| name | spark-video-cast |
| description | Scaffold and generate reference assets for characters (cast), locations (movie-set / set dressing), and key props — the three pillars of visual consistency in spark-video. Wraps bl image generate / edit for cast reference sheet creation. Use when adding new characters/locations/props or when costume/state changes are needed. |
Cast / Set / Prop Skill — spark-video Art Department (all-in-one)
You are the art department of the pipeline. Your job is to scaffold
folder structures and generate reference images for the three things
that pin visual consistency:
| Pillar | Pins | Folder pattern |
|---|
| Cast | Full-body reference sheet: face, hairstyle, costume, build | cast/<name>/ |
| Movie-set | Locations, lighting, decor | movie-set/<name>/ |
| Prop | Hero objects that recur or change state | props/<name>/ |
All three follow the same mental model: one folder = one reference
image = one frozen visual state. State changes (day→night, intact→torn,
casual→formal) = separate folders.
Set env vars:
export SPARK_VIDEO_PROJECT=<project_id>
export SPARK_VIDEO_EPISODE=<NN>
export SPARK_VIDEO_PHASE=cast-reference
Two-tier model — project vs episode
Both cast, set, and prop live under two tiers. Episode tier overrides
project tier on name collision:
projects/<p>/
├── cast/<name>/ ← project mains (shared across all episodes)
├── movie-set/<name>/ ← project recurring locations (sitcom rooms)
├── props/<name>/ ← project recurring hero objects
└── <episode>/
├── cast/<name>/ ← episode NPCs OR project-cast overrides (fork)
├── movie-set/<name>/ ← one-off locations for this episode
└── props/<name>/ ← one-off or state-overrides for this episode
Use the project tier when an asset is shared across episodes
(sitcom recurring rooms, series mains). Use the episode tier for
one-off NPCs / locations / state-changes (episode-wide costume forks, episode-only
hero items, one-off rooms).
⚠ THE ONE-FOLDER-ONE-STATE RULE (hard rule, applies to all 3)
The video model reads reference images literally. Mixing two
visual states into one folder produces a muddy averaged intermediate.
| Pillar | "Same X, different…" → separate folder |
|---|
| Cast | Episode-wide costume change (wedding dress / battle wounds / period vs modern) → fork into episode tier |
| Set | Time-of-day (day/night), season (spring/autumn), color grade (cool/warm), weather (clear/rain) |
| Prop | State (intact/creased/torn), damage (clean/bloodied), open/closed |
Naming convention: <base_name>-<discriminator>:
同福客栈大堂-白天 / 同福客栈大堂-夜晚红包-完整 / 红包-起皱 / 红包-撕碎陆辰-汉服 (forked from 陆辰 for one episode)
Procedure 1 — scaffold a cast
Cast reference image contract
Generated cast images should be character reference sheets, not
single photoreal portraits. Seedance-style providers are less likely to
reject an obviously synthetic character sheet than a realistic headshot.
Every cast-image generation prompt MUST include this intent:
a three-view drawing of a person and annotate it with an AI-generated, non-real-person watermark on the bottom
Make the sheet show the same character in front / side / back views,
with consistent face, hair, costume, build, and a plain background. The
bottom annotation is part of the image content for safety provenance;
do not use provider watermark flags. Avoid any other readable text.
1.1 Lead / project-tier character
uv run scripts/scaffold.py cast --name "陆辰"
Then generate the cast reference sheet via bl. Default to a full-body
standing character sheet / 三视图 rather than a face-only or front-only
portrait. Keep the portrait filename prefix for compatibility with the
rest of the pipeline:
./scripts/bl image generate \
--model wan2.6-t2i \
--prompt "a three-view drawing of a person and annotate it with an AI-generated, non-real-person watermark on the bottom, 28-year-old man, short hair, dark T-shirt, full-body character reference sheet, front view, side view, back view, same face and costume in all views, plain background, no other readable text, $(uv run scripts/scaffold.py mood-anchor)" \
--size 16:9 \
--out-dir projects/$SPARK_VIDEO_PROJECT/cast/陆辰/ \
--out-prefix portrait
Notes:
- Default model
wan2.6-t2i: produces stable cast reference sheets
compatible with downstream r2v. qwen-image-2.0 is newer but visual
style differs; test before switching.
- Append
lore.mood_anchor to every cast-reference prompt so the visual
style matches the rest of the production. The scaffold.py mood-anchor
helper prints lore's mood_anchor for piping.
- Prefer generated / hand-drawn character sheets over real-person
photos for providers with input privacy checks. If you drop a real
actor reference into the folder, make sure you have the right to use it
and that your target provider accepts real-person inputs; it overrides
the generated reference sheet at r2v time.
Optional: voice reference for reference-voice r2v (Wan / bl both support):
- Drop a 5–10s clean speech sample as
voice.mp3 in the cast folder.
1.2 NPC (episode-only)
uv run scripts/scaffold.py cast --name "钱夫人" --episode
./scripts/bl image generate \
--model wan2.6-t2i \
--prompt "a three-view drawing of a person and annotate it with an AI-generated, non-real-person watermark on the bottom, middle-aged woman, stout build, dark silk hanfu, gold hairpin, shrewd worldly expression, full-body character reference sheet, front view, side view, back view, same face and costume in all views, plain background, no other readable text, $(uv run scripts/scaffold.py mood-anchor)" \
--size 16:9 \
--out-dir projects/$SPARK_VIDEO_PROJECT/episode-$SPARK_VIDEO_EPISODE/cast/钱夫人/ \
--out-prefix portrait
Then re-init the merged cast.json:
uv run scripts/scaffold.py cast-init
1.3 Cast fork — episode-wide costume change
When a character needs a different outfit for THIS episode only (wedding,
period costume, battle-damaged version), DO NOT solve it in shot prompts. Fork the cast reference:
uv run scripts/scaffold.py cast --fork --name "陆辰" --drop-portraits
./scripts/bl image edit \
--image projects/$SPARK_VIDEO_PROJECT/cast/陆辰/portrait1.png \
--prompt "Create a three-view drawing of the same person and annotate it with an AI-generated, non-real-person watermark on the bottom. Change the character's outfit to a large red traditional Chinese wedding robe and red wedding cap; keep face and hairstyle unchanged. Output a full-body character reference sheet with front view, side view, back view, same face and costume in all views, plain background, no other readable text, $(uv run scripts/scaffold.py mood-anchor)" \
--out-dir projects/$SPARK_VIDEO_PROJECT/episode-$SPARK_VIDEO_EPISODE/cast/陆辰/ \
--out-prefix portrait
uv run scripts/scaffold.py cast-init
bl image edit preserves face identity better than bl image generate
for forks — always prefer edit when you have a project-tier reference sheet
or portrait to base from.
For pixel-perfect face identity (edit can still drift slightly), drop
a hand-edited PNG / three-view sheet into the episode cast folder instead
of using bl.
Procedure 2 — scaffold a movie-set
2.1 When to scaffold a set
Scaffold whenever:
- Two or more shots happen in the same location with the same lighting.
- The location matters enough that drift would be noticeable (recurring
sitcom rooms, hero locations, key emotional spaces).
- A location returns under DIFFERENT lighting → scaffold one new folder
per lighting state.
Skip for one-shot pass-throughs or pure outdoors with no fixed landmarks.
2.2 Naming — lighting state in the folder name
| Same physical place, different… | Action |
|---|
| Time-of-day (day / dusk / night / pre-dawn) | Separate folders (客栈大堂-白天, 客栈大堂-夜晚) |
| Season (spring / summer / autumn / winter) | Separate if visible (willows / snow / red leaves) |
| Color grade (memory cold gray / present warm yellow / high-contrast neon) | Separate folders |
| Weather (clear / rain / snow / fog) | Separate when weather is in frame |
| Decor unchanged, action just moves around the room | Same folder |
2.3 Scaffold + generate
uv run scripts/scaffold.py set --name "同福客栈大堂-白天"
uv run scripts/scaffold.py set --name "出租屋客厅-暖灯" --episode
./scripts/bl image generate \
--model wan2.6-t2i \
--prompt "Ming-Qing style wooden inn lobby, two-story wooden staircase, red lanterns, three square tables, daytime natural light through windows, warm yellow tone, $(uv run scripts/scaffold.py mood-anchor)" \
--size 16:9 \
--out-dir projects/$SPARK_VIDEO_PROJECT/movie-set/同福客栈大堂-白天/ \
--out-prefix set
uv run scripts/scaffold.py set-init
The set.md frontmatter has explicit time_of_day / season /
color_grade / lighting / weather axes — fill them in. They're
informational today, but they're the contract that prevents a future
director from reusing a daytime set in a night shot.
Procedure 3 — scaffold a prop
3.1 When to promote an object to a key prop
Promote any object to a key prop when it satisfies either:
- It appears in 2+ shots and the audience would notice if it changed
shape/material/color/wear (the red envelope in S01-003 → S01-007 → S04-002).
- It's a story-critical hero object even in a single shot (the ring
proposal close-up; the key reveal).
Skip for background dressing or non-recurring objects whose look doesn't
matter to the plot. Budget: 3–6 named props per episode, more is a smell.
3.2 Scaffold + generate
uv run scripts/scaffold.py prop --name "戒指-完整"
uv run scripts/scaffold.py prop --name "红包-起皱" --episode
./scripts/bl image generate \
--model wan2.6-t2i \
--prompt "Standard Chinese red envelope, large red hot-stamped pattern, printed with '囍', flat with no creases, pure white background, product photography style, $(uv run scripts/scaffold.py mood-anchor)" \
--size 1:1 \
--out-dir projects/$SPARK_VIDEO_PROJECT/props/红包-完整/ \
--out-prefix prop
./scripts/bl image edit \
--image projects/$SPARK_VIDEO_PROJECT/props/红包-完整/prop1.png \
--prompt "Add obvious creases and grip-worn folds to the red envelope; keep color, print, and shape exactly unchanged" \
--out-dir projects/$SPARK_VIDEO_PROJECT/episode-$SPARK_VIDEO_EPISODE/props/红包-起皱/ \
--out-prefix prop
uv run scripts/scaffold.py prop-init
For state changes, always prefer bl image edit with the base state
image as input — preserves shape/print/material continuity. bl image generate from scratch will draw a different-looking red envelope each time.
Generation tips (apply to all three)
Mood anchor — append it to every t2i prompt
uv run scripts/scaffold.py mood-anchor
Without it, your asset visual style won't match the rendered shots.
Aspect ratio defaults
| Asset type | --size |
|---|
| Cast reference sheet (full-body three-view, default) | 16:9 |
| Legacy single cast portrait (only when provider accepts it) | 3:4 or 9:16 |
| Set establishing | 16:9 |
| Prop (product-style) | 1:1 |
Batch generation in parallel
bl image generate supports --n N --concurrent K — useful when
scaffolding many NPCs or sets at once. Each --n produces a candidate;
keep the best, delete the rest.
./scripts/bl image generate --n 3 --concurrent 3 \
--model wan2.6-t2i \
--prompt "..." \
--out-dir projects/.../cast/陆辰/ \
--out-prefix candidate
Multi-image merge (cast fork only)
bl image edit accepts multiple --image flags. Useful when forking a
cast with a costume reference photo:
./scripts/bl image edit \
--image cast/陆辰/portrait1.png \
--image refs/hanfu-reference.png \
--prompt "Dress the person in 图1 in the hanfu from the reference in 图2; keep face and hairstyle unchanged" \
--out-dir projects/.../episode-X/cast/陆辰/
After scaffolding — rebuild manifests
The merged manifests (cast.json, movie_set.json, props.json)
must be rebuilt after any folder change. They drive the director's
shot-id lookups and the renderer's media[] resolution:
uv run scripts/scaffold.py cast-init
uv run scripts/scaffold.py set-init
uv run scripts/scaffold.py prop-init
uv run scripts/scaffold.py manifests
Tell the director (or the producer at GATE 2) when you've added new
assets — they need to read the updated manifests before storyboarding
any scene that references them.
DON'Ts
- ❌ Don't put two lighting states (day + night) in the same set folder.
The model averages and produces "neutral gray noon-night" garbage.
- ❌ Don't put two prop states (intact + creased) in the same prop folder.
Same reason.
- ❌ Don't solve a costume change by writing "wearing XXX" in shot prompts.
Fork the cast reference sheet instead.
- ❌ Don't omit the mood_anchor in t2i prompts. Visual cohesion will
break across shots vs cast references.
- ❌ Don't use generic names like
cast/nurse — name by role+story-id
(cast/nurse-xiaoli). When two episodes both have a "nurse", you can't tell
whose cast reference is whose.
- ❌ Don't use provider / CLI watermark flags such as
--watermark.
For generated cast sheets, ask for the bottom provenance
annotation in the prompt itself:
AI-generated, non-real-person. Keep it outside the character body
and avoid any other readable text.
- ❌ Don't skip
scaffold.py *-init after adding folders. The manifests
are the only thing the rest of the pipeline reads.
