storyboard

Shared foundation for turning a recurring character + a scene description into a scene-specific anchor image that LTX ia2v / flf2v / multiguide can consume. Music-video and drama-video both delegate their anchor stage to this skill; one-off renders can call the CLI directly.

Why this exists — the trap without it

Pipelines that use image: "@anchor" on their first shot often pass a raw character sheet (neutral backdrop, canonical pose) as the LTX ia2v starting frame. The shot prompt then asks for a completely different scene ("dimly lit studio, sweat, mirror, overhead light"). LTX compromises — you get the character-sheet pose wearing the sheet's neutral-backdrop lighting, with the new scene vaguely visible behind. Looks wrong. Always.

The fix: always flux2 i2i (or i2iN) the character sheet against the scene prompt first to produce a shot-specific anchor — character in the setting, new lighting, appropriate pose — and feed that into LTX. This skill owns the pattern.

Anchor types

CLI — generate_anchor.py

python3 /home/sandbox/.openclaw/skills/storyboard/scripts/generate_anchor.py \
    --out     <path>.png \
    --type    t2i|i2i|i2i2 \
    --prompt  "..." \
    [--reference  a.png]                   # i2i
    [--references a.png,b.png]             # i2i2 / i2iN
    [--subject    name=description] ...    # repeatable — {name} substituted in prompt
    [--style      editorial|indie|storybook|risograph]   # preset tail, or --style-tail "<custom>"
    [--width 1024] [--height 576] [--steps 8]

Idempotent — if --out already exists, skips. Subject tokens are expanded before the prompt goes to flux2. Valid camera-LoRA names are validated only when the caller also runs an LTX pass; this CLI just makes images.

Examples:

# Single-character scene anchor
generate_anchor.py --type i2i \
    --reference character.png \
    --prompt  "{fennie} in a late-night dance studio, overhead spot, wooden floor, mirror behind her, sweat on her brow, paused mid-balance" \
    --subject "fennie=Fennie, 17yo Korean ballet trainee, black hair in a tight bun, black leotard, pale pink tights" \
    --style   indie \
    --out     shots/001-opening-anchor.png

# Two-character meeting
generate_anchor.py --type i2i2 \
    --references fennie.png,trainer.png \
    --prompt  "{fennie} kneeling on the studio floor; {trainer} kneeling beside her, hand on her shoulder, warm key light" \
    --subject "fennie=…" --subject "trainer=older man, trimmed grey beard, charcoal wrap cardigan" \
    --style   indie \
    --out     shots/003-coaching-anchor.png

Reference sheets — project-level consistency (ingredients IC-LoRA)

Per-shot flux2 anchors keep a character consistent one shot at a time. For a whole music-video / drama where the same character, props and location must stay consistent across every shot, use an ingredients reference sheet: one composite image with a clean panel per recurring element (character face + turnaround, each prop, the location) on a black background, no text. LTX-2.3's ingredients IC-LoRA conditions each clip on that sheet (see the comfyui skill's ingredients command). This is video-level (it replaces/augments the ia2v anchor), not an image anchor type.

CLI — generate_reference_sheet.py

Generates the panels via flux2 (character turnaround via multiprompt, props/location via t2i) and tiles them with the comfyui skill's make_reference_sheet.build_sheet:

python3 /home/sandbox/.openclaw/skills/storyboard/scripts/generate_reference_sheet.py \
    --out        project_sheet.png \
    --character   marco.png \
    --character-desc "Marco, early-30s fisherman, tan skin, short black hair, stubble, red flannel" \
    --angle      "front-facing head-and-shoulders portrait" \
    --angle      "full-body three-quarter view" \
    --prop       "trident=an ornate golden three-pronged trident" \
    --prop       "lantern=an old brass oil lantern" \
    --location   "weathered wooden pier at golden hour, calm sea, distant lighthouse" \
    --width 768 --height 448          # = the project render WxH (sheet aspect must match)

• --prop name=description → a flux2 t2i panel; --prop name=path.png → uses that image directly (skip generation). --location likewise.
• Without --character it makes a single identity panel from --character-desc.
• --style "<tail>" appends a look to every panel; --grid RxC forces layout.
• Idempotent (skips if --out exists; --force to re-render).

Feed the sheet to comfy_graph.py ingredients --sheet project_sheet.png --prompt …, or set a project-level reference_sheet: + per-scene mode: ingredients in music-video / drama-video (see those SKILL.md). Author the sheet at the project render aspect — the ingredients graph stretches it to the output size, so a mismatched aspect distorts the reference. Trained buckets are landscape (768×448, 960×544); portrait is off-bucket — evaluate.

Sheet vs per-shot anchor — when to use which

Python API — lib/prompts.py

Orchestrators can also import the prompt-preprocessing helpers directly:

import sys
sys.path.insert(0, "/home/sandbox/.openclaw/skills/storyboard/lib")
from prompts import expand_subjects, apply_style_tail, STYLE_TAILS

• expand_subjects(text, subjects: dict[str,str]) -> str — {name} → value.
• apply_style_tail(text, style_id_or_custom) -> str — append a known-good style suffix.
• STYLE_TAILS: dict[str,str] — named tails (see references/style-tails.md).

Multi-guide resolver — lib/guides.py

Shared resolver that turns a yaml-friendly guides: list into (frame_idx, absolute_path, strength) tuples ready for comfy_graph.py multiguide --guides --frame_indices --strengths. Both music-video and drama-video use this so a yaml authored for one skill reads the same in the other.

import sys
sys.path.insert(0, "/home/sandbox/.openclaw/skills/storyboard/lib")
from guides import resolve_guides, ensure_frame_zero

resolved = resolve_guides(
    scene.get("guides"),              # the yaml list
    duration_sec=scene["duration_sec"],
    fps=24,
    total_frames=None,                # defaults to int(duration_sec * fps)
    project_dir=Path("./project"),    # relative image paths resolve against this
    token_resolver=lambda tok: ...,   # optional: resolve @anchor/@last → abs path
)
ensure_frame_zero(resolved, image_path=first_frame_path, strength=1.0)
# → [(0, "/abs/first.png", 1.0), (96, "/abs/mid.png", 1.0), ...]

Supported per-entry keys:

Frame indices are snapped to multiples of 8 (LTX's temporal latent quantization) and clamped to [0, ((total_frames-1)//8)*8]. The output list is sorted by frame_idx ascending. ensure_frame_zero prepends an entry at frame_idx=0 if none of the resolved guides targets frame 0 — useful when the yaml-level list only describes mid-scene anchors and the caller wants the first-frame anchor (e.g. a resolved @anchor PNG) to be included automatically.

When to add guides to a shot

• First frame can't show the character — e.g. scene opens on a prop, a landscape, an empty room. LTX will invent a random identity halfway through unless you anchor it with a mid-scene guide at strength 1.0.
• Long shots (≥13s) — single-anchor ia2v drifts by the tail of a long shot. Add a second anchor near the end (e.g. at_relative: 0.9) to keep the character consistent through the whole take.
• Specific visual beats synced to lyrics/audio — place an anchor at the exact frame where a line lands to time a character appearance to the word.

When to use what

See also

• references/prompting-guide.md — the hard-learned rules: subject tokens, action emphasis, pose-neutral descriptions, common failure modes (lineups, Ghibli-ification, identity drift).
• references/style-tails.md — preset style suffixes by aesthetic.
• references/camera-loras.md — valid LTX camera LoRA names (relevant for downstream video, not for anchor generation).
• comfyui skill — the flux2 t2i/i2i/i2i2/multiprompt subcommands this CLI wraps, plus the ingredients command and make_reference_sheet.py tiler that generate_reference_sheet.py feeds.