Update (2026-10-12): free to start, 100 JPY of GPU time per new key with no card and 400 JPY more after saving a card on a page that charges nothing (once per card), up to 500 JPY in all; then 0.1 JPY per GPU-second, and previews stay free up to 2 GPU-minutes a day. Paid use started on 2026-10-08, and this post describes the current pricing. Auto top-up and monthly plans are optional (https://render.janction.jp/pricing).
JANCTION Render (https://render.janction.jp) renders AI-written scenes and existing .blend files on a cloud GPU. It is a separate product from SmartRender, JasmyLab's distributed-GPU platform.
A 4-second Blender preview and a 4-second round trip are not the same claim. JANCTION Render publishes both, for a few simple scenes, and they answer different questions. This note is how to reproduce the published matrix and how not to quote it.
This post is from JasmyLab, which operates JANCTION Render. The numbers below are individual observations from production logs, not a speed guarantee and not a comparison against Colab, SheepIt, or another farm.
Machine-readable table: https://render.janction.jp/benchmarks.json
Human page: https://render.janction.jp/benchmarks
The two clocks
GPU seconds are worker run time. They include starting Blender and loading the scene. They are not "the GPU was busy for this many seconds" in a profiler sense, and they do not include time the job spent queued while the GPU was lent to something else.
Round trip is submission until the result is received. In the published cases the queue was empty. Input upload is not guaranteed to be inside that number for every case. Do not subtract GPU seconds from the round trip and call the remainder "upload," and do not add a guessed upload on top of GPU seconds.
Checked against the production description on 2026-10-02:
| Example | What was rendered | GPU time | Round trip |
|---|---|---|---|
| 4-frame preview | Rotating cube, four 640×360 tiles, 16 samples, scene lights | about 4 s | about 8 s |
| Short MP4 | 24 frames, 1280×720, 64 samples | about 58.7 s | about 65 s |
GPU: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96 GB. Dates on the benchmark page are in UTC. The 24-frame case is from 2026-09-30 / 2026-10-01 logs.
What you can re-run
The matrix uses small scenes you can download, including https://render.janction.jp/samples/cube_scene.py. The page's command shape is:
janction-render render cube_scene.py --frames 1-24 --size 1280x720 --samples 64 --output png --wait
Those files reproduce the matrix scenes. They are not bit-identical to every historical Claude.ai transcript that produced a similar cube. If your re-run disagrees, keep your resolution, samples, frame count, engine, Blender version, and whether the queue was empty next to the number. Cycles is the default. EEVEE is available on final jobs and adds per-job shader compilation, so a one-frame EEVEE timing is a poor way to estimate a long EEVEE animation.
render_estimate is the number to use for a new scene. After that scene has been rendered once, later estimates can use its own history. The published "about 11 seconds" for one 1920×1080 frame at 128 samples is the estimator's baseline for a light scene, not a promise.
What the matrix seemed to say, and what it does not
On those simple scenes, moving from 32 to 128 samples added roughly ten percent, and a 100-frame run was about four times a 24-frame run. Per-frame time was dominated by fixed work: scene build, denoising, file write. About 2.4 seconds per 720p frame and about 5.2 seconds per 1080p frame showed up in that table.
That pattern is about those scenes. A volume, a large texture, or a shader that actually spends its time on samples will not follow it. The benchmark page says so. Quoting "samples are free, only resolution matters" without the scene is a misquote.
The free credit is a third clock
The free allowance is counted in the same GPU seconds. Each new key starts with a welcome credit of up to 500 JPY, 5,000 GPU-seconds for 14 days (100 JPY with no card, 400 JPY more after saving a card that is not charged), covering previews and finals; after that, previews stay free up to 2 GPU-minutes (120 GPU-seconds) a day, reset at 00:00 UTC. The 24-frame example above is about one GPU-minute, so the welcome credit covers about 85 such clips only while the scene stays that cheap. A heavier scene uses it much faster. A job the free credit does not cover gets HTTP 402 payment_required with a checkout link. The job is refused, not held. Splitting it does not create more free credit.
Since 2026-10-08, GPU time beyond the free credit costs 0.1 JPY per GPU second from prepaid credit, so the 24-frame example comes to about 6 JPY.
Do not use these numbers against another service
There is no controlled Colab run in this dataset. Colab's own FAQ says GPU type, idle timeout, and lifetime vary, and that free notebooks run for at most 12 hours depending on availability. https://research.google.com/colaboratory/faq.html
There is no controlled SheepIt run either. SheepIt is free and has no API. A long animation there can finish sooner when volunteers pick up frames. That is a different system. https://www.sheepit-renderfarm.com/
Sceneplane's free plan, as published on its site on 2026-10-02, includes 30 compute minutes per month. Compute there includes scene building and export. It is not a GPU-second figure you can subtract from the table above. https://sceneplane.online/
Top comments (0)