DEV Community: GL

Seedance 2.5 Camera Control: Block Out AI Film Shots in Blender First

GL — Thu, 22 Oct 2026 13:14:59 +0000

AI video models are great at pictures and bad at taking direction. Ask for "a slow dolly-in on a mob boss behind his desk, cat on his lap, 76 mm lens" and you get a shot, not your shot. The camera drifts, the framing changes between takes, and the actor ends up wherever the model puts him.

So we stopped describing the camera and started building it. Every shot below was blocked out in Blender first, with Blender AI: a gray set, stand-in actors, real lenses and a real camera move. Then that blockout drove Seedance 2.5, so the AI video follows our camera frame for frame.

This post walks through three of those films: a Godfather-style office scene, a cyberpunk short called Hope, and a wizard-school potions class. Then it covers the exact workflow behind them.

TL;DR: camera control for AI video

Block the shot out in Blender first. Set, cast, rigs, lens and camera move, all as gray geometry. You don't need finished models.
Let the blockout drive the AI video. The AI video model gets the motion and the framing from your Blender camera instead of inventing its own.
Use the AI for the look, Blender for the direction. You keep control of composition, timing and lens; Seedance 2.5 handles light, texture and performance.
The Blender side is done by 3D-Agent, an AI agent that works inside your open Blender scene. You describe the set, it builds it.

Film 1: the Don's office (Blender blockout → Seedance 2.5)

This is the shot everyone asks about. The finished frame looks like a 1970s crime drama: a dim office, the Don behind his desk, a cat on his lap. Here's how it was made.

1. The blockout

3D-Agent built the whole shot as a gray blockout in Blender from plain-language prompts:

The room: walls, desk, chairs, window and the props that matter for the framing.
The cast: the Don, the cat and the film crew, as simple rigged stand-ins.
The camera: a 76 mm lens on a 2.7 m dolly, with a real dolly-in move keyframed on the timeline.
Motion: motion capture and rigs, so the actors move on the right beats.

None of it is pretty, and that's the point: a blockout is about where things are and how the camera moves, not how they look.

2. The camera move

Because the camera is a real Blender camera, you direct it like on set: focal length, dolly distance, height, timing. Change "76 mm" to "35 mm" and the whole perspective changes, the way it would on a real lens. Move a keyframe and the push-in starts later. AI video prompts can't give you that level of control.

3. The blockout drives the AI shot

3D-Agent then feeds that blockout to Seedance 2.5. The AI video follows the Blender camera frame for frame: the dolly, the framing and the timing all come from the blockout. Seedance supplies the look: skin, fabric, wood, smoke, light.

Film 2: Hope, a cyberpunk AI animated short

Hope is a short cyberpunk film made the same way, but taken all the way to a finished edit. The breakdown video covers:

A Blender motion reference made with 3D-Agent: a simple animation that fixes the character's movement and the camera path before any AI video is generated.
Planning character movement and camera shots in Blender, so every shot has a clear start, end and lens.
A character sheet, a first frame and a last frame, so the character stays the same across shots.
Image and video prompts written against that reference instead of from scratch.
The final edit in DaVinci Resolve.

The motion reference is what keeps the film consistent. Each AI shot starts and ends where the Blender reference says it does, so the cuts line up.

Film 3: a wizard-school potions class, with a handheld camera

Our newest one, and the busiest shot so far: a sunny potions classroom with 12 students and a teacher, a broom rider who nearly crashes through the room, and a handheld phone move that ends at the window.

Blender blockout: 3D-Agent built the classroom and the whole cast as gray stand-ins, so every student has a fixed seat and the broom rider has a real flight path.
Camera move and rig: a handheld-style phone camera, rigged and keyframed in Blender. The shake, the drift and the turn to the window are all yours, not the model's.
The blockout drives the AI shot: Seedance 2.5 followed that camera frame for frame, near-crash included.

A dozen people in one moving shot is where prompt-only AI video falls apart: people swap places, appear and disappear. In the blockout they're pinned in 3D space, so they stay where they belong.

Why a Blender blockout beats a longer prompt

Inside Blender: the rigged character (left) and the clay render (right).

AI filmmaking tools usually offer camera control in one of two ways: preset moves ("dolly in", "orbit left") or a text description. Both leave the model guessing about:

What you want	Prompt only	Blender blockout first
Exact lens and framing	Approximate	Exact: a real focal length on a real camera
Camera path and timing	Model's interpretation	Your keyframes, frame for frame
Actor positions	Random per take	Fixed in 3D space
Consistency across shots	Hard	Same set and cameras for every shot
Iteration	Re-roll and hope	Move a keyframe, re-render

A blockout is a few minutes of work in Blender, and much less with an AI agent building it for you. It turns the AI video model from a slot machine into a renderer that takes direction.

The workflow, step by step

Describe the set to Blender AI. "A 1970s office: desk facing the camera, window behind it, two leather chairs, a lamp." 3D-Agent builds it in your open Blender scene as editable objects.
Add the cast as stand-ins. Simple rigged figures are enough. Pose them, or add motion for the beats that matter.
Set the camera like a cinematographer. Pick the lens (say 76 mm), place it on a path or dolly, and keyframe the move and its timing.
Check it in the viewport. Scrub the timeline. If the framing or timing is wrong, fix it here, where it's cheap.
Generate the AI shot from the blockout. Seedance 2.5 follows your camera and composition and adds the look.
Repeat per shot and edit. Keep the same set and character sheet for continuity, then cut it together.

More shots made this way

"The Letter": from a simple Blender scene to a cinematic AI short, using Blender motion references to guide timing, camera movement and action.

FAQ

Can you control the camera in Seedance 2.5?
Yes. Instead of describing the move in text, block the shot out in Blender with a real camera and lens, then let that blockout drive the AI video. Seedance 2.5 follows the Blender camera frame for frame.

Do I need to be a Blender expert?
No. With Blender AI you describe the set, the cast and the camera in plain English and the agent builds them in Blender. You can still adjust everything by hand, because it's all normal Blender objects.

Is this an AI movie maker?
It's a workflow, not a single button. Blender gives you direction (camera, lens, staging, timing) and an AI video model gives you the look. Together they work like an AI movie maker that takes direction.

Does the blockout need finished models or textures?
No. Gray boxes and simple rigs are enough; the AI video model supplies the materials and lighting. The blockout only needs correct positions, scale, motion and camera.

What does 3D-Agent cost?
It's a paid desktop app for macOS and Windows that works with Blender 4.2 or later: Basic $19/month, Pro $29/month, Ultra $89/month, no free tier. See 3d-agent.com/download.

Try it

Everything in the blockout is ordinary, editable Blender geometry with clean quad topology.

Pick one shot you've been trying to get out of an AI video model with prompts alone. Block it out in Blender first: set, stand-ins, lens, one camera move. Then generate. The difference in control is obvious from the first try.

Blender AI guide: 3d-agent.com/blender-ai
AI animation in Blender: 3d-agent.com/blender-ai/animation
More films and builds: 3d-agent.com/showcase
Download 3D-Agent: 3d-agent.com/download
YouTube: @3D-Agent

Which shot should we break down next? Tell me in the comments.

Roblox AI in Blender: How We Built the Roblox Backrooms with 3D-Agent and Claude Opus 5.5

GL — Wed, 30 Sep 2026 18:53:00 +0000

I'm not a Roblox dev. I've never modeled a Backrooms level by hand, and I've never rigged a Roblox character without watching three tutorials first.

But over the last few weeks we made a whole Roblox Backrooms series, four episodes, and nobody on our team pushed a single vertex. Every room, light, character and animation was built in Blender by prompting 3D-Agent. This is how Roblox AI actually looks when it works inside a real 3D tool, and I want to show you exactly what it did.

Here's the full story so far:

First, what is 3D-Agent?

Quick context so the rest makes sense. 3D-Agent is a desktop app for macOS and Windows that works on your open scene in Blender 4.2 or later. You describe what you want, and it builds it step by step with Blender's own tools, so everything you get is normal, editable Blender geometry.

It has its own agent (it's not a Blender MCP server), it installs its own Blender extension, and there's no API key to set up. For this series it ran on Claude Opus 5.5 + GPT-6 Astra.

Roblox Backrooms, built in Blender with prompts

The idea: Noob, Bacon and Hazmat get trapped in Backrooms Level 0. The walls keep moving, an entity hunts them, and someone wrote 3D-AGENT.COM on the wall. Then they find the computer that's building the rooms.

The yellow corridors, the ceiling lights, the Roblox-style characters, the entity in the dark: all prompted. Later episodes go to the Poolrooms and to AI copies of the characters.

The part I like most: nothing is a baked video or a fake render. Every wall is still a real object in Blender. You can open the scene and keep editing it.

Roblox map maker: one screenshot, one prompt

This one surprised me. We gave 3D-Agent one screenshot of a Roblox map and one prompt. It rebuilt the whole thing in Blender: 13 cottages, seven watchtowers, a windmill and a bell tower. Real, editable geometry, not a generated shell.

If you've ever tried to recreate a map from a reference by hand, you know how long that takes.

Roblox Blender rigging in 3 prompts

Rigging is where I usually give up. Here it took three prompts:

Build the character as 15 separate body parts from one reference picture.
One bone per part, then a six-second walk cycle at 30 fps.
Light it, render it, and export an FBX for Roblox Studio.

No manual weight painting.

Roblox Studio AI: getting it into Studio

The classic route still works: export FBX or OBJ from Blender and bring it in with the Studio 3D Importer. We wrote a guide for that: Blender to Roblox.

The newer part: 3D-Agent can now connect to Roblox Studio directly. It can inspect your project, edit scripts, insert assets, run playtests and read the console. There's also a Roblox plugin that repairs scenes, attachments and rigs, and publishes through Open Cloud. More here: Roblox Studio AI with 3D-Agent.

What it's not good at (being honest)

It's not a one-shot "Roblox AI model generator" that spits out a finished asset from one sentence. It works like an artist you brief: step by step, and you steer it. It's great at maps, props, buildings and blocky characters. Photoreal faces and very organic sculpts are still better done by hand.

Episode 4 is out

The finale: Noob, Bacon and Hazmat find the server room where the Backrooms get built, turn to gray clay, and wake up trapped inside Blender. Someone keeps typing...

Want to try it?

Download: 3d-agent.com (macOS and Windows, Blender 4.2+)
Plans start at $19/month (Basic, 100 prompts). Pro is $29 and Ultra is $89.
Come hang out on Discord. I'm there a lot.

So, what should we build next in the Backrooms? Drop it in the comments. The best idea goes into Episode 5.

So happy to have 3D-Agent.com on the Top 7!

GL — Tue, 02 Dec 2025 21:14:09 +0000

3D-Agent and Blender-MCP: Both Amazing Blender AI Tools

GL — Mon, 01 Dec 2025 16:49:05 +0000

From Blender-MCP to 3D-Agent: The Evolution of AI-Powered Blender Modeling

GL ・ Dec 1

#blender #mcp #ai #design

Blender MCP vs 3D-Agent: Two Ways to Put AI in Blender (2026 Update)

GL — Mon, 01 Dec 2025 16:24:21 +0000

Update, September 2026: I rewrote this post. A lot changed in Blender + AI this year: Blender now has its own official MCP server, Anthropic shipped a Claude connector for Blender, and 3D-Agent became a desktop app with its own agent. The old version compared tools that no longer look like that, so here is the current picture.

If you're exploring AI inside Blender, you'll run into two different ideas that sound similar: Blender MCP and 3D-Agent. They solve the same problem in different ways, and picking the right one depends on how you like to work.

What is Blender MCP?

Blender MCP is the Model Context Protocol applied to Blender. An MCP server exposes Blender's operations as tools, and an AI client (Claude, Cursor, Codex and others) calls those tools to build or change your scene. The AI client does the thinking; the MCP server gives it hands inside Blender.

Today there are three main ways to get it:

Blender Lab MCP Server, the official one from the Blender project (blender.org/lab/mcp-server). It needs Blender 5.1 or later, an MCP add-on and an AI client of your choice.
Claude's Blender connector, from Anthropic's Claude for Creative Work (April 2026). It's the Claude Desktop route to Blender's own MCP server.
The open-source blender-mcp on GitHub (now also called "MCP for Blender"). Free and community-maintained, and it works with many MCP clients.

With any of these you bring your own AI client and subscription, and you set up the connection yourself.

What is 3D-Agent?

3D-Agent is not an MCP server. It's a desktop app for macOS and Windows with its own AI agent, and it works on your open scene in Blender 4.2 or later. You describe what you want in plain language, and the agent plans the steps and runs real Blender operations, checking each result in the viewport before moving on.

It installs and updates its own Blender extension, so there's no add-on to install by hand, no MCP server to configure and no API key. Paid plans include prompts.

Check the clean wireframes!

A real example: a spiral staircase

Same prompt: "Create a spiral staircase with metal railings". Here is what 3D-Agent built:

It took about 8 minutes, and I exported it straight to FBX for Unity. The thing I care about most is the topology: clean geometry is what decides whether you can actually keep working on the output.

Blender MCP vs 3D-Agent

	Blender MCP (official server, Claude connector, or open source)	3D-Agent
What it is	An MCP server your AI client talks to	A desktop app with its own agent
Who does the 3D work	Your AI client (Claude, Cursor, Codex...)	3D-Agent's agent
Setup	Install the server or add-on, connect an MCP client	Install the app; it sets up its Blender extension
Blender version	5.1+ for the official server	4.2 or later
AI subscription / API key	Your own	Not needed; prompts are included
Cost	Free server; you pay for your AI client	From $19/month
Roblox Studio	No	Yes: inspect projects, edit scripts, insert assets, run playtests

I keep a more detailed, sourced version of this table on 3d-agent.com/blender-mcp.

Which one should you choose?

Choose Blender MCP if you already use Claude, Cursor or Codex, want to pick your own model, like wiring things up yourself, or want something free and open source. Setup guides and the original open-source server: blendermcp.org.
Choose 3D-Agent if you want an agent that works on your open scene with no MCP setup and no API keys, on macOS or Windows, with the prompts included in a monthly plan.

And you don't have to pick just one. On the Ultra plan, 3D-Agent runs its own local MCP server, so Claude Code, Claude Desktop, Cursor or Codex can hand whole Blender tasks to 3D-Agent. The setup is in the external agents docs.

Try it

Download 3D-Agent for macOS or Windows (Blender 4.2 or later)
Plans: Basic $19/month (100 prompts), Pro $29 (200), Ultra $89 (800). There's no free tier.
Come say hi on Discord. I read everything there.

Questions?

Drop a comment if you have questions about choosing between the two, specific use cases or how it works under the hood. I read every comment.

How I Built Cursor for Blender!

GL — Thu, 27 Nov 2025 13:49:59 +0000

How I Built a Blender AI Plugin That Generates 3D Models from Text: 3D-Agent

GL ・ Nov 26

#blender #3dagent #ai #3dprinting

How I Built a Blender AI Plugin That Generates 3D Models from Text: 3D-Agent

GL — Wed, 26 Nov 2025 17:18:02 +0000

Update, September 2026: a few things in this post were out of date, so I fixed them. 3D-Agent is now a desktop app for macOS and Windows that works on your open scene in Blender 4.2 or later. It installs its own Blender extension (no addon .zip, no API key), runs its own agent (it's not a Blender MCP server), and it's paid: plans start at $19/month, and there's no free tier.

TLDR: I built Cursor for Blender: a Blender AI plugin that generates editable 3D models, keeps wireframes clean, and can help clean up topology from other AI-generated assets.

Want an AI assistant inside your Blender session? Start with the 3D-Agent Blender AI plugin.

Let me start with a confession.

I'm a terrible 3D artist. Like, genuinely bad. Every time I opened Blender, I'd spend three hours trying to model something simple and end up with a geometry crime that looked like it was designed by someone who has never seen a real object before.

But here's the thing. I needed 3D assets for my projects. Game prototypes, architectural mockups, product visualizations. And hiring a 3D artist for every little thing, especially to send proposals and make drafts, was not in the budget.

So I did what any developer with too much free time and not enough common sense would do.

I built an AI that does it for me.

Actually, I googled first for Blender AI plugin and Blender AI tools and mostly found Blender-MCP experiments, which were exciting but did not meet my expectations for a daily production workflow.

May 2026 Update: Blender AI Is Not a Hypothetical Anymore

Since this post was first published, the conversation around AI inside Blender has moved from "cool experiment" to real industry debate. Anthropic announced Claude for Creative Work with Blender connection workflows, and the Blender community discussion around AI made one thing obvious: artists care about control, transparency, and whether AI tools respect the creative process.

That is exactly the gap I want 3D-Agent to fill. It is not just a one-shot text-to-3D generator. 3D-Agent is a standalone company building a multi-agent Blender workflow around editable scenes, cleaner topology, and practical production use inside Blender.

If you want the current product page, use the Blender AI page. If you want to see how it works inside Blender, start with the Blender AI plugin page.

The "Aha" Moment

It started when I was playing around with AI assistants and Blender's Python API.

I thought: what if an AI could actually use Blender, the way I do?

What if I could just type "spiral staircase" and have it actually appear in my scene? Like an AI 3D model generator?

Turns out, you can. And it works way better than I expected.

Meet 3D-Agent

3D-Agent is the result of that experiment. It's a desktop app for macOS and Windows that works on your open scene in Blender: you describe what you want, and it models it with Blender's own tools.

You type what you want. The AI figures out how to build it. Blender renders it. Done.

No more dragging vertices around for hours. No more watching YouTube tutorials at 2am trying to understand subdivision surfaces. Just describe your vision and let the machine do the heavy lifting.

How It Actually Works Under the Hood

Since this is dev.to and you probably want the technical details:

The Blender Side

3D-Agent installs its own Blender extension, which hooks into Blender's Python API. When you submit a prompt, the agent plans the steps and runs real Blender operations one by one, checking the result in the viewport as it goes.

Think of it like a very sophisticated macro system, but instead of recording your clicks, an AI is generating the commands based on your natural language input.

The AI Side

3D-Agent runs its own multi-agent workflow on leading AI models, with a Blender-aware tool layer. You don't bring your own API key; prompts are included in the plan. The model has context about Blender's capabilities, common 3D modeling patterns, and best practices for clean topology.

When you say "make a medieval tower," it doesn't just create a cylinder and call it a day. It actually understands architectural concepts and applies them.

The Output

Models export to standard formats like OBJ, FBX, GLB, and USDZ. The topology is clean enough to use in production. I've shipped game prototypes with assets generated entirely through prompts.

Why Not Just Use [any_tool_here]?

Fair question. There are other text-to-3D tools out there. I've tried most of them.

The difference is that 3D-Agent works directly inside Blender. No uploading to a web service. No downloading weird file formats and importing them. No waiting in queues.

You stay in your workflow. The AI comes to you, not the other way around.

Also, setup is simple: download the app, open Blender 4.2 or later, sign in and start prompting. No config files, no API keys.

Others generate files full of triangles and millions of vertices, which usually means a bad wireframe and painful topology cleanup.

What It's Good At And What It's Not

Let me be honest because I hate when people oversell their tools.

Works great for:

Concept art and rapid prototyping. If you need to visualize an idea quickly, this is perfect. I've used it to mock up entire game levels in an afternoon. This use case is awesome for AI 3D model generation.

Low to medium poly assets. Characters, buildings, props, environments. Anything where "stylized" or "game ready" is the goal.

Learning 3D modeling concepts. Watching the AI build things actually taught me a lot about how experienced modelers approach problems.

Not ideal for:

Photorealistic organic models. If you need a hyperrealistic human face, you still need a skilled artist.

Extremely precise CAD-style work. This is creative modeling, not engineering.

Replacing an entire art pipeline. It's a tool, not a magic wand. Think of it as having a really fast junior 3D artist on your team.

Let's Build Something Together

Here's a more complex example. I'm generating an entire architectural scene from a single prompt:

The whole thing took about 10 minutes. Try doing that manually as a beginner or with other Blender AI tools. I'll wait.

Want to Try It?

If you're a developer who needs 3D assets but would rather write code than push vertices around:

Start here: 3D-Agent Blender AI plugin

Main site: 3D-Agent

It's a paid app: Basic is $19/month (100 prompts), Pro $29 (200) and Ultra $89 (800). Works with Blender 4.2 and up, on macOS and Windows.

And if you run into issues or have feature requests, come hang out in the Discord. I'm pretty active there and always looking for feedback.

What's Next?

I'm constantly adding new features based on what users actually need. Right now I'm working on:

Better batch generation for creating multiple assets at once
Improved topology optimization for game engines
More export options

If there's something you'd love to see, drop a comment. I actually read them.

So here's my question for you: Have you tried using AI for any part of your creative workflow? What worked? What flopped spectacularly?

I'm genuinely curious because this space is evolving so fast. Let's chat in the comments.

Related searches: Blender AI plugin, Blender AI tools, AI 3D model generator, text to 3D Blender, AI-powered Blender modeling, 3D-Agent.

DEV Community: GL

Seedance 2.5 Camera Control: Block Out AI Film Shots in Blender First

TL;DR: camera control for AI video

Film 1: the Don's office (Blender blockout → Seedance 2.5)

1. The blockout

2. The camera move

3. The blockout drives the AI shot

Film 2: Hope, a cyberpunk AI animated short

Film 3: a wizard-school potions class, with a handheld camera

Why a Blender blockout beats a longer prompt

The workflow, step by step

More shots made this way

FAQ

Try it

Roblox AI in Blender: How We Built the Roblox Backrooms with 3D-Agent and Claude Opus 5.5

First, what is 3D-Agent?

Roblox Backrooms, built in Blender with prompts

Roblox map maker: one screenshot, one prompt

Roblox Blender rigging in 3 prompts

Roblox Studio AI: getting it into Studio

What it's not good at (being honest)

Episode 4 is out

Want to try it?

So happy to have 3D-Agent.com on the Top 7!

Top 7 Featured DEV Posts of the Week

Jess Lee for The DEV Team ・ Dec 2

3D-Agent and Blender-MCP: Both Amazing Blender AI Tools

From Blender-MCP to 3D-Agent: The Evolution of AI-Powered Blender Modeling

GL ・ Dec 1

Blender MCP vs 3D-Agent: Two Ways to Put AI in Blender (2026 Update)

What is Blender MCP?

What is 3D-Agent?

A real example: a spiral staircase

Blender MCP vs 3D-Agent

Which one should you choose?

Try it

Questions?

How I Built Cursor for Blender!

How I Built a Blender AI Plugin That Generates 3D Models from Text: 3D-Agent

GL ・ Nov 26

How I Built a Blender AI Plugin That Generates 3D Models from Text: 3D-Agent

May 2026 Update: Blender AI Is Not a Hypothetical Anymore

The "Aha" Moment

Meet 3D-Agent

How It Actually Works Under the Hood

Why Not Just Use [any_tool_here]?

What It's Good At And What It's Not

Let's Build Something Together

Want to Try It?

What's Next?