A grey playblast, rendered without a renderer
A 12-second animation shot that starts as a raw untextured playblast: correct performance, correct camera, no lighting, no materials, no set dressing. Nothing here is re-animated. The playblast keeps the acting and the camera exactly as blocked, and everything a render farm would normally spend hours on is generated as reference and applied on top. Built in Studio, so the whole node graph is open to read.
Stage 1 - decide the look on two boards
The look is decided before a single frame of video is generated. One frame is pulled from the playblast and turned into two boards: a character sheet and the same room re-rendered empty. Boards are cheap to judge and cheap to redo, and once approved they hold the character and the set still for the whole shot.
Stage 2 - let the animation drive the render
The shot is described in plain language, a language model turns those notes into a render prompt under strict written rules, and the playblast itself drives the render. The performance and the camera come from the animation. The character, the set and the light come from the boards.
Step by step
- 01
Start from the playblast, not from a description
The source is a plain viewport playblast: grey shading, flat light, no textures. It is also the only part of this shot that a human animated, and it is the part that has to survive untouched. Everything downstream is arranged so that the acting and the camera are inputs, never things the generator gets a vote on.
The playblast: right performance, no render. - 02
Pull one frame and build a character sheet from it
A single frame is extracted and used as the identity reference for a four-view character sheet: front, three-quarter, profile and a close portrait. Asking for a sheet rather than a pretty picture is the point. Four angles of the same face is what stops the character drifting once the frame count starts climbing.
The prompt for this stepShow promptHide
Using the attached CG playblast frame as the exact character reference, create a multi-angle character sheet of this girl rendered at polished animated-feature quality. Same character, same proportions, same outfit: round glasses, long brown hair, green-sleeved cream raglan tee with red letters. Four views on one sheet: front, three-quarter, profile, and a close-up portrait. Feature-film 3D render: subsurface-scattered skin, detailed eyes, individual hair strands, woven fabric texture, soft studio key light, neutral grey background.
The character board, generated from one playblast frame. - 03
Re-render the room from the same frame, with nobody in it
The same extracted frame drives a second board, this time as a layout reference: keep the counter, the tables, the stools, the door and the window exactly where they are, and render the room at feature quality with the character removed. Separating the set from the character means each can be judged, and fixed, without disturbing the other.
The prompt for this stepShow promptHide
Using the attached CG playblast frame as the exact layout reference, re-render this cafe interior at polished animated-feature quality. Keep the same layout: wooden bar counter in the foreground, round cafe tables with pink stools, white-paneled door and window at left, soft pink walls, wine glasses on the counter. Feature-film 3D render: warm afternoon window light, soft bounce fill, shallow depth of field, physically based materials, gentle film grain. No characters in the frame.
The environment board: same layout, real light, empty. - 04
Write the shot down in plain language first
Before any prompt engineering, the shot is described the way it would be described in dailies: who is in frame, what they do and in what order, what the camera is doing, what the room feels like. Short, specific, and free of any render vocabulary. This is the human judgement in the pipeline, and it is deliberately kept separate from the prompt that gets sent.
The prompt for this stepShow promptHide
Shot: a girl with round glasses and long brown hair sits at a cafe bar counter with a glass of red wine. She grins broadly and leans forward, rests her hand on her face in a bashful beat, then settles back with a content smile, hands near the glass on the counter. Single locked-off medium shot, eye level, 24fps. Cafe interior with round tables, pink stools, and soft window light.
- 05
Have a language model write the render prompt, under rules
The notes are expanded into the render prompt by a language model working from a fixed instruction set: name the subject plainly, specify the render quality, specify the lighting, and state twice over that motion and camera follow the reference video while character and set follow the reference images. The instructions also forbid camera moves, since a locked-off shot that gets described as drifting will start drifting. Writing the rules once and reusing them is what makes the result repeatable rather than lucky.
The prompt for this stepShow promptHide
You write render prompts for Seedance 2.0 reference-conditioned video generation. The motion, framing, camera, and timing are locked by a reference video; reference images define the character and environment look. Rewrite the user shot notes into one flowing render prompt that: (1) names the subject and action plainly, (2) specifies polished 3D animated-feature render quality: subsurface skin, strand-level hair, woven fabric, physically based materials, (3) specifies cinematic cafe lighting: warm window key, soft bounce, shallow depth of field, (4) states that motion, camera, and timing must follow the reference video exactly, and (5) states the character and set must match the reference images. No camera moves that contradict a locked-off shot. No style names of other studios. Under 1500 characters. Return only the prompt.
- 06
Render with the animation in the driving seat
The playblast goes in as the reference video, the two boards go in as reference images, and the written prompt goes in as the description. What comes out has the same performance, the same timing and the same locked framing as the blocking, now carrying skin, hair, fabric, glass and window light. The lesson for a comp department is that this output is an element, not a delivery: it still wants grain, grade and a real edge treatment before it belongs anywhere.
The prompt for this stepShow promptHide
A 3D animated-feature render of a young girl with round glasses and long brown hair sitting at a wooden cafe bar counter next to a glass of red wine. In a single locked-off medium shot at eye level, she grins broadly, leans forward to rest her cheek on her hand bashfully, and then settles back with a content, warm smile. The character design, facial features, clothing, and the cafe set - complete with round tables, pink stools, and background details - must exactly match the reference images. The character's performance, facial expressions, gestures, and the locked camera framing must perfectly track the reference video. The scene is rendered in premium animated-feature quality, featuring realistic subsurface scattering on her skin, individual strand-level hair dynamics, woven fabric textures on her shirt, and physically based materials for the glass, wine, and wooden counter. The lighting is cinematic and cozy, with a warm key light streaming from a side window, soft ambient bounce, and a shallow depth of field that gently blurs the cafe interior behind her.
The rendered 12-second shot.
How it was wired
Wired in Studio as: the playblast to the reference video input, the character board and the environment board to the two reference image inputs, the enhanced prompt to the prompt input, 12 seconds, 720p, 16:9. Open the graph to see every node with its real output still attached.