3D with Blender
Why a rough 3D scene gives you more control over AI video, the five ways to use it, and when to skip it.
A video model guesses the camera, the set and the blocking from your prompt. A rough 3D scene in Blender lets you decide them instead. You block the shot with boxes and a keyed camera, render it, and give the render to the video model as a reference. Claude Opus 5.5 can build that scene for you through the official Blender connector, so you do not need to know Blender well.
Why 3D helps AI video
- Camera control. Text camera moves are vague. A clay video rendered from a real camera path is exact. ByteDance now documents this officially for Seedance 2.5 as "3D clay model video reference". Creators report the camera is followed closely but not perfectly.
- Consistency. A set or product built once is the same geometry from every angle. Every reference still, location plate, first frame and last frame comes from one source, so shots match.
- Fewer regenerations. A bad blockout costs a re-render on your own computer. A bad clip costs credits. One creator claims blocking first made the work about 3 times cheaper with fewer regenerations, but that post asks readers to comment for a guide and nobody has measured it independently.
The workflows at a glance
| Workflow | What you build in Blender | What the video model gets | Page |
|---|---|---|---|
| Previz | Boxes, capsules and a keyed camera | A clay video for camera, positions and timing | Previz |
| Consistent sets and products | One set or product model and a fixed camera kit | Reference stills, location plates, first and last frames | Consistent scenes |
| Character proxies | A colour coded stand in, or a rigged generated mesh | Colour mapped stand ins, posed renders, turnarounds | Consistent scenes |
| Full 3D | The finished animation, rendered in Eevee or Cycles | Nothing, or the render as input for a restyle | Full 3D |
| three.js instead | A web page that Opus writes, no Blender install | A captured video or stills | Full 3D and Video from code |
Start with previz. It is the cheapest route and the one with official support.
Why Opus 5.5
Artificial Analysis launched AA-3D on 12 Oct 2026. Each model works as an agent in an offline sandbox and builds a three.js scene from one of 55 briefs, and people vote blind on the results. On the board fetched on 15 Oct 2026:
| Model | Score | Note |
|---|---|---|
| Sonnet 5.5 | 1194 (plus or minus 34) | Tied at the top |
| Opus 5.5 | 1190 (plus or minus 35) | Tied at the top, ranked 1 to 3 |
| GPT-6 Astra | 1165 | Tied at the top |
| Fable 5.1 | 1131 | |
| Gemini 3.8 Flash | 965 | Best Gemini on the board |
Opus 5.5's 95% interval runs from 1155 to 1225 and includes both Sonnet 5.5 and GPT-6 Astra, so treat the three as tied. Two limits matter. The sandbox has no GPU and the output is three.js, so AA-3D measures the code route more than Blender work. And Gemini 4 Argon is not on the board, because it rolled out around 13 Oct 2026, after the snapshot.
Every Argon versus Opus comparison so far is an anecdote from X. In those tests Argon is much cheaper and faster (for example 14 minutes and 0.95 USD against 48 minutes and 12.51 USD for the same scene), while most testers still prefer Opus for taste, lighting and motion. A few identical "Argon wins" posts look like promotion. No blind Blender comparison exists yet. Use Opus 5.5 for scenes you will keep and try Argon for cheap drafts.
When not to use Blender
- A single shot with a simple camera. Text or image to video is enough. A storyboard sheet covers most beginner projects.
- Faces and close up acting. Faces are the weakest part of any 3D route. Keep faces with the video model and your image references.
- Loose, organic motion. Video models copy stiff blockout animation and camera easing literally. If you want the model to invent the movement, a blockout fights it.
- A budget of around 25 USD. One Opus 5.5 Blender scene costs from a few dollars to tens of dollars in tokens, before any video credits. See Time and cost.
- A computer with sensitive data and no spare machine. The connector runs AI written Python in Blender with no real sandbox. See Setup.
Start small
Block one shot of 5 to 8 seconds, render it, and run one cheap 480p draft. If the camera is followed well enough for your project, build the next shot. If not, a storyboard and image references may be all you need.
Pages in this section
Setup
Blender 5.1 or newer, the official connector, Claude Desktop or Claude Code, and the traps that stop it connecting.
Previz
Block the camera with simple shapes, render a clay video and let Seedance 2.5 follow it.
Consistent scenes
Build a set, product or character proxy once and render every angle and frame pair from it.
Full 3D
Render the whole shot in Blender, restyle it with video to video, or use three.js instead.
Prompts
A rules prompt and six task prompts for Opus, with checks for every step.
Hand off
Render settings for guide frames and what each video model accepts.
Troubleshooting
Failure modes, time and cost, and which posts to trust.
Sources
- Seedance 2.5 prompt guide (3D clay model video reference)
- Artificial Analysis AA-3D launch article and leaderboard
- Gemini 4 Argon model page
- Blender Lab MCP server
- Argon versus Opus tests on X: notjazii, Avenoxai, xikhar
- Previz cost claim: pounddz (organic, comment funnel)