Video to 2D Animation
An approved still sets identity and geometry, a video model supplies motion, and deterministic
tools prepare it for the engine. Video is not promised to be smaller, seamless or cheaper than sheets.
Use when
- Animating a character or creature: every action comes from one approved master still, one clip per action.
- Fluid motion from an accepted master: idle breathing, walks, runs, attacks, casts, hair, capes, creatures.
- A supplied clip must become transparent, registered, looped frames and engine video.
Do not use when
Capability check (once per session)
Run python "<skill-dir>/../generate2dmedia/scripts/forge_doctor.py" --host-tools <tools> --save <output>/doctor.json, where <tools> lists the media tools in your own tool list (image_gen, image_edit, image_to_video) or none. Its ROUTES block shows which API keys are configured (yes or no), the local readiness and the resolved order. If encoding.stdout fails, set PYTHONUTF8=1. python "<skill-dir>/scripts/video2dsprite.py" doctor checks ffmpeg (5.1+, VP9 alpha round trip).
Art source
The master still comes from generate2dsprite (image generation, approved by the user) or from the user. Motion is image-to-video, one clip per action from that master; record the route as art_source and name it in your reply:
- API, when a key is configured: xAI Grok Imagine Video (
XAI_API_KEY), BytePlus Seedance (ARK_API_KEY) or fal.ai (FAL_KEY: Kling v3, Veo 3.1, Luma Ray 3.2, MiniMax H3, Wan 3.0, Vidu Q3, LTX) (route-media.md). The configured key is the owner's consent, so there is no per-call question. --last-frame <still> pins idle and attack clips to end on their first pose.
- Local: your own image-to-video tool (Grok's native
image_to_video) when you have one, otherwise the user's Grok CLI in ACP mode (it cannot pin a last frame; the result says lastFrameUsed).
- A clip the user supplies (no account needed).
- codeart2d only when the user explicitly asks for code-drawn animation, or when
route_media.py prints no-route (exit 3); otherwise prepare the job and prompt, explain the gap, and never invent a successful generation.
Every clip goes through one command: python "<skill-dir>/../generate2dmedia/scripts/route_media.py" video --prompt-file <job>/prompt.txt --reference <job>/input.png --last-frame <job>/input.png --duration 6 --resolution 720p --out-dir <new>. It prints one JSON line with route, artifact, sha256, estimateUsd and lastFrameUsed. The finish is HD by default; the pixel finish is an option when the user asks for pixel art.
A whole sprite set from one still
For a character's game set, use sprite_set.py: one approved master (generate2dsprite master.json) in, one image-to-video clip per action out, keyed, gated, registered, cut, finished and packaged. For one clip or a supplied video, use the pipeline below. Details: sprite-set.md. Prompts and fix clauses: motion-prompts.md.
- Plan.
python "<skill-dir>/scripts/sprite_set.py" plan --master <art>/master.json --output-dir <set>. Defaults: --actions idle,walk,run,attack,jump,hurt, --finish from the master (hd default), --target-height 256 hd or 80 pixel. Jump gets a 3:4 canvas with 34% headroom and attack a 16:9 one; idle and attack pin the end frame to the master. Read each action's prompt in <set>/set_plan.json, and edit motion, negatives or gates there before the run if needed.
- Run.
python "<skill-dir>/scripts/sprite_set.py" run --plan <set>/set_plan.json. Clips come from generate2dmedia route_media.py, then:
- soft-matte keying;
- numeric gates: area 0.72-1.32, feet, identity (a warning unless the design is lost), zoom, turning around, edge, background, extra objects, motion, end pose, colour bleed, one-shot timing;
- up to 3 takes, each retake carrying the fix clause for the gate that failed; if every take fails, the best usable window is kept;
- registration, then a loop, or a one-shot retimed to game length from its motion energy (attack 0.7 s with the hit at 40%, jump 0.9 s, hurt 0.5 s, cast 0.9 s; set in the plan's
timing);
finish_frames.py hd or pixel, with the colour lock to the master on by default (plan --no-colour-lock);
- png, webm and packed packages, verified.
Run it again to resume; nothing is generated twice.
- No route (exit 3). Make the clip with your own image-to-video tool from the input and prompt that the summary names, save it as the named
media/clip.mp4, and run again.
- Review.
python "<skill-dir>/scripts/sprite_set.py" review --plan <set>/set_plan.json, then LOOK at every -sheet.png, every -final.png and the lineup (Read in Claude Code, view_image in Codex). The numbers cannot see identity, facing, a changed weapon, text or a covered face.
- Decide each action.
- Approve:
accept --plan <set>/set_plan.json --action <id> --take <n>. Add --window START:END to keep part of a failed take.
- Reject:
retake --plan <set>/set_plan.json --action <id> --fix <clause id or text>, then run again. Clause ids include never-turn, weapon-shape, face-visible, no-text, no-flares and keep-colours.
- Report.
python "<skill-dir>/scripts/sprite_set.py" report --plan <set>/set_plan.json lists the routes, takes, QC numbers, loops, timing and package paths. In your reply, name the route that made each clip. Never call an action approved before you have looked at its sheet.
Host notes
- Codex: no native image-to-video tool;
route_media.py takes the xAI API or the Grok CLI. Look at review images with view_image.
- Claude Code: no media tools;
route_media.py takes the route. Look at review PNGs and contact sheets with Read; run tools with Bash; <skill-dir> is ${CLAUDE_SKILL_DIR}.
- Grok: its native
image_to_video tool is your own tool; the same grok executable is the "Grok (local CLI)" route for other hosts.
Commands
Run each tool as one line from the user's project root: python "<skill-dir>/scripts/<tool>.py" .... <skill-dir> is this skill's folder; sibling skills sit beside it. Keep inputs and outputs inside the project. Every verb writes a new --output-dir: it refuses an existing one and publishes only after its checks (--strict fails on residue). Success prints one JSON line; errors print error: ... and exit 1; usage errors exit 2. Needs Python 3.10+, numpy, Pillow and ffmpeg 5.1+. --help lists every flag.
Pipeline (one clip)
- Key plan.
video2dsprite.py key-plan --master <master.png>: paste background_sentence into the prompt; protect or recolour design_colours_at_risk.
- Prepare.
prepare_i2v_input.py prepare --master <art> --action <idle|walk|run|attack|cast|guard|hurt|victory|defeat|ambient|fx> --output-dir <job>; send input.png and prompt.txt; registration_job.json keeps the transform. An opaque master on a flat magenta, green or blue backdrop is keyed first (--master-key names another backdrop; the job records masterKeying). --subject is the whole identity phrase; the background line names its head ("adventurer with a scarf" gives "behind the adventurer"). Check a hand-written prompt with prepare_i2v_input.py lint.
- Generate with
route_media.py video (Art source above).
- Accept the take.
register_clip.py qc --job <job>/registration_job.json --video <take>; then video2dsprite.py triage --video <take> --output-dir <new>. A border touch inside the action means regenerate, never pad.
- Key.
video2dsprite.py process --video <take> --output-dir <new> --reference <master.png> (soft matte); read frames-clean/matte-report.json and look at frames over light and dark. Same keying for every clip of a character: --matte-profile <character-profile.json>. A design colour near the key: --protect-color #rrggbb. Already decoded frames (frame_*.png or raw_*.png, one frame keys a still): video2dsprite.py clean --raw-dir <dir> --output-dir <new>.
- Register.
register_clip.py apply --job <job>/registration_job.json --frames <out>/frames-clean --output-dir <new>; feet drift or jumps: --lock feet; spell or hit FX: --profile fx; one scale per character: register_clip.py profile, then --character-profile.
- Loop or retime. Walks, runs, idles:
gait_loop.py select --frames-dir <reg>/frames --fps N --output-dir <new> (--kind idle or hover); play aids/loop3x.gif. Attacks, casts, FX: retime.py --frames-dir <reg>/frames --fps N --output-dir <new> --spans <spans> (impact, hold, ticks), or for a slow-motion one-shot --range 0:N --auto-oneshot --duration 700 --key hit=<impact>@0.4+80 (cuts frozen holds, pins the hit, compresses to game length); walk cadence: gait_loop.py measure-stride.
- Finish and package. Finish the registered frames (Finish, below), then
engine_export.py package --clean-dir <fin>/frames --output-dir <new> --selection <selection.json> --source-size W,H --source-anchor X,Y --formats png,webm,packed --tiers actor (W,H and X,Y from finish.json). Refused for key residue: re-key; --allow-key-residue only with a recorded override.
- Verify.
engine_export.py verify --package <dir>, then validate_animation.py <character dir> --require-states idle,walk --require-verify. --require-states matches each package's --name (a package named hero-run needs --require-states hero-run, not run).
Details: pipeline.md, keying matte.md, review and cuts animation-review.md, prompts prompt-rules.md.
Finish (HD default, pixel option)
Finish every registered action before packaging. Finished frames keep the source names, so gait_loop and retime selections still apply. Details: finishing.md.
- HD (default):
python "<skill-dir>/scripts/finish_frames.py" hd --frames <reg> --output-dir <new> --target-height <px> --role hero --character <id>. The rest pose (frame 0 or --rest-frame) sets one scale. Box downscale with premultiplied alpha on a grid pinned to the feet, then clean alpha. Never nearest, never an upscale. --display-sizes 1x,2x rebuilds each size from the source.
- Same character, other actions:
--scale-ref <idle finish dir> instead of --target-height (scale x sqrt(rest-area ratio), clamped to +-3%). A warning means the rest frame is not the base still.
- Pixel (on request): first
finish_frames.py palette build --frames <every action's reg dir> --colors 32 [--reserve #hex] --out <palette.json>, then per action finish_frames.py pixel ... --target-height <1/8 to 1/12 of the source> --palette <palette.json>. The default is bold pixel art: a crisp 4x4-cluster downscale (no blurred in-between colours; lines and eyes survive), a light contrast and saturation lift, at most --colors (32) colours per frame, lone-pixel cleanup that keeps eyes, a 1 px selective outline (--outline dark|none), temporal hysteresis, binary alpha, no dither. --canvas 48x64 gives fixed cells. Use --loop-policy oneshot for attacks and --indexed-sheet for an indexed sheet.
- Colour lock:
--colour-lock <keyed master.png> keeps every design colour on the master's in OKLab (chroma only, lightness kept: no posterising) and smooths still pixels over time; colour_lock.py measure reports hue flips and region spread.
- Cast check:
finish_frames.py lineup --sets <finished dirs> --out <lineup.png>. Roles come from --role or ROLE=DIR; the default rules are hero:1.0, mob:<=1.0, boss:~2.0, spirit:~1.0 (by area). Look at the x1/x3 sheet; a failed rule exits 1.
- Review: look at
review-contact.png and read qa in the JSON line (flipsPerPair, specksRemoved, alphaBinary). The status stays needs-visual-review.
- Package:
engine_export.py package --clean-dir <fin>/frames --source-size W,H --source-anchor X,Y, taking W,H and X,Y from size and anchor in finish.json. Do not pass --registration; pixel finishes add --pixel-art --sampling nearest.
Motion and geometry rules
- One clip, one action. The game owns translation, hit timing, damage and state changes.
- One fixed transform per clip; never crop or resize frames to their own alpha bounds (it pumps size and pins airborne feet). Never freeze the still's alpha over moving RGB.
- All actions of one character share one body height, taken from the master's rest pose; never upscale a finished sprite.
- Choose cycles with
gait_loop.py select, never by eye alone, then confirm by eye. Walks never ping-pong; attacks never recover by playing frames backwards; retime actions to their impact and hold. Walks ship cadence and stride.
- Keying claims quote the matte report's numbers and stay
needs-visual-review. --matte binary --despill-mode off reproduces the cfed170 keyer.
Runtime and acceptance
- Acceptance order: package, then verify, then
validate_animation.py. Deliver animation.json (3.0, keeps every 2.0 key; whole-ms durationsMs, exact fpsRational), QA, poster, PNG fallback and the requested transports. Geometry stays in source units (sourceSize, sourceAnchor, sourceRect).
- Packed H.264 stores RGB left and alpha right; it is not a transparent MP4. Draw the compositor's
lease.drawable (packed-alpha-runtime.js over references/runtime/packed-alpha-webgl.mjs); show the poster until a decoded frame exists. Distance-driven walks, hit-stop and transitions: references/runtime/forge-runtime.mjs.
- Inspect at game scale: face and silhouette stability, contact, moving alpha edges over light and dark, loop seam, readability. Verify real iPhone playback before claiming alpha or frame rate; compare bytes and decode cost before replacing sheets.
Data contracts: references/schemas/video.schema.json.