Seedance 2.5 Blockout to Film: Rendering White-Model Previz into Final Shots
How Seedance 2.5 turns gray-box previz into finished footage — coarse vs. fine blockouts, both prompt templates, and a fitness-demo worked example with before/after video.
Previz exists because camera language is hard to say in words. A director can storyboard "the camera cranes down as the two groups collide", but a text-to-video model guessing at that spatial logic gets it right sometimes. Seedance 2.5's answer is to accept the previz itself: upload a white-model animation — gray boxes and stand-in geometry from any 3D tool — as a reference video, and the model renders finished footage on top of it, honoring three things text alone can't pin down:
- Camera language — dolly, crane, orbit; the 3D camera path is read from the blockout and reproduced.
- Spatial composition — who occludes whom, foreground vs. background, held as staged.
- Blocking — each stand-in's position, posture and movement path becomes the character's performance.
This is the previz half of the VFX workflow; the keying half is covered in the companion piece on green-screen editing. Both run in Pixo's Playground with Seedance 2.5 — up to 30 seconds at 720p.
Coarse vs. fine: two blockout modes
| Coarse blockout | Fine blockout | |
|---|---|---|
| What you upload | Simple geometry (spheres, cylinders, boxes) standing in for people and props | Complete 3D models with clear structure and materials |
| What it provides | A motion skeleton: trajectories, camera moves, blocking, light changes, cut timing | A finished animation awaiting materials, lighting and style |
| What the model does | Generates a new video on top of the skeleton | Re-renders the model directly into the target look |
| Current strength | Better today — the official guide recommends coarse | Best when the modeling is already done and only the look changes |
The coarse-blockout template
[Reference declaration] Reference @video 1's {camera moves | actions |
movement paths | light changes | blocking | pacing}.
[Mapping] Replace the {color/shape} stand-in in @video 1 with @image X's
{character}; the {color/shape} stand-in with @image Y's {character};
the {color/shape} stand-in with @image Z's {prop / set piece}.
[Plot detail] The story in timeline order: actions, expressions,
dialogue, light changes.
[Scene treatment] Text description, or "use @image W's {scene} as the
background".
[Global wrap] Style, image-quality bar, consistency constraints
("no clipping, no face drift"), audio treatment.The mapping block is the load-bearing part — it's how the model knows the blue cylinder is your elf queen and the gray box is her throne. A minimal real example from the official guide:
Reference @video 1's dynamic camera moves and physical light changes.
Precisely replace the white stand-in in the video with @image 1's elf
queen. Her look must strictly follow the image — long silver hair, an
ornate white-gold embroidered robe, a staff crowned with a glowing
crystal — with full character consistency, no clipping, no face drift.
The scene is the elven great hall; keep @video 1's set as is.Coarse blockouts with limbs or wings need the full limb action sequence written out, or the motion goes stiff. The official guide's best practice is blunter: don't use limbed stand-ins at all — let the mapping block and plot detail carry the performance.
The fine-blockout template
[Render instruction] Render @video 1's white-model animation into a
final film.
[Per-segment description] 0–{N}s: {environment, palette and mood,
character materials, lighting}. At {N}s, {transition trigger and type}:
{the new segment's render description}.
[Scene treatment] Text, or background from @image W.
[Global wrap] Whether character rendering stays fixed or follows the
scene; audio treatment.Before exporting a fine blockout from your 3D tool, strip the trajectory lines, coordinate axes and camera cones from the viewport — the model reads them as picture content and they leak into the render.
Here's the fine template running for real — a wireframe ocean with a red boat marker, rendered into a storm at sea:
Render @video 1's white-model animation into a final film.
0–7s: photoreal ocean night render. Deep gray-blue storm sea, huge waves
rolling continuously, cold low-saturation cinematic palette; waves with
real volume, fluid detail and layered foam. The floating red-and-white
lighthouse keeps its metal and worn-paint materials, pitching and heeling
naturally with the swell. Heavy cloud cover with fine rain; faint distant
lightning as fill. HDR lighting, volumetric fog, real environment
reflections, wet speculars and dynamic shadows — materials and light must
stay continuous through the camera move.
Scene: the storm sea is the only setting; the sea dominates the frame,
horizon and clouds blending naturally. The lighthouse stays near frame
center, driven by the swell — real pitch and heel, no deformation or size
change. Rain, spray and fog share one wind direction and physics; the
camera may bob gently with the waves.
Wrap: keep the scene, render style, lighthouse materials, wave fluids and
lighting consistent. Character rendering: unchanged (no characters).Fine blockout
Rendered result
Note how much of that prompt is material talk — foam layers, worn paint, wet speculars. In the fine mode, motion is already solved; the prompt's whole job is the look.
Create AI videos with Pixo
Turn any idea into a publish-worthy video. One sentence is all it takes.
Worked example: fitness demo, blockout to final
From the official guide's test set — a CG stand-in performing a cable-row sequence (with a glowing form-check effect), rendered into a live-action shot:
Generate a fitness technique explainer: render @video 1. The performer
references @image 1 — strictly execute the movements in the video. Her
face, build and outfit must exactly match @image 1, with no face drift
and no altered movements throughout. Render the scene as @image 2's gym
environment, and keep the glowing effects from the original video.Blockout reference
Rendered result
Watch the ✓ form-check glow: the prompt asked for the original video's effects to survive the render, and they do — same timing, same position, now tracking a real performer. That's the blockout promise in one detail: the motion layer is authored once and kept; only the look is regenerated.
Worked example: one skeleton, four worlds
The most striking case in the official set pushes that promise to its limit. The reference is a single gray mannequin trudging forward against resistance. The prompt renders it into four different films in sequence — a hard cut every five seconds, a new first-frame reference image per segment:
Follow @video 1's motion throughout; fixed camera; no BGM, sound
effects on.
First frame per @image 1 — 0–5s: a heavy vintage copper-helmet diver
drags a glowing yellow-green sample sphere, swimming slowly forward, the
helmet lamp cutting clean light shafts; top right, a giant ice-blue
jellyfish pulses in slow breaths; schools of glowing fish cruise past.
Hard cut. First frame per @image 2 — 6–10s: a lone astronaut in a
weathered suit struggles left-to-right across cracked wasteland, body
pitched deep against hurricane headwind, broken cables and cloth
streaming behind; debris whips past the lens; a vast glowing black-hole
accretion disk turns slowly in the sky.
Hard cut. First frame per @image 3 — 11–15s: every step heavy, cracked
earth kicking up dust, robes and hair blown back; a taut chain drags a
huge teal legendary blade behind, striking sparks and green energy from
the ground; golden sunset backlight, runes glowing in the cracks.
Hard cut. First frame per @image 4 — 16–20s: low-gravity steps raising
slow-settling moon dust; the towed capsule scrapes a gray trail; wrecked
ships on the horizon; a giant Jupiter turning slowly overhead.Mannequin blockout
Four-world render
One motion performance, four productions. This is also the counterexample to the no-limbs rule: a limbed mannequin can work — the prompt earns it by describing each segment's body mechanics ("body pitched deep against the headwind", "low-gravity steps") instead of leaving the limbs to fate. The official verdict: rendered exactly to prompt and references.
Worked example: the minimal version
And at the other end of the effort scale, sometimes the render instruction is one line:
Render @video 1. The walking figure references @image 1; render the scene
as ancient battlefield ruins, backlit; strictly keep the original
video's camera moves.Blockout
Rendered result
White stairs become sun-struck stone; the walk cycle, the staging and the camera survive untouched. Character reference, scene, light, camera constraint — four clauses is a complete render prompt when the blockout carries the rest.
Where this pays off
- Previz to pitch — a director or storyboard artist turns gray-box previz into lit, textured, cinematic frames in one pass; blocking conversations happen on real-looking footage.
- Complex action and staging — multi-character fights, vehicle chases, big spatial moves: anything where camera logic must be exact is authored in 3D and rendered by the model.
The general prompting foundations — character formulas for your mapped-in performers, timestamp scripting for per-segment renders — are in the Seedance 2.5 prompt guide. Reference sweet spots live on the model page.
Try it with a gray box
Export a five-second blockout — even primitive shapes on a camera path — map one character image onto it, and render. The first time the model honors your crane move is the moment this clicks.
Run this on Pixo
- Sign up at pixo.video — new accounts start with 200 free credits.
- Open the Playground and pick Seedance 2.5 as the model.
- Upload your blockout video and character images, paste a render prompt from this post, and generate — up to 30 seconds at 720p.
Generate AI summary
From idea to finished video.
In one conversation.
Start CreatingNo credit card required • Free 200 credits
