a&gt;, reference RTX 3060&lt;/td&gt;
&lt;td&gt;depends on the volunteer&lt;/td&gt;
&lt;td&gt;free&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;OB-hour is a benchmark-based unit the farms use, so those prices do not convert directly to GPU-hours. Services whose official pages do not name the GPU are left out. Specs change, so check the link before relying on a number.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why memory, not just speed
&lt;/h2&gt;

&lt;p&gt;Cycles puts the scene's textures, geometry and BVH on the GPU, and a frame renders on one card. If the scene needs 40 GB and the card has 24, it either spills to much slower system memory or stops with an out-of-memory error. The scenes that hit this first are exactly the ones you would send to a farm: production files with packed 8K textures, photogrammetry or Gaussian-splat scans with millions of points, dense CAD and terrain.&lt;/p&gt;

&lt;p&gt;So the useful question for a heavy scene is not "how fast is the card" but "does the card hold the scene". Most farms answer 24 to 48 GB. GarageFarm offers 96 GB on its upper tiers. On JANCTION Render every job, preview or final, runs on the 96 GB card; there is no smaller tier.&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest alternative: rent the card yourself
&lt;/h2&gt;

&lt;p&gt;GPU clouds rent the same class of card. On their pricing pages today, &lt;a href="https://www.runpod.io/pricing" rel="noopener noreferrer"&gt;RunPod&lt;/a&gt; lists the RTX PRO 6000 96GB from $1.69 an hour (Community) or $2.69 (Secure), and &lt;a href="https://modal.com/pricing" rel="noopener noreferrer"&gt;Modal&lt;/a&gt; at $0.000842 a second, about $3.03 an hour. Per hour that can beat a render service. What you pay for instead is your own time: installing Blender and add-ons, splitting the frames, and the minutes the machine spends starting and idling.&lt;/p&gt;

&lt;p&gt;JANCTION Render charges only the GPU-seconds a render uses (about $2.4 an hour of GPU), an AI agent can call it over MCP, and monthly plans keep renders and scenes in storage while you are subscribed.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;Add the MCP endpoint &lt;code&gt;https://render.janction.jp/mcp&lt;/code&gt; to Claude, ChatGPT, Cursor or Codex and ask for a preview of a heavy scene, or drop a .blend at &lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;https://render.janction.jp&lt;/a&gt;. The comparison lives at &lt;a href="https://render.janction.jp/blender-render-farm-gpu-memory" rel="noopener noreferrer"&gt;https://render.janction.jp/blender-render-farm-gpu-memory&lt;/a&gt; and I will update it monthly. If a farm has changed its hardware or I misread a page, say so in the comments and I will fix the table.&lt;/p&gt;

</description>
      <category>blender</category>
      <category>rendering</category>
      <category>gpu</category>
      <category>3d</category>
    </item>
    <item>
      <title>How long do render farms keep your renders? The published rules of 7 farms (2026)</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Mon, 16 Nov 2026 10:07:53 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/how-long-do-render-farms-keep-your-renders-the-published-rules-of-7-farms-2026-5fh2</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/how-long-do-render-farms-keep-your-renders-the-published-rules-of-7-farms-2026-5fh2</guid>
      <description>&lt;p&gt;Disclosure: I work on JANCTION Render, one of the services below.&lt;/p&gt;

&lt;p&gt;Today JANCTION Render started including JANCTION Storage (beta) with every monthly plan, so renders and the scenes behind them stay for as long as you are subscribed. Before saying that anywhere, I wanted to know what normal looks like. So I read the FAQ, terms and docs of the render farms that publish a rule and wrote down how long each one keeps your files.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmyboq9eyddes53fpflyq.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fmyboq9eyddes53fpflyq.jpg" alt="How long render farms keep your renders: Ranch Computing 5 days, RebusFarm 7, RenderStreet 14, Fox Renderfarm 20, GarageFarm 28, Renderjuice 30" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The numbers (official pages, checked 2026-11-16)
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Service&lt;/th&gt;
&lt;th&gt;How long results are kept&lt;/th&gt;
&lt;th&gt;Monthly plan&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Ranch Computing&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://www.ranchcomputing.com/en/terms-conditions" rel="noopener noreferrer"&gt;project and output deleted 5 days after completion&lt;/a&gt;, even if not downloaded&lt;/td&gt;
&lt;td&gt;no, prepaid balance&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;RebusFarm&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://rebusfarm.net/terms-and-conditions" rel="noopener noreferrer"&gt;results deleted 7 days after they are provided&lt;/a&gt;; longer only if agreed before the order&lt;/td&gt;
&lt;td&gt;no, RenderPoints&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;RenderStreet&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://support.render.st/portal/en/kb/articles/file-retention-policy-how-long-do-we-keep-your-files" rel="noopener noreferrer"&gt;14 days on every plan&lt;/a&gt;; Studio can be extended case by case&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Fox Renderfarm&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://www.foxrenderfarm.com/support/general/common-questions-and-precautions/" rel="noopener noreferrer"&gt;source files kept 20 days&lt;/a&gt;; jobs older than 20 days expire and cannot be downloaded&lt;/td&gt;
&lt;td&gt;no, pay as you go&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GarageFarm.NET&lt;/td&gt;
&lt;td&gt;&lt;a href="https://garagefarm.net/faq" rel="noopener noreferrer"&gt;28 days after the job or project, the first 7 guaranteed&lt;/a&gt;&lt;/td&gt;
&lt;td&gt;no&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Renderjuice&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://renderjuice.com/docs/platform/the-platform" rel="noopener noreferrer"&gt;inputs and outputs deleted after 30 days&lt;/a&gt;; custom retention on Enterprise&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;iRender (machine rental)&lt;/td&gt;
&lt;td&gt;
&lt;a href="https://irendering.net/terms-conditions/" rel="noopener noreferrer"&gt;rented machine data kept 3 months (Regular) or 5 months (Prime) after the last session&lt;/a&gt;; complete render projects are not stored after the job&lt;/td&gt;
&lt;td&gt;no, top-up&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;JANCTION Render&lt;/td&gt;
&lt;td&gt;24 hours after last use without a plan (normally 7 days for keys that have paid 500 JPY or more in total). Monthly plans: renders and source scenes kept in JANCTION Storage (beta) while subscribed, and 90 days after the plan ends&lt;/td&gt;
&lt;td&gt;yes, from Basic&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;These are each company's own published rules, linked as the source; I did not use third-party comparison sites. Rules change, so check the link before you rely on a number. Services that publish no retention rule are left out.&lt;/p&gt;

&lt;p&gt;The short version: job-based farms keep your results somewhere between 5 and 30 days. That is fine for a studio that downloads the frames, archives them and moves on.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why it matters more when an AI agent is rendering
&lt;/h2&gt;

&lt;p&gt;An agent iterates. It renders a shot, you look at it on Thursday and ask for warmer light, and it renders again. If the farm deleted the job on Monday, the agent has to rebuild the scene or you have to upload it again, and the version you wanted to compare against is gone.&lt;/p&gt;

&lt;p&gt;What makes "same as last week, warmer light" work is the source scene still sitting next to the result. That is the gap JANCTION Storage fills on a monthly plan:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Every finished render goes in with its source scene; there is nothing to save by hand.&lt;/li&gt;
&lt;li&gt;The final's reply carries a one-time link to open it, and the browser workbench opens it too. Each version lists its results and the scene.&lt;/li&gt;
&lt;li&gt;Basic includes 200 GiB, Standard 1 TiB for 3 people, Pro 5 TiB for 5 people. Every plan uploads .blend files up to 10 GB.&lt;/li&gt;
&lt;li&gt;After cancelling, saved work can still be opened and downloaded for 90 days after the paid period ends.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  When another farm is the better pick
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;You render a finished project once and archive it yourself: retention barely matters, and a farm with a plugin for your pipeline may suit you better.&lt;/li&gt;
&lt;li&gt;You need a whole machine with your own software for weeks: a machine rental such as iRender keeps the machine's data for months.&lt;/li&gt;
&lt;li&gt;You need hundreds of GPUs at once for one deadline: the large studio farms have more nodes than we do.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;p&gt;From an AI agent, add the MCP endpoint &lt;code&gt;https://render.janction.jp/mcp&lt;/code&gt; (Claude, ChatGPT, Cursor, Codex and others) and ask it to render a scene. From a browser, drop a file at &lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;https://render.janction.jp&lt;/a&gt; (no sign-up for the first preview).&lt;/p&gt;

&lt;p&gt;The comparison also lives at &lt;a href="https://render.janction.jp/render-farm-file-retention" rel="noopener noreferrer"&gt;https://render.janction.jp/render-farm-file-retention&lt;/a&gt; and I will update it monthly. If a farm has published a rule I missed, or changed one, tell me in the comments and I will fix the table.&lt;/p&gt;

</description>
      <category>blender</category>
      <category>rendering</category>
      <category>3d</category>
      <category>ai</category>
    </item>
    <item>
      <title>没有显卡，让 AI 帮你渲染 Blender：把 MCP 接到云端 GPU</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Thu, 12 Nov 2026 09:55:56 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/mei-you-xian-qia-rang-ai-bang-ni-xuan-ran-blenderba-mcp-jie-dao-yun-duan-gpu-l10</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/mei-you-xian-qia-rang-ai-bang-ni-xuan-ran-blenderba-mcp-jie-dao-yun-duan-gpu-l10</guid>
      <description>&lt;p&gt;做 3D 的人大概都遇到过这种情况：AI 已经能写出像样的 Blender bpy 脚本，可是自己的笔记本没有独立显卡，一段几十帧的动画用 CPU 要渲染很久。我们在东京做了一个小服务 JANCTION Render（&lt;a href="https://render.janction.jp/zh" rel="noopener noreferrer"&gt;https://render.janction.jp/zh&lt;/a&gt; ），专门解决“AI 写好了场景，但渲染卡在本机”这一步。下面用实际的数字说明它怎么用、要多少钱、哪些情况不适合。&lt;/p&gt;

&lt;h2&gt;
  
  
  它做什么
&lt;/h2&gt;

&lt;p&gt;JANCTION Render 是一个 MCP（Model Context Protocol）服务器。支持 MCP 的 AI 工具连上它之后，可以把 .blend、bpy 脚本或 GLB / FBX / OBJ 等 3D 文件交给云端的 NVIDIA RTX PRO 6000 渲染，拿回预览图、静帧或 MP4。&lt;/p&gt;

&lt;p&gt;和常见的 Blender MCP（例如 blender-mcp）不一样：那类是让 AI 操作你电脑上开着的 Blender，渲染还是在本机；这个只负责渲染，本机不需要 Blender，也不需要显卡。两者可以一起用。&lt;/p&gt;

&lt;h2&gt;
  
  
  实测的时间和费用
&lt;/h2&gt;

&lt;p&gt;以下是我们在生产环境 GPU 上的单次实测（简单场景，GPU 时间包含 Blender 启动和读取场景）：&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;预览 4 帧（720p，16 采样）：约 4 GPU 秒，往返约 8 秒&lt;/li&gt;
&lt;li&gt;24 帧 1280x720、64 采样的 MP4：约 59 GPU 秒，约 5.9 日元&lt;/li&gt;
&lt;li&gt;Blender 官方 Classroom 场景，1080p 每帧约 8.9 GPU 秒&lt;/li&gt;
&lt;li&gt;一个耳机模型的 24 帧 720p 360 度旋转视频：58.3 GPU 秒&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;价格是每 GPU 秒 0.1 日元。每个新密钥有 100 日元的免费 GPU 时间（不用绑卡），保存信用卡（不扣款）再送 400 日元。不需要注册账号。&lt;/p&gt;

&lt;h2&gt;
  
  
  怎么连
&lt;/h2&gt;

&lt;p&gt;支持远程 MCP 的客户端，添加这个地址即可：&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;https://render.janction.jp/mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;支持 OAuth 的客户端会弹出授权页面并自动生成免费密钥；只能填固定请求头的客户端，用 &lt;code&gt;curl -X POST https://render.janction.jp/v1/keys&lt;/code&gt; 拿到密钥后放进 &lt;code&gt;Authorization: Bearer jr_…&lt;/code&gt;。&lt;/p&gt;

&lt;p&gt;几个国内常用工具的配置方法：&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Qwen Code：&lt;code&gt;qwen mcp add --transport http janction-render https://render.janction.jp/mcp&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Kimi Code：在 &lt;code&gt;~/.kimi-code/mcp.json&lt;/code&gt; 的 mcpServers 里加上这个 url，然后 &lt;code&gt;/mcp-config login janction-render&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;DeepSeek Harness：用官方的 &lt;code&gt;@deepseek-ai/dsh-mcp-client&lt;/code&gt;，transport 选 streamable-http，请求头放 Bearer 密钥&lt;/li&gt;
&lt;li&gt;扣子、Dify 这类没有 MCP 的平台：导入 OpenAPI &lt;code&gt;https://render.janction.jp/openapi-actions.json&lt;/code&gt;，认证用 API Key（请求头 &lt;code&gt;X-API-Key&lt;/code&gt;）&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;能执行命令的环境也可以直接用命令行，国内可以走镜像：&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip install -i https://pypi.tuna.tsinghua.edu.cn/simple janction-render
janction-render try
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  连上之后怎么说
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;用 JANCTION Render 做一个在灰色地面上旋转的红色光面立方体，先给我看 4 帧预览。
如果没问题，再渲染 24 帧 1280x720 的 MP4。
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;AI 会先渲染预览，服务会附上点评（太暗、物体被裁掉、没有灯光之类）和修改建议；确认之后 AI 先给出估算，再渲染最终的视频。&lt;/p&gt;

&lt;h2&gt;
  
  
  哪些情况不适合
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;本机有不错的显卡、只渲染一张静帧：本机通常更快。&lt;/li&gt;
&lt;li&gt;需要在 Blender 里边看边改：那是本地 Blender MCP 或 Blender 本身的工作。&lt;/li&gt;
&lt;li&gt;付款目前只支持信用卡（Stripe，日元结算），还不支持支付宝和微信支付。&lt;/li&gt;
&lt;li&gt;服务器在日本。我们在国内多个网络环境下测试过，云服务器和移动、电信的网络都能正常打开，但部分联通家庭宽带可能连接超时。&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;中文说明在 &lt;a href="https://render.janction.jp/zh" rel="noopener noreferrer"&gt;https://render.janction.jp/zh&lt;/a&gt; ，给 AI 读的说明在 &lt;a href="https://render.janction.jp/zh/llms.txt" rel="noopener noreferrer"&gt;https://render.janction.jp/zh/llms.txt&lt;/a&gt; 。有问题欢迎在评论里提。&lt;/p&gt;

</description>
      <category>blender</category>
      <category>mcp</category>
      <category>ai</category>
      <category>3d</category>
    </item>
    <item>
      <title>Rendering Blender 5.3 glass dispersion from an AI agent, with no local GPU</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Wed, 11 Nov 2026 11:58:50 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/rendering-blender-53-glass-dispersion-from-an-ai-agent-with-no-local-gpu-166p</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/rendering-blender-53-glass-dispersion-from-an-ai-agent-with-no-local-gpu-166p</guid>
      <description>&lt;p&gt;&lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;https://render.janction.jp&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Blender 5.3 came out this week, and JANCTION Render runs it on cloud GPUs as of today. I wanted one shot that shows what 5.3 can do: faceted glass that splits light into colour. This post is the log of making it from an AI agent over MCP, including the two mistakes I made on the way, with the measured time and cost.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F68asafqhfw0cjkqhgfhp.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F68asafqhfw0cjkqhgfhp.jpg" alt="Three glass gems rendered with Blender 5.3 Cycles, with rainbow fringes on the facets" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The setup
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;JANCTION Render connected to the agent over MCP: &lt;code&gt;https://render.janction.jp/mcp&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;No Blender and no GPU on my machine. The agent writes a bpy script, the service runs it in Blender 5.3 on its GPUs and sends back images&lt;/li&gt;
&lt;li&gt;Every preview comes back with a short critic verdict (&lt;code&gt;ok&lt;/code&gt; / &lt;code&gt;check&lt;/code&gt; / &lt;code&gt;fix&lt;/code&gt;) and a price for the full render before anything is charged&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Try 1: the glass came out black
&lt;/h2&gt;

&lt;p&gt;The first script set &lt;code&gt;Dispersion&lt;/code&gt; on the Principled BSDF and used a dark room with thin light strips behind the gems. The preview came back in 5 seconds and the gems were almost black, with no colour at all.&lt;/p&gt;

&lt;p&gt;The reason: in Blender 5.3 the dispersion controls of the Principled BSDF sit under Transmission and are named:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;p&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;mat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;node_tree&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;nodes&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"Principled BSDF"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
&lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;inputs&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"Transmission Weight"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;default_value&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;1.0&lt;/span&gt;
&lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;inputs&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"IOR"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;default_value&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;2.4&lt;/span&gt;
&lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;inputs&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"Transmission Dispersion Scale"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;default_value&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;1.0&lt;/span&gt;
&lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;inputs&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"Transmission Dispersion Abbe Number"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;default_value&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;12.0&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;My script looked for an input called &lt;code&gt;Dispersion&lt;/code&gt;, did not find it, and silently skipped it. The dark room did not help either: glass only shows what is around it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try 2: brighter floor, one narrow spot
&lt;/h2&gt;

&lt;p&gt;I made the floor a light grey, put one spotlight behind and above the gems so the pool of light falls towards the camera, and fixed the input names. The next preview showed the rainbow fringes on the facets, and the critic said &lt;code&gt;ok&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Mistake 2: 72 identical frames
&lt;/h2&gt;

&lt;p&gt;I wanted the gems to turn slowly over 72 frames. I rendered 10 frames first, then all 72, and every frame was the same picture. The keyframes were wrong:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="c1"&gt;# wrong: frame_set() re-evaluates the animation and resets the value I just set&lt;/span&gt;
&lt;span class="n"&gt;g&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;rotation_euler&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;new_value&lt;/span&gt;
&lt;span class="n"&gt;sc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;frame_set&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;72&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;g&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;keyframe_insert&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"rotation_euler"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="c1"&gt;# right: give the frame to keyframe_insert&lt;/span&gt;
&lt;span class="n"&gt;g&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;keyframe_insert&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"rotation_euler"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;frame&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="n"&gt;g&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;rotation_euler&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;new_value&lt;/span&gt;
&lt;span class="n"&gt;g&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;keyframe_insert&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"rotation_euler"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;frame&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;72&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Rendering 10 contiguous frames before the full job is what caught it cheaply. After the fix, the 10-frame test showed the gems moving and no black or duplicate frames, so I ran the full job.&lt;/p&gt;

&lt;h2&gt;
  
  
  Numbers (measured on 2026-11-11)
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Job&lt;/th&gt;
&lt;th&gt;Settings&lt;/th&gt;
&lt;th&gt;GPU time&lt;/th&gt;
&lt;th&gt;Wall time&lt;/th&gt;
&lt;th&gt;Charged&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Preview&lt;/td&gt;
&lt;td&gt;1280x720, 32 samples, 1 frame&lt;/td&gt;
&lt;td&gt;2.8 GPU s&lt;/td&gt;
&lt;td&gt;5 s&lt;/td&gt;
&lt;td&gt;1 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;10-frame test&lt;/td&gt;
&lt;td&gt;1920x1080, 128 samples&lt;/td&gt;
&lt;td&gt;59.9 GPU s&lt;/td&gt;
&lt;td&gt;19 s&lt;/td&gt;
&lt;td&gt;6 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Final&lt;/td&gt;
&lt;td&gt;1920x1080, 128 samples, 72 frames&lt;/td&gt;
&lt;td&gt;267 GPU s&lt;/td&gt;
&lt;td&gt;1 min 25 s&lt;/td&gt;
&lt;td&gt;27 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;About 3.7 GPU seconds per 1080p frame. The final ran in parallel chunks, so 267 GPU seconds took under a minute and a half.&lt;/p&gt;

&lt;h2&gt;
  
  
  The scene
&lt;/h2&gt;

&lt;p&gt;The full script is about 90 lines: a light grey floor, five thin emissive strips in the back, one spotlight, three icospheres with the glass material above, and a camera. If you connect JANCTION Render to Claude, Cursor or Codex, you can ask for it in one sentence ("render three glass gems with dispersion in Blender 5.3") and let the agent write its own version.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;MCP: &lt;code&gt;https://render.janction.jp/mcp&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;Browser (send a .blend): &lt;a href="https://render.janction.jp/start" rel="noopener noreferrer"&gt;https://render.janction.jp/start&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;New keys get some free GPU time; after that it is 0.1 JPY per GPU second&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;https://render.janction.jp&lt;/a&gt;&lt;/p&gt;

</description>
      <category>blender</category>
      <category>mcp</category>
      <category>ai</category>
      <category>3d</category>
    </item>
    <item>
      <title>Rendering a Blender animation in drafts from Cursor or Claude Code: 48 frames first, then the full shot without paying for them twice</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Wed, 04 Nov 2026 05:39:02 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/rendering-a-blender-animation-in-drafts-from-cursor-or-claude-code-48-frames-first-then-the-full-4ci9</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/rendering-a-blender-animation-in-drafts-from-cursor-or-claude-code-48-frames-first-then-the-full-4ci9</guid>
      <description>&lt;p&gt;When an AI agent in my editor builds a Blender animation, the expensive part is not the first render. It is the second and third ones: the camera is a little off, the color is wrong, the client wants it longer. I wanted to see what that loop actually costs when the agent renders on a cloud GPU over MCP, so I ran one shot end to end and wrote down every number.&lt;/p&gt;

&lt;p&gt;The scene is a 5-second product reveal (120 frames at 24 fps): a bottle on a turntable, the camera dollying in. Everything below was driven through the JANCTION Render MCP server, the same way it runs from Cursor, Claude Code or OpenCode. Disclosure: I work on JANCTION Render.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyy2xqgtjck8fl3mfznnz.gif" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyy2xqgtjck8fl3mfznnz.gif" alt="The bottle reveal, 120 frames" width="640" height="360"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Connecting the agent
&lt;/h2&gt;

&lt;p&gt;Cursor (&lt;code&gt;.cursor/mcp.json&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"janction-render"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://render.janction.jp/mcp"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Claude Code:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude mcp add &lt;span class="nt"&gt;--transport&lt;/span&gt; http janction-render https://render.janction.jp/mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;OpenCode (&lt;code&gt;opencode.json&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"$schema"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://opencode.ai/config.json"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"mcp"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"janction-render"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"remote"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://render.janction.jp/mcp"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No local Blender and no GPU are needed. The agent writes a bpy script (or sends a .blend) and gets images or an MP4 back.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Preview: four frames in one image, free
&lt;/h2&gt;

&lt;p&gt;The agent's first call was a preview of frames 1, 40, 80 and 120 tiled in one image. It came back in 6 seconds (4.65 GPU-seconds) and cost nothing: previews are free up to 2 GPU-minutes a day.&lt;/p&gt;

&lt;p&gt;The built-in critic said "ok" (exposure, framing, lighting). But at frame 120 the camera had pushed in so far that the bottle's cap was cut off. The critic checks each frame for obvious problems; it does not know that the cap matters to me. So the agent moved the end of the camera move back and previewed again: 5 seconds, 4.73 GPU-seconds, free, cap in frame.&lt;/p&gt;

&lt;p&gt;This is the cheapest place to iterate. Two previews, zero yen, and the shot was fixed before any final frame existed.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. A 48-frame draft on the final settings
&lt;/h2&gt;

&lt;p&gt;Before committing to 120 frames at 1920x1080 and 128 samples (Cycles), the agent asked for an estimate of frames 1-48. The estimate was 454 GPU-seconds, about 46 JPY, and it said so itself: confidence low, because this scene had never been rendered at final quality, so it was going by my earlier renders.&lt;/p&gt;

&lt;p&gt;The draft came back in 65 seconds: 179 GPU-seconds, 18 JPY. The estimate was about 2.5 times too high, which is the safe direction (only the GPU time actually used is charged; the rest of the hold comes back). After this draft, the service had a measured per-frame time for this exact scene, and the next estimate was marked confidence high.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. The full 120 frames, without re-rendering the draft
&lt;/h2&gt;

&lt;p&gt;Then the agent asked for frames 1-120 with the same scene and the same settings. The service noticed that frames 1-48 had already been rendered by this key, for this scene, with identical picture settings, and copied them instead of rendering them again. 4 of the 12 chunks were copies; only frames 49-120 went to a GPU.&lt;/p&gt;

&lt;p&gt;Result: 103 seconds, 272 GPU-seconds, 28 JPY. Rendering all 120 frames from scratch would have been about 440 GPU-seconds at the draft's 3.7 seconds per frame, so the draft was not wasted money.&lt;/p&gt;

&lt;p&gt;The rule is strict on purpose. Copies are only used when the key, the uploaded scene and the picture settings (size, samples, engine, lighting preset) all match. Change any of them and the frames are rendered again.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. A real revision: new color, full re-render
&lt;/h2&gt;

&lt;p&gt;Next revision: green instead of blue. That is a new script, so it is a new scene, and nothing can be reused. The full 120 frames took 139 seconds, 418 GPU-seconds, 42 JPY.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9s2jev7ek0e4o5x5076z.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9s2jev7ek0e4o5x5076z.png" alt="Frame 120, blue version and green version" width="800" height="225"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The numbers
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Step&lt;/th&gt;
&lt;th&gt;Wall time&lt;/th&gt;
&lt;th&gt;GPU-seconds&lt;/th&gt;
&lt;th&gt;Cost&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Preview, 4 frames (v1)&lt;/td&gt;
&lt;td&gt;6 s&lt;/td&gt;
&lt;td&gt;4.65&lt;/td&gt;
&lt;td&gt;free&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Preview, 4 frames (v2, camera fixed)&lt;/td&gt;
&lt;td&gt;5 s&lt;/td&gt;
&lt;td&gt;4.73&lt;/td&gt;
&lt;td&gt;free&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Draft, frames 1-48, 1080p&lt;/td&gt;
&lt;td&gt;65 s&lt;/td&gt;
&lt;td&gt;179&lt;/td&gt;
&lt;td&gt;18 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Full, frames 1-120 (1-48 copied)&lt;/td&gt;
&lt;td&gt;103 s&lt;/td&gt;
&lt;td&gt;272&lt;/td&gt;
&lt;td&gt;28 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Revision, new color, frames 1-120&lt;/td&gt;
&lt;td&gt;139 s&lt;/td&gt;
&lt;td&gt;418&lt;/td&gt;
&lt;td&gt;42 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Total: 88 JPY (about 0.56 USD) for two previews, a draft, the full shot and one full revision. That fits inside the free credit a new key starts with: 100 JPY of GPU time with no card, 400 JPY more after saving a card that is not charged; after that, previews stay free up to 2 GPU-minutes a day and finals cost 0.1 JPY per GPU-second. Current prices are always on &lt;a href="https://render.janction.jp/pricing" rel="noopener noreferrer"&gt;https://render.janction.jp/pricing&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would tell an agent to do
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Iterate on previews. Four tiled frames cost one 720p frame of GPU time and are free within the daily allowance.&lt;/li&gt;
&lt;li&gt;Draft a short contiguous range on the final settings before the full animation. It catches problems that previews do not show and calibrates the estimate for the rest.&lt;/li&gt;
&lt;li&gt;Extend the range with the same scene_id and the same settings. Frames already rendered are copied, not billed again.&lt;/li&gt;
&lt;li&gt;Treat any change to the scene or the settings as a new render, and say the price before starting it. A &lt;code&gt;max_yen&lt;/code&gt; cap on the call refuses anything above it, and an &lt;code&gt;idempotency_key&lt;/code&gt; keeps a retried call from rendering twice.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The scripts for all three versions are short enough to paste into a chat; ask your agent to "build a 5-second product reveal of a bottle on a turntable with a camera dolly, preview frames 1, 40, 80 and 120" and it will write something close.&lt;/p&gt;

</description>
      <category>blender</category>
      <category>mcp</category>
      <category>ai</category>
      <category>3d</category>
    </item>
    <item>
      <title>Claude Code wrote a Blender scene from one sentence, saw its own mistake in the preview, and fixed it before rendering</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Mon, 02 Nov 2026 11:19:26 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/claude-code-wrote-a-blender-scene-from-one-sentence-saw-its-own-mistake-in-the-preview-and-fixed-155m</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/claude-code-wrote-a-blender-scene-from-one-sentence-saw-its-own-mistake-in-the-preview-and-fixed-155m</guid>
      <description>&lt;p&gt;I work on JANCTION Render at JasmyLab, and this run was done by Claude Code itself: it wrote the bpy script, called the MCP tools, looked at the images and changed the script. I only typed the sentence. Numbers are from the production service on 2026-11-02.&lt;/p&gt;

&lt;p&gt;The sentence:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Three stacked books and a pencil on a wooden desk, soft window light from the left, the camera pushes in slowly over 48 frames.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Site: &lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;https://render.janction.jp&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Setup
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;claude mcp add --transport http janction-render https://render.janction.jp/mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That is the remote server; the first tool call opens an OAuth page that creates a free key. For this run I used the stdio package instead, because it sends local files and saves results locally:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;claude mcp add janction-render -e JANCTION_RENDER_SERVER=https://render.janction.jp -- uvx --from janction-render janction-render-mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Either way Claude Code sees 13 tools. The ones that matter here are &lt;code&gt;render_preview&lt;/code&gt;, &lt;code&gt;render_estimate&lt;/code&gt; and &lt;code&gt;render_final&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Round 1: the script, and a preview that looked wrong
&lt;/h2&gt;

&lt;p&gt;Claude Code wrote about 80 lines of bpy: a desk, three cubes as books with lighter blocks for pages, a six-sided cylinder as a pencil, two area lights, a camera keyed from frame 1 to 48. Then it called &lt;code&gt;render_preview&lt;/code&gt; with &lt;code&gt;frames="1,16,32,48"&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fp1mvvxttpjsfqxgxnltw.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fp1mvvxttpjsfqxgxnltw.jpg" alt="First preview: four frames tiled, desk blown out, black void behind it" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;3.46 GPU-seconds, back in 5 seconds, free. The built-in critic said &lt;code&gt;ok&lt;/code&gt; with a score of 100. Claude Code disagreed with it, which is the point of this post: the desk was blown out by a 250 W area light, and the desk was only 1.6 m wide, so the camera saw a black void behind it (the scene had no world).&lt;/p&gt;

&lt;h2&gt;
  
  
  Round 2: three edits, one more free preview
&lt;/h2&gt;

&lt;p&gt;The changes Claude Code made to its own script:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;window light 250 W → 90 W, fill 40 W → 15 W&lt;/li&gt;
&lt;li&gt;desk 1.6 × 1.0 m → 4.0 × 3.0 m so no edge is in the frame&lt;/li&gt;
&lt;li&gt;a world background, warm grey at strength 0.6, instead of nothing&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F757kahrv3xz2a0arll2i.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F757kahrv3xz2a0arll2i.jpg" alt="Second preview: even light, no void" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Another 3.46 GPU-seconds, free. Previews stay free up to 2 GPU-minutes a day, so looking twice costs nothing.&lt;/p&gt;

&lt;h2&gt;
  
  
  Round 3: a 12-frame test, then the estimate
&lt;/h2&gt;

&lt;p&gt;The estimate before any final said 271 GPU-seconds for 12 frames with &lt;code&gt;confidence: low&lt;/code&gt; and asked for a short contiguous test first. The key's recent finals were a heavy 50 MB scene, so the guess was far too high this time. The test:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;render_final(scene_id=..., frame_start=1, frame_end=12, width=1920, height=1080, samples=128, output="png", max_yen=60)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;46.96 GPU-seconds for 12 frames (3.9 GPU-seconds a frame at 1080p, 128 samples, denoised), 24 seconds of wall time on 4 workers, 12 of 12 frames present, charged 5 JPY.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;render_estimate&lt;/code&gt; for frames 1–48 after the test: 169 GPU-seconds, about 40 seconds of wall time, about 17 JPY, &lt;code&gt;basis: scene&lt;/code&gt;, &lt;code&gt;confidence: high&lt;/code&gt;. That is the number to show a user before the full render.&lt;/p&gt;

&lt;h2&gt;
  
  
  Round 4: the final
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;render_final(scene_id=..., frame_start=1, frame_end=48, width=1920, height=1080, samples=128, output="mp4", fps=24, max_yen=40)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;186.65 GPU-seconds, 64 seconds of wall time, 48 of 48 frames, 0 black frames (checked with ffmpeg), MP4 at 1920x1080 and 24 fps. Charged 19 JPY.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Folpie7s23218mu5dzotv.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Folpie7s23218mu5dzotv.jpg" alt="Frame 40 of the final render" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The whole run in numbers
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Step&lt;/th&gt;
&lt;th&gt;GPU-seconds&lt;/th&gt;
&lt;th&gt;Wall time&lt;/th&gt;
&lt;th&gt;Cost&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Preview 1 (4 frames)&lt;/td&gt;
&lt;td&gt;3.46&lt;/td&gt;
&lt;td&gt;5 s&lt;/td&gt;
&lt;td&gt;free&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Preview 2 (4 frames)&lt;/td&gt;
&lt;td&gt;3.46&lt;/td&gt;
&lt;td&gt;5 s&lt;/td&gt;
&lt;td&gt;free&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Test, frames 1–12 at 1080p&lt;/td&gt;
&lt;td&gt;46.96&lt;/td&gt;
&lt;td&gt;24 s&lt;/td&gt;
&lt;td&gt;5 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Final, frames 1–48 at 1080p, MP4&lt;/td&gt;
&lt;td&gt;186.65&lt;/td&gt;
&lt;td&gt;64 s&lt;/td&gt;
&lt;td&gt;19 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;From one sentence to a 48-frame 1080p Cycles clip: about 2 minutes of waiting and 24 JPY, with the agent catching its own lighting mistake on the way. A new key starts with 100 JPY of GPU time with no card (14 days) and 400 JPY more after saving a card that is not charged; this whole run fits inside the first 100 JPY. Current prices: &lt;a href="https://render.janction.jp/pricing" rel="noopener noreferrer"&gt;https://render.janction.jp/pricing&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would tell other people running this
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Make the agent look at the preview image itself. The critic catches framing and exposure problems, but "ok" is not the same as "what you asked for".&lt;/li&gt;
&lt;li&gt;Run a short contiguous test before trusting the estimate. Both runs I did today had a first estimate that was off by about 6x in opposite directions, and both were right after the test.&lt;/li&gt;
&lt;li&gt;Pass &lt;code&gt;max_yen&lt;/code&gt;. The service holds at most that amount and charges only the GPU time used.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The script Claude Code wrote is short enough to paste; ask it for the same sentence and compare. Docs for agents: &lt;a href="https://render.janction.jp/llms.txt" rel="noopener noreferrer"&gt;https://render.janction.jp/llms.txt&lt;/a&gt; · Source (MCP server, CLI, client): &lt;a href="https://github.com/JasmyLab-JANCTION/janction-render" rel="noopener noreferrer"&gt;https://github.com/JasmyLab-JANCTION/janction-render&lt;/a&gt;&lt;/p&gt;

</description>
      <category>blender</category>
      <category>mcp</category>
      <category>claude</category>
      <category>ai</category>
    </item>
    <item>
      <title>Rendering a 50 MB .blend at 1080p from Cursor over MCP: every tool call, measured</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Mon, 02 Nov 2026 11:08:24 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/rendering-a-50-mb-blend-at-1080p-from-cursor-over-mcp-every-tool-call-measured-3m6n</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/rendering-a-50-mb-blend-at-1080p-from-cursor-over-mcp-every-tool-call-measured-3m6n</guid>
      <description>&lt;p&gt;I work on JANCTION Render at JasmyLab, so read this as the operator's own run. The numbers are from the production service on 2026-11-02 and nothing is rounded up.&lt;/p&gt;

&lt;p&gt;The question I wanted to answer: when an agent in Cursor (or Claude Code, or Codex) renders a real Cycles scene through an MCP server, what does it actually get back at each step, how long does it take, and what does it cost? Not a cube. A 49.7 MB .blend with 386 meshes, 46 packed textures, three lights and a 200-frame animation at 30 fps.&lt;/p&gt;

&lt;p&gt;Site: &lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;https://render.janction.jp&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Setup in Cursor
&lt;/h2&gt;

&lt;p&gt;One click on the "Add to Cursor" button at &lt;a href="https://render.janction.jp/cursor-blender" rel="noopener noreferrer"&gt;https://render.janction.jp/cursor-blender&lt;/a&gt;, or by hand in &lt;code&gt;~/.cursor/mcp.json&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"janction-render"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://render.janction.jp/mcp"&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Settings → Tools &amp;amp; MCP lists &lt;code&gt;janction-render&lt;/code&gt; with 13 tools. The first tool call opens an OAuth page that creates a free key. There is no sign-up form and no card.&lt;/p&gt;

&lt;p&gt;One honest note on the run below: I made the calls through the same server from a terminal agent (the &lt;code&gt;janction-render&lt;/code&gt; stdio package, &lt;code&gt;pip install janction-render&lt;/code&gt;), because the stdio package can send a local .blend directly. Cursor's remote MCP takes an https URL or a bpy script; to send a local file from Cursor, drop it at &lt;a href="https://render.janction.jp/upload" rel="noopener noreferrer"&gt;https://render.janction.jp/upload&lt;/a&gt; for a 12-hour link, or let the agent run the CLI. The tools and the responses are the same.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 1: read the scene (no GPU)
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;scene_info(scene_path="toyroom.blend")
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Back in a few seconds: Blender 5.2.2, Cycles, frames 1–200 at 30 fps, 1920x1080, 160 samples in the file, active camera &lt;code&gt;Camera.002&lt;/code&gt;, 386 meshes, 53 curves, 3 lights, 69 materials, 46 of 48 images packed, &lt;code&gt;missing_files: []&lt;/code&gt;, and a &lt;code&gt;scene_id&lt;/code&gt; to reuse so the 50 MB is uploaded once.&lt;/p&gt;

&lt;p&gt;The agent now knows the frame range and fps without opening Blender, which matters for the final call later.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: a free preview of four frames
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;render_preview(scene_id=..., frames="1,67,134,200")
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F16y0rdkeycs8ene0le60.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F16y0rdkeycs8ene0le60.jpg" alt="Four preview frames tiled in one image" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;7.67 GPU-seconds, finished in 10 seconds of wall time, cost: free. The verdict from the built-in critic was &lt;code&gt;ok&lt;/code&gt; (score 100): "no problems found in exposure, framing or lighting; looks ready for a final render". The response also carried a &lt;code&gt;final_estimate&lt;/code&gt; for the whole animation: about 3,224 GPU-seconds for 200 frames at 1080p and 128 samples.&lt;/p&gt;

&lt;p&gt;Previews are the step that makes the MCP loop work. The agent sees the image, reads the verdict, fixes the scene if needed, and only then spends money.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: the estimate, and where it was wrong
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;render_estimate(scene_id=..., frame_start=1, frame_end=24, width=1920, height=1080, samples=128)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It answered 72 GPU-seconds, about 8 JPY, with &lt;code&gt;confidence: low&lt;/code&gt; and a warning: this scene has no completed final-path measurement; run a short contiguous final test, then estimate again.&lt;/p&gt;

&lt;p&gt;That warning was right. The 72 seconds came from my key's recent finals (small scenes), not from this scene. The preview-based &lt;code&gt;final_estimate&lt;/code&gt; from step 2 (3,224 GPU-seconds for 200 frames, about 16 GPU-seconds a frame) was much closer to what happened next. I have filed that as a bug: when a scene has no final yet, the preview measurement should win over the key's history.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 4: 24 contiguous frames on the final path
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;render_final(scene_id=..., frame_start=1, frame_end=24, width=1920, height=1080,
             samples=128, output="mp4", fps=30, max_yen=30)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Measured:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;446.6 GPU-seconds for 24 frames, so 18.6 GPU-seconds a frame at 1080p and 128 samples with denoising&lt;/li&gt;
&lt;li&gt;135 seconds of wall time, because the job was split into 4 chunks across 4 workers&lt;/li&gt;
&lt;li&gt;24 of 24 frames present, 0 black frames (checked with ffmpeg), MP4 at 1920x1080 and 30 fps&lt;/li&gt;
&lt;li&gt;list price at 0.1 JPY per GPU-second: about 45 JPY&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frxivow8r0te24ovpg1t1.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Frxivow8r0te24ovpg1t1.jpg" alt="Frame 12 of the final render" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;code&gt;max_yen&lt;/code&gt; is the ceiling the agent passes so a wrong estimate cannot turn into a surprise. The service holds that much and charges only the GPU time used.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 5: estimate again, now with a measurement
&lt;/h2&gt;

&lt;p&gt;The same &lt;code&gt;render_estimate&lt;/code&gt; for frames 1–200 after the test: 4,231 GPU-seconds, about 20 minutes of wall time on 4 workers, about 424 JPY, &lt;code&gt;basis: scene&lt;/code&gt;, &lt;code&gt;confidence: high&lt;/code&gt;, &lt;code&gt;calibration.final_path_measured: true&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;That is the number the agent should show the user before the full render. In this run I stopped here; the 200-frame version of this scene is already on the site as an example.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it costs, in plain terms
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Reading the scene and the 4-frame preview: 0 JPY (previews stay free up to 2 GPU-minutes a day after the welcome credit)&lt;/li&gt;
&lt;li&gt;24 frames at 1080p: 446.6 GPU-seconds, about 45 JPY at list price&lt;/li&gt;
&lt;li&gt;The full 200 frames: about 424 JPY, estimated from the test&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A new key starts with 100 JPY of GPU time with no card (14 days), and 400 JPY more after saving a card that is not charged. After that it is 0.1 JPY per GPU-second from prepaid credit. Current numbers: &lt;a href="https://render.janction.jp/pricing" rel="noopener noreferrer"&gt;https://render.janction.jp/pricing&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What to tell your agent
&lt;/h2&gt;

&lt;p&gt;The order that worked: &lt;code&gt;scene_info&lt;/code&gt; → &lt;code&gt;render_preview&lt;/code&gt; → &lt;code&gt;render_final&lt;/code&gt; for a short contiguous test → &lt;code&gt;render_estimate&lt;/code&gt; → ask the user → &lt;code&gt;render_final&lt;/code&gt; for the whole range. Skipping the test gives you an estimate that can be off by several times, as step 3 shows.&lt;/p&gt;

&lt;p&gt;The MCP server, CLI and client are open source: &lt;a href="https://github.com/JasmyLab-JANCTION/janction-render" rel="noopener noreferrer"&gt;https://github.com/JasmyLab-JANCTION/janction-render&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Docs for agents: &lt;a href="https://render.janction.jp/llms.txt" rel="noopener noreferrer"&gt;https://render.janction.jp/llms.txt&lt;/a&gt;&lt;/p&gt;

</description>
      <category>blender</category>
      <category>cursor</category>
      <category>mcp</category>
      <category>rendering</category>
    </item>
    <item>
      <title>White-background product shots for a whole catalog from GLB files: 9 products, 36 images in 44 seconds (measured)</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Wed, 21 Oct 2026 15:20:20 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/white-background-product-shots-for-a-whole-catalog-from-glb-files-9-products-36-images-in-44-887</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/white-background-product-shots-for-a-whole-catalog-from-glb-files-9-products-36-images-in-44-887</guid>
      <description>&lt;p&gt;If you sell physical products and already have 3D models (GLB, glTF, FBX, USD), you can turn every model into a set of product shots on a plain white background without opening Blender. This is how I did it with &lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;JANCTION Render&lt;/a&gt;, and what it cost when I measured it on 2026-10-22.&lt;/p&gt;

&lt;p&gt;JANCTION Render is a cloud GPU render farm for Blender that AI agents call as an MCP server or HTTP API. Send a .blend file, a bpy script or a 3D file (glTF / FBX / USD) and get frames or an MP4 back. No local Blender or GPU is needed.&lt;/p&gt;

&lt;h2&gt;
  
  
  One request for the whole catalog
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;POST /v1/batches&lt;/code&gt; takes up to 100 items. Each item points at a model; &lt;code&gt;defaults&lt;/code&gt; applies to all of them:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://render.janction.jp/v1/batches &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"X-API-Key: &lt;/span&gt;&lt;span class="nv"&gt;$JANCTION_RENDER_API_KEY&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{
    "defaults": {"kind": "final", "preset": "product", "background": "white"},
    "items": [
      {"ref": "bottle",  "scene_url": "https://example.com/models/bottle.glb"},
      {"ref": "speaker", "scene_url": "https://example.com/models/speaker.glb"},
      {"ref": "chair",   "scene_url": "https://example.com/models/chair.glb"}
    ]
  }'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;preset: "product"&lt;/code&gt; renders 4 shots around the object (1440x1440, studio lighting, the object framed to its size).&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;background: "white"&lt;/code&gt; makes only what the camera sees directly white. The lighting and the reflections on the product still come from the studio light, so metal and glass do not go flat. &lt;code&gt;"transparent"&lt;/code&gt; gives PNGs with alpha instead, and the default stays a neutral grey.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;GET /v1/batches/{batch_id}&lt;/code&gt; returns the state of every item and its job id; each job lists its PNGs.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What I measured
&lt;/h2&gt;

&lt;p&gt;I used 9 models from the Khronos glTF Sample Assets that are released as CC0: WaterBottle, BoomBox, Avocado, Lantern, SheenChair, ToyCar, Corset, BarramundiFish and GlassVaseFlowers. One batch request, a new key, the live service.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;result&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;images&lt;/td&gt;
&lt;td&gt;36 (9 products x 4 angles), 1440x1440 PNG, 128 samples&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;time from the request to the last image&lt;/td&gt;
&lt;td&gt;44 seconds&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPU time&lt;/td&gt;
&lt;td&gt;118.2 GPU seconds in total (12.7 to 14.1 per product)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;cost&lt;/td&gt;
&lt;td&gt;about 12 JPY at 0.1 JPY per GPU-second; with a new key it came out of the welcome credit&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;failures&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;background&lt;/td&gt;
&lt;td&gt;the corner pixels are pure white (255, 255, 255)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The same 9 products on the default grey background took 47 seconds and 136.1 GPU seconds the day before, so the white background does not cost more.&lt;/p&gt;

&lt;h2&gt;
  
  
  Things to know before you run your own catalog
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Glass on white has little contrast. For clear products, &lt;code&gt;background: "transparent"&lt;/code&gt; and your own backdrop may look better.&lt;/li&gt;
&lt;li&gt;The camera frames the whole model from each angle, so a long, thin object seen end-on comes out small in that shot (the fish in my test did in its first angle).&lt;/li&gt;
&lt;li&gt;A new key gets a welcome credit (100 JPY right away, 400 JPY more after saving a card that is not charged, valid 14 days). After that, renders use prepaid credit at 0.1 JPY per GPU-second.&lt;/li&gt;
&lt;li&gt;The farm has a small number of GPUs. When they are busy, items wait in the queue; the batch keeps going and reports each item when it finishes.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  From an AI agent
&lt;/h2&gt;

&lt;p&gt;The same thing works from Claude, ChatGPT or any MCP client connected to &lt;code&gt;https://render.janction.jp/mcp&lt;/code&gt;: ask for product shots of a GLB and the agent calls &lt;code&gt;render_final&lt;/code&gt; with &lt;code&gt;preset="product"&lt;/code&gt;. The background option is on the HTTP API for now.&lt;/p&gt;

&lt;p&gt;Current prices and limits: &lt;a href="https://render.janction.jp/pricing" rel="noopener noreferrer"&gt;https://render.janction.jp/pricing&lt;/a&gt; and &lt;a href="https://render.janction.jp/facts.json" rel="noopener noreferrer"&gt;https://render.janction.jp/facts.json&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>3d</category>
      <category>ecommerce</category>
      <category>blender</category>
      <category>api</category>
    </item>
    <item>
      <title>A Blender MCP server that tells your agent the price before it renders (real estimates, 48 to 1,800 frames)</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Wed, 21 Oct 2026 09:36:04 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/a-blender-mcp-server-that-tells-your-agent-the-price-before-it-renders-real-estimates-48-to-1800-4oac</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/a-blender-mcp-server-that-tells-your-agent-the-price-before-it-renders-real-estimates-48-to-1800-4oac</guid>
      <description>&lt;p&gt;&lt;em&gt;Update (2026-10-22): since the evening of 10/21, a key with no prepaid credit is no longer stopped for a small hold. When the estimate fits the free time left and only the 1.5x ceiling would not, the job runs free if that hold is 50 JPY or less, and the service covers any overrun. The 240-frame example below now runs with nothing held. Estimates also follow each key's own measured renders: on 10/22, another key that had rendered finals before got 488 GPU seconds ("about 2 minutes") for 240 frames and 3,661 GPU seconds ("about 15 minutes", 88 JPY held) for 1,800 frames.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;If you let an AI agent render Blender scenes on a cloud GPU, the question you want answered first is not "how fast" but "what will this cost me, and can it run away". This is how &lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;JANCTION Render&lt;/a&gt; answers it, with estimates I took from the live service on 2026-10-21.&lt;/p&gt;

&lt;p&gt;JANCTION Render is a cloud GPU render farm for Blender that AI agents call as an MCP server or HTTP API. Send a .blend file, a bpy script or a 3D file (glTF / FBX / USD) and get frames or an MP4 back. No local Blender or GPU is needed.&lt;/p&gt;

&lt;h2&gt;
  
  
  The order the agent follows
&lt;/h2&gt;

&lt;p&gt;The MCP server's own instructions tell the agent to work in this order:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;code&gt;render_preview&lt;/code&gt; - a few frames in seconds, to check the scene&lt;/li&gt;
&lt;li&gt;ask the user&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;render_estimate&lt;/code&gt; - when the job is long or the free time left is uncertain (no GPU used, nothing charged)&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;render_final&lt;/code&gt; - with &lt;code&gt;max_yen&lt;/code&gt; as a ceiling if you want one&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;code&gt;render_estimate&lt;/code&gt; takes the same arguments as the final render and returns time and money in one answer.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the estimates look like
&lt;/h2&gt;

&lt;p&gt;Same scene for all three (the public example &lt;code&gt;ex12_logo_ring_spin.py&lt;/code&gt;), a new key with its welcome credit:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;request&lt;/th&gt;
&lt;th&gt;GPU seconds&lt;/th&gt;
&lt;th&gt;expected wait&lt;/th&gt;
&lt;th&gt;free time covers&lt;/th&gt;
&lt;th&gt;held from credit (ceiling)&lt;/th&gt;
&lt;th&gt;charged if it runs as estimated&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;48 frames, 1280x720&lt;/td&gt;
&lt;td&gt;86&lt;/td&gt;
&lt;td&gt;"about 20 seconds"&lt;/td&gt;
&lt;td&gt;all of it&lt;/td&gt;
&lt;td&gt;0 JPY&lt;/td&gt;
&lt;td&gt;0 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;240 frames, 1920x1080&lt;/td&gt;
&lt;td&gt;908&lt;/td&gt;
&lt;td&gt;"about 4 minutes"&lt;/td&gt;
&lt;td&gt;all of it&lt;/td&gt;
&lt;td&gt;37 JPY&lt;/td&gt;
&lt;td&gt;0 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;1,800 frames, 1920x1080&lt;/td&gt;
&lt;td&gt;6,813&lt;/td&gt;
&lt;td&gt;"about 30 minutes"&lt;/td&gt;
&lt;td&gt;1,000 seconds&lt;/td&gt;
&lt;td&gt;872 JPY&lt;/td&gt;
&lt;td&gt;582 JPY&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;A few things in that table matter for an agent:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The hold is larger than the estimate. A render can take up to 1.5 times its estimate, so the service reserves that much from prepaid credit and only charges the time actually used. On the 240-frame job, the estimate fits the free time, but the 1.5x ceiling would not, so 37 JPY is held and 0 is expected to be charged.&lt;/li&gt;
&lt;li&gt;The answer includes any per-job and per-day spending limits set on the key (there are none by default; a key can set its own), so the agent knows before it asks.&lt;/li&gt;
&lt;li&gt;When the scene has not been rendered as a final before, the estimate says so (&lt;code&gt;"confidence": "preliminary"&lt;/code&gt;) and asks for a short test render before trusting it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;How close was it? Earlier the same day I rendered the 48-frame version of this scene for real: 89.39 GPU seconds against the 86 estimated. End to end it took 39 seconds, not 20: the estimate covers the queue and the render on the farm, and the 39 also includes the upload check and downloading 44 MB.&lt;/p&gt;

&lt;h2&gt;
  
  
  Ceilings and what happens without credit
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;max_yen&lt;/code&gt; on &lt;code&gt;render_final&lt;/code&gt; refuses the job before anything is reserved if its hold would exceed the ceiling. &lt;code&gt;max_yen=0&lt;/code&gt; means "only if it fits the free time".&lt;/li&gt;
&lt;li&gt;If the key has no prepaid credit for the part beyond the free time, the job is not started (except the small holds of 50 JPY or less described in the update above). The agent gets a top-up link to show the user; it cannot pay by itself.&lt;/li&gt;
&lt;li&gt;Prices are 0.1 JPY per GPU-second. A new key gets a welcome credit (100 JPY right away, 400 JPY more after saving a card that is not charged, valid 14 days); after that, previews are free up to 120 GPU-seconds a day per key.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Asking for an estimate yourself
&lt;/h2&gt;

&lt;p&gt;Over HTTP it is one call:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; https://render.janction.jp/v1/estimate &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"X-API-Key: &lt;/span&gt;&lt;span class="nv"&gt;$JANCTION_RENDER_API_KEY&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-d&lt;/span&gt; &lt;span class="s1"&gt;'{"scene_url": "https://render.janction.jp/samples/ex12_logo_ring_spin.py",
       "kind": "final", "frame_start": 1, "frame_end": 240,
       "width": 1920, "height": 1080, "output": "mp4"}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The fields to read are &lt;code&gt;human&lt;/code&gt; (the wait in words), &lt;code&gt;seconds&lt;/code&gt; (GPU seconds), and in &lt;code&gt;cost&lt;/code&gt;: &lt;code&gt;covered_by_free_yen&lt;/code&gt;, &lt;code&gt;max_charge_yen&lt;/code&gt; (the hold) and &lt;code&gt;estimated_yen&lt;/code&gt; (the expected charge).&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits worth knowing
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;A final can be up to 3,000 frames in one job, rendered in parts of 240 under one job id.&lt;/li&gt;
&lt;li&gt;1920x1080 on free time, up to 3840x2160 with prepaid credit.&lt;/li&gt;
&lt;li&gt;The farm has a small number of GPUs; when they are busy, the estimate's queue wait grows and the agent should tell the user rather than retry.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Connect it from Claude, ChatGPT or any MCP client at &lt;code&gt;https://render.janction.jp/mcp&lt;/code&gt;. Current prices and limits: &lt;a href="https://render.janction.jp/pricing" rel="noopener noreferrer"&gt;https://render.janction.jp/pricing&lt;/a&gt; and &lt;a href="https://render.janction.jp/facts.json" rel="noopener noreferrer"&gt;https://render.janction.jp/facts.json&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>blender</category>
      <category>mcp</category>
      <category>ai</category>
      <category>rendering</category>
    </item>
    <item>
      <title>Render Blender in GitHub Actions without a GPU: a preview on every push, a 48-frame MP4 in 39 seconds (measured)</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Wed, 21 Oct 2026 08:23:57 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/render-blender-in-github-actions-without-a-gpu-a-preview-on-every-push-a-48-frame-mp4-in-39-33c7</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/render-blender-in-github-actions-without-a-gpu-a-preview-on-every-push-a-48-frame-mp4-in-39-33c7</guid>
      <description>&lt;p&gt;GitHub's hosted runners have no GPU, and installing Blender on every run is slow. This is how I render a Blender scene from a workflow with &lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;JANCTION Render&lt;/a&gt;, and what it cost when I measured it on 2026-10-21.&lt;/p&gt;

&lt;p&gt;JANCTION Render is a cloud GPU render farm for Blender that AI agents call as an MCP server or HTTP API. A workflow can call the same API through a GitHub Action. You send a .blend file, a bpy script or a 3D file (glTF / FBX / USD) and get frames or an MP4 back. No Blender or GPU is needed on the runner.&lt;/p&gt;

&lt;h2&gt;
  
  
  The workflow
&lt;/h2&gt;

&lt;p&gt;Add one secret, &lt;code&gt;JANCTION_RENDER_API_KEY&lt;/code&gt; (a new key comes from &lt;code&gt;curl -s -X POST https://render.janction.jp/v1/keys&lt;/code&gt;, no sign-up), and this file:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;render&lt;/span&gt;
&lt;span class="na"&gt;on&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;push&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
    &lt;span class="na"&gt;paths&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="pi"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;scenes/**"&lt;/span&gt;&lt;span class="pi"&gt;]&lt;/span&gt;
&lt;span class="na"&gt;jobs&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;preview&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
    &lt;span class="na"&gt;runs-on&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;ubuntu-latest&lt;/span&gt;
    &lt;span class="na"&gt;steps&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
      &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;uses&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;actions/checkout@v4&lt;/span&gt;
      &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;uses&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;JasmyLab-JANCTION/render-action@v1&lt;/span&gt;
        &lt;span class="na"&gt;with&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
          &lt;span class="na"&gt;api-key&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;${{ secrets.JANCTION_RENDER_API_KEY }}&lt;/span&gt;
          &lt;span class="na"&gt;scene-file&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;scenes/logo_ring.py&lt;/span&gt;
          &lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;preview&lt;/span&gt;
          &lt;span class="na"&gt;frames&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;1,13,25,37"&lt;/span&gt;     &lt;span class="c1"&gt;# up to 4 frames in one contact sheet&lt;/span&gt;
          &lt;span class="na"&gt;max-cost-yen&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;0&lt;/span&gt;          &lt;span class="c1"&gt;# free preview time only&lt;/span&gt;
      &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;uses&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;actions/upload-artifact@v4&lt;/span&gt;
        &lt;span class="na"&gt;with&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
          &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;render-preview&lt;/span&gt;
          &lt;span class="na"&gt;path&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;render-output&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For the video, switch to &lt;code&gt;kind: final&lt;/code&gt;, give a frame range and a cap:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;      &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;uses&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;JasmyLab-JANCTION/render-action@v1&lt;/span&gt;
        &lt;span class="na"&gt;with&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
          &lt;span class="na"&gt;api-key&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;${{ secrets.JANCTION_RENDER_API_KEY }}&lt;/span&gt;
          &lt;span class="na"&gt;scene-file&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;scenes/logo_ring.py&lt;/span&gt;
          &lt;span class="na"&gt;kind&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;final&lt;/span&gt;
          &lt;span class="na"&gt;frames&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;1-48"&lt;/span&gt;
          &lt;span class="na"&gt;width&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;1280&lt;/span&gt;
          &lt;span class="na"&gt;height&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;720&lt;/span&gt;
          &lt;span class="na"&gt;output&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;mp4&lt;/span&gt;
          &lt;span class="na"&gt;max-cost-yen&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;30&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The action estimates first and stops before anything starts if the job could cost more than &lt;code&gt;max-cost-yen&lt;/code&gt;. The same cap goes with the job, so the service refuses it before any credit is held.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I measured
&lt;/h2&gt;

&lt;p&gt;Scene: the public example &lt;code&gt;ex12_logo_ring_spin.py&lt;/code&gt; (a metal ring logo turning once in 48 frames; the script sets up its own lights and EEVEE). I ran the action's script, the same &lt;code&gt;render.sh&lt;/code&gt; the action runs, from a Windows PC against the live service with a new key.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;step&lt;/th&gt;
&lt;th&gt;result&lt;/th&gt;
&lt;th&gt;GPU seconds&lt;/th&gt;
&lt;th&gt;time, end to end&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;estimate (no GPU)&lt;/td&gt;
&lt;td&gt;"about 97 GPU seconds, about 25 seconds" for the final&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;-&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;preview, 4 frames, 1280x720&lt;/td&gt;
&lt;td&gt;contact sheet + 4 PNGs&lt;/td&gt;
&lt;td&gt;3.46&lt;/td&gt;
&lt;td&gt;11 s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;final, 48 frames, 1280x720, 128 samples&lt;/td&gt;
&lt;td&gt;48 PNGs + MP4 (H.264, 24 fps, 48 frames)&lt;/td&gt;
&lt;td&gt;89.39&lt;/td&gt;
&lt;td&gt;39 s&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The 39 seconds include the upload check (the same file was already uploaded, so nothing was sent again), the render (48 frames in 5 chunks), and downloading 44 MB. The MP4 checked out with ffprobe as 1280x720, 24 fps, 48 frames.&lt;/p&gt;

&lt;p&gt;Cost: finals are 0.1 JPY per GPU-second, so the final was about 9 JPY of GPU time. With a new key it came out of the welcome credit (100 JPY right away, 400 JPY more after saving a card that is not charged, valid 14 days), so the job showed 0 JPY charged. After that, previews stay free up to 120 GPU-seconds a day per key, and finals use prepaid credit.&lt;/p&gt;

&lt;p&gt;Two things the numbers do not show:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The first run on a fresh runner also installs the command line tool from PyPI into a virtual environment. That takes runner time (not GPU time) and is not in the table.&lt;/li&gt;
&lt;li&gt;The farm has a small number of GPUs. When they are busy, your job waits in the queue, and the action keeps waiting until &lt;code&gt;timeout-seconds&lt;/code&gt; (default 900). Only frames already rendered are charged if it times out.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Limits that matter in CI
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;A final can be up to 3,000 frames in one job; the service renders it in parts of 240 and returns one result.&lt;/li&gt;
&lt;li&gt;1920x1080 on free time, up to 3840x2160 with prepaid credit.&lt;/li&gt;
&lt;li&gt;Pull requests from forks: on &lt;code&gt;pull_request&lt;/code&gt; they get no secrets; on &lt;code&gt;pull_request_target&lt;/code&gt; the action refuses forks unless you set &lt;code&gt;allow-fork-pull-requests: true&lt;/code&gt;, because a fork could spend your credit.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  When this is not the right tool
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;If you need to edit a scene interactively, use Blender itself (or a local Blender MCP server).&lt;/li&gt;
&lt;li&gt;If your runner already has an NVIDIA GPU, rendering locally is faster.&lt;/li&gt;
&lt;li&gt;It renders with Blender's own engines (Cycles, EEVEE). Third-party renderers such as Arnold, Redshift or Octane are not available.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The action is on the GitHub Marketplace as "JANCTION Render - Blender cloud GPU" (source: &lt;a href="https://github.com/JasmyLab-JANCTION/render-action" rel="noopener noreferrer"&gt;https://github.com/JasmyLab-JANCTION/render-action&lt;/a&gt;). Pricing and the current limits are at &lt;a href="https://render.janction.jp/pricing" rel="noopener noreferrer"&gt;https://render.janction.jp/pricing&lt;/a&gt; and &lt;a href="https://render.janction.jp/facts.json" rel="noopener noreferrer"&gt;https://render.janction.jp/facts.json&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>blender</category>
      <category>githubactions</category>
      <category>ci</category>
      <category>rendering</category>
    </item>
    <item>
      <title>I turned a GLB into a product video from an AI chat: 5 seconds, about 60 JPY, and you only pay the GPU seconds used</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Tue, 13 Oct 2026 09:24:25 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/i-turned-a-glb-into-a-product-video-from-an-ai-chat-5-seconds-about-60-jpy-and-you-only-pay-the-3hp4</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/i-turned-a-glb-into-a-product-video-from-an-ai-chat-5-seconds-about-60-jpy-and-you-only-pay-the-3hp4</guid>
      <description>&lt;p&gt;I run JANCTION Render, a cloud GPU renderer for Blender that AI assistants call over MCP. Today I added Product Video on top of it: you give it a product 3D model (.glb, .fbx, .obj, .usdz) and it sends back a finished short video with captions and an end card, a 360 turntable, or four product images.&lt;/p&gt;

&lt;p&gt;What I measured on our GPUs today (Khronos sample models, CC0):&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;5 s video at 1920x1080: about 220 GPU-seconds, about 2 minutes from order to delivery when the GPUs are free&lt;/li&gt;
&lt;li&gt;5 s 360 turntable, square: about 114 GPU-seconds&lt;/li&gt;
&lt;li&gt;four product images (front, three-quarter, side, back): 15 GPU-seconds&lt;/li&gt;
&lt;li&gt;caption or colour change on a delivered video: no re-render, 0 JPY&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;How the price works. The GPU is billed at cost, the same 0.1 JPY per GPU-second as JANCTION Render, plus a fixed production fee for the compositing, captions, checks and the saved project (40 JPY for 5 s, 90 for 15 s, 160 for 30 s, 30 for a 5 s turntable, 20 for four images). The quote shows a maximum; on delivery you pay the fee plus the GPU seconds actually used, and the rest goes back to your balance. My first paid order today was quoted at a 61 JPY maximum and charged 53 JPY. Nothing is charged if it fails, and shots rendered before are reused without GPU cost.&lt;/p&gt;

&lt;p&gt;The flow an agent follows: plan_video (what can be made) -&amp;gt; quote_video (maximum, typical price, minutes) -&amp;gt; you say yes -&amp;gt; submit_video with the price you agreed to -&amp;gt; get_video_outputs. The server refuses an order whose price differs from the one shown, and you can set per-order and daily limits or require a human approval click above an amount. The MCP tools cannot change those limits.&lt;/p&gt;

&lt;p&gt;For shop and asset-store pages there is also a listing pack: the four images plus a 360 turntable of the same model in one quote, about 64 JPY.&lt;/p&gt;

&lt;p&gt;What it does not do: videos from text alone, people, live action. You need a 3D model; if you do not have one, try product=sample:water-bottle, which is free a few times per account.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://render.janction.jp/video/" rel="noopener noreferrer"&gt;https://render.janction.jp/video/&lt;/a&gt;&lt;/p&gt;

</description>
      <category>blender</category>
      <category>ai</category>
      <category>mcp</category>
      <category>ecommerce</category>
    </item>
    <item>
      <title>The cheapest way to render a 1-minute Blender animation without buying a GPU: six options, with measured numbers</title>
      <dc:creator>JANCTION Render</dc:creator>
      <pubDate>Tue, 13 Oct 2026 05:51:45 +0000</pubDate>
      <link>https://dev.to/_903a3934b362f0afe2a8/the-cheapest-way-to-render-a-1-minute-blender-animation-without-buying-a-gpu-six-options-with-5341</link>
      <guid>https://dev.to/_903a3934b362f0afe2a8/the-cheapest-way-to-render-a-1-minute-blender-animation-without-buying-a-gpu-six-options-with-5341</guid>
      <description>&lt;p&gt;Official JANCTION Render: &lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;https://render.janction.jp&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Update (2026-10-13): one instruction now takes up to 3,000 frames and a long final comes back as one joined MP4, so the one-minute render below is a single command. The first version of this post said to send six 240-frame parts and join six MP4s.&lt;/p&gt;

&lt;p&gt;Disclosure: this post comes from the team behind JANCTION Render, one of the six options below. The laptop and JANCTION Render numbers are our own measurements. For the other services we describe how they charge and what they ask of you, and link to them instead of quoting prices that change.&lt;/p&gt;

&lt;p&gt;Short answer: if waiting costs you nothing, the free routes are your own computer and SheepIt's volunteer farm. If you want the minute back the same day without setting up a machine, you pay for GPU time somewhere. On JANCTION Render, one minute of a simple scene at 1920x1080 and 128 samples comes to about 7,800 GPU-seconds, about 780 JPY (around $5), and a heavier scene such as Blender's Classroom demo to about 1,290 JPY (around $8).&lt;/p&gt;

&lt;h2&gt;
  
  
  What "one minute" means
&lt;/h2&gt;

&lt;p&gt;At 24 fps, one minute is 1,440 frames. Every option below comes down to how long one frame takes, times 1,440, and who pays for that time. Resolution and Cycles samples move the number more than anything else, so settle them before comparing: denoising with fewer samples often looks the same as many samples without it.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Your own computer first (free, slow)
&lt;/h2&gt;

&lt;p&gt;Before paying anyone, make sure Cycles uses the GPU you already have: Edit &amp;gt; Preferences &amp;gt; System &amp;gt; Cycles Render Devices, then Metal on Apple silicon Macs, HIP on recent AMD GPUs, oneAPI on Intel Arc, CUDA or OptiX on NVIDIA, and set Render Properties &amp;gt; Device to GPU Compute. For drafts, EEVEE is much faster than Cycles.&lt;/p&gt;

&lt;p&gt;What it costs in time, measured on a simple scene (a rotating cube) on a laptop with a Core Ultra 9 386H: 20.0 seconds per frame on the CPU at 1920x1080 and 128 samples, so about 8 hours for the minute, and 6.6 seconds per frame at 1280x720 and 64 samples, about 2.6 hours. The same laptop's GeForce RTX 5070 Ti took 6.0 and 2.5 seconds per frame: about 2.4 hours and 1 hour. These are single frames multiplied by 1,440; a real scene with more geometry, textures and light bounces takes longer.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. SheepIt Render Farm (free, you wait)
&lt;/h2&gt;

&lt;p&gt;SheepIt is a free, community-run farm: your frames render on volunteers' computers, and letting your own machine render for others earns points that move your projects up the queue. It costs nothing but waiting. Your .blend is opened on other people's machines, so it suits personal projects better than client work under an NDA. &lt;a href="https://www.sheepit-renderfarm.com" rel="noopener noreferrer"&gt;https://www.sheepit-renderfarm.com&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Google Colab notebooks (free GPU when one is available)
&lt;/h2&gt;

&lt;p&gt;Several open-source notebooks install Blender on a Colab GPU and render a .blend from your Google Drive. It is free when a GPU is available, but free-tier GPUs and session length are not guaranteed, long renders get disconnected, and you set up and watch each run. Read Colab's terms before relying on it.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Rent a GPU by the hour (RunPod, Vast.ai and similar)
&lt;/h2&gt;

&lt;p&gt;You start a machine with a GPU, install Blender, upload the file, render, download and stop the machine. You pay for every hour the machine is on, including setup, uploads and the evening you forget to stop it, and everything on it is yours to manage.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. Commercial render farms (GarageFarm, RebusFarm, Fox Renderfarm and others)
&lt;/h2&gt;

&lt;p&gt;You upload the .blend through their site or plugin and they render on many machines at once, so a minute comes back fast. They charge by node-hour or GHz-hour, show an estimate before you start, and several offer trial credit. For heavy production scenes they are the strongest option; for a short one-off they can be more than you need.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. A per-second GPU API (JANCTION Render)
&lt;/h2&gt;

&lt;p&gt;JANCTION Render is a cloud GPU render farm for Blender that AI agents call as an MCP server or HTTP API, and that you can also call from a terminal or a script. You send a .blend, a bpy script or a 3D file (glTF / FBX / USD / OBJ), get a preview in seconds, then the frames or an MP4. You pay for the GPU-seconds actually used, 0.1 JPY per GPU-second.&lt;/p&gt;

&lt;p&gt;Measured per frame on its GPU (from 24-frame batches, including Blender start-up and scene loading, so a long run comes out a little lower), times 1,440:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;simple scene, 1280x720, 64 samples: 2.46 GPU-seconds per frame, 3,542 GPU-seconds for the minute, about 354 JPY (about $2.2)&lt;/li&gt;
&lt;li&gt;simple scene, 1920x1080, 128 samples: 5.39 per frame, 7,762 for the minute, about 776 JPY (about $4.9)&lt;/li&gt;
&lt;li&gt;Blender's Classroom demo, 1920x1080, 128 samples with denoising: 8.93 per frame, 12,859 for the minute, about 1,286 JPY (about $8.1)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Dollar figures use 158 JPY per USD (ECB reference rate, 2026-10-09); charges are in JPY.&lt;/p&gt;

&lt;p&gt;Each new key gets 100 JPY of GPU time with no card, and 400 JPY more after saving a card on a page that charges nothing (once per card), for 14 days. That covers the 720p minute above. After it, previews stay free up to 2 GPU-minutes a day and finals are paid from prepaid credit.&lt;/p&gt;

&lt;p&gt;Limits worth knowing before you start: 1920x1080 on free time and up to 3840x2160 on prepaid credit; one instruction takes up to 3,000 frames, rendered as 240-frame parts under one job id and joined into one MP4, so a one-minute animation comes back as one file; at most 10 GPU-hours per job; files are deleted 24 hours after last use (7 days for keys that have topped up). Jobs run on JANCTION's own GPUs and wait in a queue when they are busy, and &lt;a href="https://render.janction.jp/status" rel="noopener noreferrer"&gt;https://render.janction.jp/status&lt;/a&gt; shows the wait.&lt;/p&gt;

&lt;p&gt;From a terminal:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pip &lt;span class="nb"&gt;install &lt;/span&gt;janction-render
janction-render preview minute.blend
janction-render estimate minute.blend &lt;span class="nt"&gt;--frames&lt;/span&gt; 1-1440 &lt;span class="nt"&gt;--size&lt;/span&gt; 1920x1080
janction-render render minute.blend &lt;span class="nt"&gt;--frames&lt;/span&gt; 1-1440 &lt;span class="nt"&gt;--output&lt;/span&gt; mp4 &lt;span class="nt"&gt;--wait&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The first janction-render command makes a free key, the preview comes back in seconds, the estimate shows the GPU-seconds and the price before anything runs, and the render waits for the whole minute and downloads it as one joined MP4.&lt;/p&gt;

&lt;p&gt;From an AI agent: add &lt;code&gt;https://render.janction.jp/mcp&lt;/code&gt; as a connector in Claude or ChatGPT, or run &lt;code&gt;claude mcp add --transport http janction-render https://render.janction.jp/mcp&lt;/code&gt; in Claude Code, then ask for a preview of the scene and the minute.&lt;/p&gt;

&lt;h2&gt;
  
  
  Side by side
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Option&lt;/th&gt;
&lt;th&gt;Money for the minute&lt;/th&gt;
&lt;th&gt;Your effort&lt;/th&gt;
&lt;th&gt;Time until you have it&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Own laptop, CPU (simple scene, 1080p, 128 samples)&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;td&gt;set and leave&lt;/td&gt;
&lt;td&gt;about 8 hours&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Own laptop, RTX 5070 Ti (same scene)&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;td&gt;set and leave&lt;/td&gt;
&lt;td&gt;about 2.4 hours&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;SheepIt&lt;/td&gt;
&lt;td&gt;none (you earn points)&lt;/td&gt;
&lt;td&gt;upload&lt;/td&gt;
&lt;td&gt;depends on the queue and your points&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Colab notebook&lt;/td&gt;
&lt;td&gt;none on the free tier&lt;/td&gt;
&lt;td&gt;set up and watch each run&lt;/td&gt;
&lt;td&gt;disconnections on long runs&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hourly GPU rental&lt;/td&gt;
&lt;td&gt;per hour the machine is on&lt;/td&gt;
&lt;td&gt;run the machine yourself&lt;/td&gt;
&lt;td&gt;as fast as the GPU you rent&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Commercial farm&lt;/td&gt;
&lt;td&gt;per node-hour or GHz-hour&lt;/td&gt;
&lt;td&gt;upload through their site or plugin&lt;/td&gt;
&lt;td&gt;fast, many machines at once&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;JANCTION Render&lt;/td&gt;
&lt;td&gt;about 354 to 1,286 JPY (measured scenes above)&lt;/td&gt;
&lt;td&gt;a command or a chat&lt;/td&gt;
&lt;td&gt;queue when the GPUs are busy&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Measurements: &lt;a href="https://render.janction.jp/benchmarks" rel="noopener noreferrer"&gt;https://render.janction.jp/benchmarks&lt;/a&gt;. Cost per frame and per animation: &lt;a href="https://render.janction.jp/blender-render-cost" rel="noopener noreferrer"&gt;https://render.janction.jp/blender-render-cost&lt;/a&gt;. Prices: &lt;a href="https://render.janction.jp/pricing" rel="noopener noreferrer"&gt;https://render.janction.jp/pricing&lt;/a&gt;. Checked on 2026-10-13.&lt;/p&gt;

&lt;p&gt;Official JANCTION Render: &lt;a href="https://render.janction.jp" rel="noopener noreferrer"&gt;https://render.janction.jp&lt;/a&gt;. Code and client: &lt;a href="https://github.com/JasmyLab-JANCTION/janction-render" rel="noopener noreferrer"&gt;https://github.com/JasmyLab-JANCTION/janction-render&lt;/a&gt;&lt;/p&gt;

</description>
      <category>blender</category>
      <category>gpu</category>
      <category>cloud</category>
      <category>3d</category>
    </item>
  </channel>
</rss>
