What Is the Seedance 2.5 Blender Workflow? Four Steps From White Model to Take

TL;DR — The Seedance 2.5 Blender workflow uses a white model in four steps to direct camera paths and timing across 4 to 30 seconds of video. It gives the model a visual reference clip to follow, which a text prompt can't describe. If a take is wrong, change the white model geometry for motion errors or change the prompt for surface mistakes.

You keyframed a dolly through a doorway in Blender, described the shot in a prompt, and got a different room and a different move. You hit re-roll, and the whole set changed again.

A white model is what you reach for, because it's the one thing you can hand over that a sentence can't carry. It tells the generator where the camera starts, where it ends, and what's in front of what.

Here's how the four steps work, what each one controls, and what came back from three of our own paid runs, including the parts that missed.

What the Seedance 2.5 Blender workflow is

In run 57d24591, we put simple grey figures inside an untextured boxing ring in Blender. The model returned fighters with faces, trunks, gloves, and the venue and its lighting.

A white model is untextured geometry: boxes, cylinders and rough shapes standing in for the real set and the real actors. Most creators call the same thing a blockout.

We checked, and no Blender build of Seedance exists. As Toonkit noted on 9 September 2026: "There is no Blender-specific version released by ByteDance. It is simply what creators call the production method of using a Blender 3D blockout as a reference."

Left is the Blender clay pass; right is the finished take from Seedance 2.0 Mini on Seadanse at 480p and 10 s. The video was made from a 1280 × 720 clay pass at 24 fps (run 57d24591, 7 September 2026).

The four steps, and what each one controls

Each step owns a distinct part of the shot.

Build the white model in Blender

You set up your layout using basic, untextured shapes to stand in for your actors and environment.

You position shapes to block the action, set your focal length, and keyframe your camera along the path you want.

Render the clay pass

You write out a clay pass (a flat grey render of your scene, lit but untextured) from your active camera as an H.264 mp4 file.

You keep studio lighting and cast shadows in the shot so structure and camera movement stay readable.

The Seadanse add-on runs in Blender 5.2.1 LTS on macOS (10 September 2026), installed from a zip. It shows the finished clay pass and the server quote on the Generate button before anything is submitted. Account details are blurred.

Attach the clay pass as your reference video

You feed the clip into the generator as a motion reference input.

You can upload the mp4 manually through your browser, or submit it straight from Blender using the Seedance 2.5 Blender add-on. The add-on makes the clay pass in your sidebar, uploads it, and imports the finished mp4 back into your scene on a plane.

Write the prompt and generate

You describe the textures, lighting, materials, and character details you want applied over the grey shapes.

You can attach still appearance images to fix costumes, faces, or architectural styles. Check our Seedance 2.5 prompt guide for ways to balance text descriptions against motion guides.

Here's what owns what in the final take:

• Geometry and camera: These own the path, the framing you intend, and the timing.
• The prompt: Text descriptions own surfaces, light, character faces, and clothing.
• Appearance reference images: Still images own the visual look and material palettes.
• Contact and weight: Nothing owns these. Gloves still went through each other on blocks.

Try the Seedance 2.5 Blender workflow

You can run this process in a browser, or from inside Blender.

We give you two free downloads with published SHA-256 checksums: a Blender add-on and an agent skill. The add-on is a self-hosted zip, not a listing on Blender's extensions platform.

You can test reference clips directly in the Seedance 2.5 generator to see how geometry directs motion.

Check what Seedance 2.5 costs to run across different durations before running larger jobs. New accounts get free starter credits on sign-up, with no card needed.

The Seadanse web composer on 10 September 2026 with Seedance 2.5 picked and 57d24591-clay.mp4 attached as the reference clip before generating.

Why the clay pass beats a prompt

A text prompt can't say where the camera starts, where it ends, or what blocks what.

Using a visual guide lets you keep the camera move you keyframed without fighting prompt drift.

ByteDance's Dreamina motion reference guide noted on 9 September 2026 when a white model reference beats texture: "White model references are useful when space, camera path, and relative position matter more than final texture."

We tested camera tracking in run b2b93be2. The reference clip flew past a windmill, turned toward a castle, passed a tower, and rose to an overhead view.

The video kept that camera order from start to finish. A single appearance image gave the fields and stone their materials and light.

The castle close-up still read as CG, and local composition, proportion and detail drifted. The camera follow passed; the live-action look didn't.

Left is a cropped public Blender screen recording with audio stripped, used as the reference clip. Right is the 14 s, 480p take generated on Seedance 2.0 Mini on Seadanse. It used a single appearance image for materials and lighting (run b2b93be2, 7 September 2026).

What Seedance 2.5 accepts on Seadanse

Seedance 2.5 on Seadanse has a fixed set of limits, on top of what Seedance 2.5 is in general:

• Resolution: Output can be 480p, 720p, or 1080p. There's no 4K on Seedance 2.5 here.
• Web generator durations: You can pick 5, 10, 15, 20, 25, or 30 seconds.
• Add-on durations: The add-on supports 4 to 30 whole seconds on Seedance 2.5, and 4 to 15 whole seconds on Seedance 2.0 Mini at 480p.
• Aspect ratios: Exactly six options — 16:9, 9:16, 1:1, 4:3, 3:4, and 21:9.
• Audio defaults: Audio is on by default.
• Reference clips on Seedance 2.5: Up to 10 clips, each 2 to 30 seconds, and under 30 seconds combined.
• Reference clips on Seedance 2.0 Mini: Up to 3 clips, each 2 to 15 seconds, and under 15 seconds combined.
• Upload limits: The add-on allows 100 MiB per clip on Seedance 2.5 and 50 MiB on Seedance 2.0 Mini.
• Appearance images on Seedance 2.5: Up to 30 PNG, JPG, or WebP files, at up to 30 MB each. They never change what the job costs.
• Appearance images on Seedance 2.0 Mini: Up to 9 images. They don't change what the job costs either.
• Reference audio: Reference audio isn't wired for Seedance on Seadanse, so you must describe voices instead of cloning them.
• File privacy: The model reads video files only, and your .blend file never leaves your computer.
• Prompt capacity: The composer box takes up to 10,000 characters.
• Upstream constraints: The published schema for Seedance 2.5 reference-to-video accepts clips with 24 to 60 frames a second and sizes from 300 to 6,000 pixels on a side.
• Frame rate: A playblast written at 12 fps sits under the 24 fps floor. Check yours before you upload.
• Reference cost structure: Attaching video references adds a 0.6 multiplier to the base rate for the input duration, which you can check on our Seedance 2.5 cost calculator.

Four ways to run Seedance 2.5 from Blender, compared

There are four routes from Blender to Seedance 2.5, and our Seadanse AI video tools are one of them. We broke down how to access Seedance 2.5 previously, but here's every way to run Seedance 2.5 from Blender, compared side by side.

Dreamina's documentation on 9 September 2026 says you "continue your Seedance 2.5 video workflow in Dreamina", which means the generation happens in your browser.

Higgsfield's published plugin requirements on 9 September 2026 stop at Blender 4.2 through 5.1 and reserve full model access for higher plans. Mixar is a 3D editor built on Blender rather than an add-on for it, so using it means leaving your own Blender.

A dash (—) would mean we could not verify the cell; there are none in this table. Every row was read or run on 2026-09-09.

What this changes for previz, walkthroughs and action beats

In run 0c29a9f2, we animated two cars moving 36 metres in 5 seconds while holding 11 metres apart.

The take on Seedance 2.0 Mini kept the red sports car front-right, the black SUV rear-left, and preserved the gap and background parallax. It left grey clay facades in the background until we corrected the prompt.

In run 5eced261, we pushed a camera through a 4-metre doorway into a living room.

The move was continuous, the door frame left the frame naturally, and we saw no wall clipping, cuts, or floating furniture.

In runs 57d24591 and ed811a27, our agent skill drove Blender over MCP to choreograph an action sequence.

The third version of our clay pass held 24 punch events and one kick across ten real seconds. The generated video kept the short punches, recoveries, and fast same-side camera cuts.

What still misses

The output came back framed tighter than the clay pass in both runs, cropping the fighters' feet.

Fast contact also showed clipping: boxing gloves interpenetrated during blocks, and moving from Seedance 2.0 Mini to Seedance 2.5 didn't solve it.

Model choice changed how some actions resolved. On Seedance 2.0 Mini, a kick read as a knee lift and the model added shoes despite a barefoot prompt direction.

On Seedance 2.5 in run ed811a27, the kick extended further and the barefoot direction survived.

Body weight and repetitive on-the-spot punches didn't improve. One paired run doesn't tell you Seedance 2.5 is generally better than Seedance 2.0 Mini.

Choreography takes real prep work in Blender before you make your clay pass. Our first try felt dull, with half-second extensions and flat camera angles.

We had to run a correction pass before the action felt intense. We planned strikes and recoveries separately, answered every attack with a hit or block, and built fast camera moves around events.

At 4.71 seconds, Seedance 2.5 on the right (run ed811a27) holds the extended front kick on a bare foot. Seedance 2.0 Mini on the left (run 57d24591) has already placed the foot down in a red boot. Both 10 s, 480p takes ran on Seadanse on 7 September 2026 using the exact same clay pass and prompt.

You're directing the model, not driving its camera.

In our early testing, two runs came back marked success, were billed, and ignored the reference clip entirely because of an internal field mismatch that's now fixed. That's what to check when Seedance 2.5 ignores your white model.

Seedance 2.5 Blender workflow FAQ

Do you need a plugin, or can you just export and upload?

You don't need a plugin. You can write out an H.264 mp4 yourself, open the Seedance 2.5 video generator in a browser, attach the file, and generate the shot.

Does the .blend file go to Seedance 2.5?

No. The model only reads video files like MP4 or MOV clips. Your .blend file stays on your machine.

Which shot works best — an orbit, a dolly or a walk-through?

We haven't ranked shot types against each other. In our testing, we verified four distinct moves:

• A flight past a windmill and castle in run b2b93be2.
• A push through a doorway in run 5eced261.
• A tracking shot of two moving cars in run 0c29a9f2.
• A fast boxing sequence in run 57d24591.

Does the take come back as geometry you can edit?

No. The generator returns a flat video file. It doesn't export 3D meshes, camera curves, or textured objects back into Blender.

The verdict

The Seedance 2.5 add-on for Blender takes the camera move you built in Blender and hands it to the model. It gives animators, previz artists, and directors a practical way to direct camera speed and scene timing before generating takes.

This process is for creators who want specific control over camera moves and scene staging in 3D. If you don't know how to set keyframes in 3D software and just want quick AI video clips, typing a prompt is faster.

We ran this on real Blender against the real model in September 2026, we paid for every take, and we kept the ones that missed.

Related articles

• Seedance 2.5 Not Following Your White Model — the failure where the take succeeds, gets billed, and ignores your clip.
• Ways to Run Seedance 2.5 From Blender — every route in one list, with what each one installs.