Blender, Opus 5.5 and Seedance: a workflow for the one-take

I wanted to test a workflow: can a video model be guided by a scene from Blender so that a camera moves precisely through a room for a full minute, without a single cut? No video model manages that from text alone. So we build the scene in Blender first, driven by Claude, and give the render to Seedance as a motion reference.

The test needed a scene that could carry a camera like that: an interrogation. A woman has killed her violent husband, a police commissioner, a big name in the building. The detective questioning her understands her. And she knows: she was on his list too. The camera follows the drama of this dialogue, in the end as one single continuous shot of 60 seconds.

The result first. On top, the film from Seedance; below, the Blender render it moves by.

The ingredients

• Nano Banana 2 for the character sheets of the two women (front, back, close-up), three set images of the interrogation room and a prop sheet for the dictaphone.
• Blender 4.4, built by Claude. Claude Opus 5.5 works in Claude Code and is connected straight to Blender via MCP. Claude builds the room, the furniture, the blocking of the characters, the camera move and the renders, with its own Python scripts for Steadicam, window and prop.
• Seedance 2.5 in BOFF.create, our own wrapper, for the actual video, with the Blender render as motion reference.
• ElevenLabs for fine-tuning the voices.

From room to continuous shot

1. Rebuild the room. From the reference images, Claude builds the interrogation room in Blender: one-way mirror, door, frosted window, table, pendant lamp and the observation room behind the glass.
2. A first version with cuts. Nine classic set-ups: the view through the glass, top-down, over-the-shoulder, close-ups, a dolly zoom, an orbit that crosses the line. The guiding idea: the mirror stands for the police apparatus. As soon as the detective stops the recording, the camera jumps the line and the mirror disappears from the frame.
3. Condense to 60 seconds. Seedance generates at most 30 seconds in one go, so two blocks of 30 seconds.
4. Rebuild as a one-take. One continuous camera. The seam between the two generations hides in a whip pan; the motion blur covers the transition, as in «Rope» or «Birdman».
5. Make the camera feel real. A single 32 mm lens, as on a Steadicam, no zoom. High mode instead of a ceiling camera, framing on the thirds with lead room. A Gaussian filter smooths the whole camera path organically; only the whip pan stays hard.
6. A prop as story. Instead of a microphone, the detective pulls a micro-cassette dictaphone from her trouser pocket and presses REC; the red light comes on. In the second block she presses STOP: the moment she changes sides.
7. Write the prompts. In English, with fixed blocks: references, cast, motion, room geometry, prop, camera, dialogue, acting, sound, light, negative. Acting restrained and naturalistic, sound only dialogue and diegetic noise, no music.

It took eight versions of the Blender pass, v1 to v8, until the motion held.

The render only provides the scaffold: who stands, walks or sits where. Seedance plays the rest.

What we learned

• Scaffold, not acting. The render sets the room's dimensions, the camera move and the positions. Head turns, nods, hands and facial expression come from the prompt, via Seedance.
• The more abstract the figures, the more natural the result. With jointed mannequins, Seedance copied stiff steps and a robotic sit-down one to one. With simple columns, no arms or legs, it invents real movement.
• Render reduced. Workbench, grey tones, no faces, no blur. That way the look of the render doesn't seep into the video.
• Define the walls clearly. Seedance ignores grey walls and invents windows, doors or even a corridor. The fix is colour-coded architecture: window bright, door dark brown in a light frame, mirror as dark glass in a frame, walls with a dark skirting. And the prompt says that the video defines the room as well.
• References have to match. A single set image with the door on the wrong wall threw the geometry off. Better to leave it out.
• Props separately. If an object can't be cleanly removed from the set images, empty the set images and supply the object as a prop sheet of its own.
• Word choice matters. Even «corridor murmur» in the sound block can put a corridor in the frame.
• Hide the seams. Don't solve transitions between two generations with start and end frames; hide them in a movement.
• Know the limits. Reference videos at most 30 seconds in total, so no voice samples as video. Prompts can be up to 20,000 characters long.

The result

The very first run in Seedance already delivered a lot: characters, light, acting, the dictaphone and the seam in the whip pan all held. The weak spot was the room geometry, with window and door in the wrong places. Version v8 answers that with colour-coded architecture and more consistent references.

Eight versions of the Blender pass, two generations of 30 seconds each at 24 fps and 16:9. One camera, one lens, no cut.

The Blender skill to copy

This is how we have Claude build the motion pass in Blender. The skill sums up the rules we worked out over eight versions: room, placeholders, camera, seams and render. It works with Claude Code and an MCP connection to Blender.

name: blender-motion-pass
description: Use when a scene for a video model (Seedance 2.5, Omni-Reference) needs a motion reference: build the room, the blocking and the camera move in Blender via MCP and render a reduced motion pass.

# Blender motion pass for Seedance

The render is a scaffold, not a performance. It defines only where and when: room dimensions, camera path, who stands, walks or sits where. Head turns, hands and facial expression come from the video prompt.

## 1. Room
- Rebuild the set from the reference images, true to scale: walls, doors, windows, mirrors, furniture, lights.
- Use only reference images that agree spatially. One image with a door on the wrong wall breaks the geometry; leave it out.
- Colour-code the architecture so the video model cannot misread it: window bright, door dark brown in a light frame, mirror as dark glass in a frame, walls with a dark skirting. Plain grey walls get ignored and replaced with invented windows, doors or corridors.
- Props that cannot be removed cleanly from the set images: empty the set images and give the object a prop sheet of its own.

## 2. Characters
- Placeholders, not mannequins: simple columns without arms or legs, one colour per character taken from the costume, a sphere as head in the hair colour.
- Jointed figures get copied one to one: stiff steps, robotic sitting. Abstract columns let the video model invent real movement.

## 3. Camera
- One lens for the whole take, e.g. 32 mm as on a Steadicam. No zoom: closer means moving closer.
- High mode instead of a ceiling camera. Framing on the thirds, with lead room toward the gaze.
- Smooth the whole camera path with a Gaussian filter for an organic operator feel. Keep only intended fast moves, such as a whip pan, hard.

## 4. Length and seams
- The video model generates at most 30 s per block: split the take into blocks of 30 seconds.
- Hide every seam inside a movement, e.g. a whip pan with motion blur. Never join blocks on start or end frames.

## 5. Render
- Workbench, grey tones, no faces, no depth of field, no blur. Nothing of the render's look may leak into the video.
- One clip per block, frame range in the file name (e.g. Block_A_0001-0720), 24 fps, 16:9.

## 6. Hand-over to the prompt
- List which placeholder colour is which character, and which colours mean window, door and mirror.
- State that the video defines the room as well as the motion, and that bodies must not copy the gliding placeholders.
- Iterate the pass, not the video: check each version against the references before generating.

What this changes

Here, Blender isn't an animation tool but the director's plan. Camera, blocking and timing are fixed before a single frame is generated. The video model plays the scene, but it no longer decides where the camera stands.