🎬 Seedance 2.5 Previz Workflow: 3D Blockouts to Finished Shots

Seedance 2.5's headline features were native 4K and 30-second takes. The one that changed how we actually work is quieter: you can hand it a rough 3D blockout and it will treat that geometry as a spatial scaffold for the shot. After a year of producing AI video where "camera control" meant writing adjectives and hoping, this is the first tool that makes previz mean what it means on a real set.

By the numbers

Capability	Seedance 2.5
Released	Early July 2026
Max continuous take	30 seconds
Max resolution	Native 4K
Reference slots	Up to 50 multimodal
3D previz input	Lightweight OBJ / FBX blockout
Source tools	Blender, C4D, Maya, SketchUp
Audio	Included in base rate

Why a 30-second take changes the economics

When clips were five seconds, re-rolling was the workflow. You generated eight variations, picked one, moved on. Nobody planned, because planning cost more than regenerating.

At 30 seconds continuous, that inverts. A single take is now a meaningful commitment of time and credits, and the failure modes get worse — a camera move that drifts off-axis at second 22 wastes the whole generation, not a fifth of it. Which means the discipline that governs live-action production suddenly pays here too: decide the shot before you shoot it.

That's what the blockout feature is for. It moves the expensive decisions — composition, camera path, spatial relationships — to a stage where fixing them costs you a few minutes of dragging boxes around.

The workflow

1. Block out the space, not the objects. Open Blender (or whatever you know) and place primitives. A box for the car, a cylinder for the actor, a plane for the ground, walls where walls go. Grey, untextured, no detail. You are defining volumes and their relationships — where things are relative to each other and to the camera.

2. Set the camera move. This is the part worth spending time on. Animate a camera through your blockout: the push-in, the arc, the crane down. Play it back. If the move feels wrong on grey boxes, it will feel wrong photoreal — and you've now learned that for free.

3. Export lightweight. OBJ or FBX, low poly. Resist the urge to add detail; the model needs the scaffold, not your subdivision surfaces. Heavy files buy you nothing here.

4. Attach it as a reference, plus a style reference. The blockout goes into one reference slot; a style frame — a still that carries the grade, lens character and lighting you want — goes into another. With up to 50 multimodal slots you also have room for character identity references, which is how you keep faces consistent across a sequence. Our consistent characters guide covers that side in depth.

5. Prompt for what the geometry can't say. The blockout handles space and camera. Your text prompt should carry everything geometry cannot: time of day, weather, emotional register, material qualities, wardrobe. Don't re-describe the camera move in words — you've already specified it, and redundant instructions fight each other. Our Seedance 2.5 prompt library has the phrasing patterns that survive this handoff.

6. Generate at low resolution first. Confirm the model respected your blockout before committing to a 4K render. If subjects are ignoring your volumes, the fix is in the blockout, not the prompt.

What the blockout actually fixes

Three specific problems, from our own production logs:

Camera drift. The single most common re-roll cause. Text-specified moves are suggestions; geometry-tracked moves are repeatable. If you need the same push-in across six shots in a sequence, this is the only reliable way to get it.

Scale collapse. Models are notoriously bad at relative size — a small character next to a large one drifts toward matching heights over a take. Volumes in a blockout are unambiguous. A cylinder that's half the height of another cylinder stays half the height.

Axis breaks. If your character is screen-left in shot A, keeping them screen-left in shot B is a continuity rule the model doesn't know. Laying out both shots against the same blockout makes the 180-degree line a property of the scene rather than something you have to remember.

Where it doesn't help

• Performance and micro-expression. Geometry says nothing about acting. That's still prompt and reference work.
• Fast, chaotic action. Blockouts stabilise deliberate camera work. Whip pans and handheld chaos are better handled with prompt language and cut rhythm.
• One-off shots. If you need a single clip and don't care about matching anything, the blockout overhead isn't worth it. Just generate.
• Anyone without 3D basics. Be honest about this. If opening Blender is a wall, use a storyboard grid instead — a 3×3 sheet of your nine key frames gets you a lot of the compositional discipline without the software.

How we tested this

We ran the blockout path against our normal prompt-only workflow on the same shot briefs — a character walking a defined path through a defined space with a specified camera move — and counted re-rolls to an acceptable take. The pattern we care about isn't a benchmark number; it's where the failures moved. With blockouts, failures shifted from "camera did something else" to "performance wasn't right," which is a much cheaper category of problem because it doesn't require rebuilding the shot. Figures on model capability are ByteDance's published specs as of the early-July release; treat pricing and tier details as moving targets.

Pros and cons

Strong for: sequences that need continuity, product and architectural visualisation, anything where a client approved a specific camera move, and multi-shot narrative work where spatial consistency across cuts matters.

Weak for: speed. This is a slower front end than typing a prompt. It pays off across a sequence and costs you time on a single clip.

The broader point: AI video is drifting toward the workflows film production already solved. Previz, style frames, continuity ledgers — these aren't legacy overhead, they're what you invent when takes get expensive. Seedance 2.5 is the first model where that convergence is explicit in the tooling. If you're weighing it against the alternatives for a specific project, our Kling 3.0 comparison covers where each one earns its render cost, and the full Seedance 2.5 guide has the model's other capabilities.

ByteDance's model documentation is at seed.bytedance.com.