Animation Frames — video-first
The image generator draws poses; the VIDEO generator understands motion — leg alternation, arc continuity, cloth and fire dynamics come free because video must animate them. So don't ask the image model to imagine mid-motion poses: animate the base and harvest real frames.
App-builder execution (read first)
This skill is doctrine (motion laws, loop QC, when to use video). In the
app-builder sandbox, do not run a freeform ffmpeg harvest or invent a
random keyable #hex background.
| Step | Do this |
|---|---|
| Production sprites / fixed grids | generate2dsprite — solid #FF00FF magenta sheets + chroma scripts |
| Denser locomotion from video | video2dsprite — base still on #FF00FF → imagine_image_to_video → skill scripts (ffmpeg + chroma) |
| Keyable background | Always #FF00FF when using either pipeline (required for chroma) |
| This skill | Loop / flip-test / motion laws below — apply after the pipeline runs |
Open .grok/skills/video2dsprite/SKILL.md or .grok/skills/generate2dsprite/SKILL.md
and follow their workflows for generation and postprocess. Then apply the
laws and flip test here before shipping frames into the game.
Default pipeline (intent)
- Base frame. Subject in neutral/starting pose, full style words, side /
game-appropriate view, solid
#FF00FFbackground (app-builder chroma key). game-asset-core defaults apply. - Animate.
imagine_image_to_videofrom the base: one clear motion, in place, static camera ("the knight walks in place, side view, camera locked", 6s). Keep the shot simple — one subject, one motion. Prefer running this throughvideo2dspriteso harvest + chroma are consistent. - Harvest + clean. Use
video2dspritescripts (not ad-hocffmpeg -i … fps=12alone). Magenta flood-fill / despill lives there; do not re-key with a different flat#hexunless you leave the magenta pipeline entirely. - Select. Pick frames that (a) capture the motion's distinct phases and (b) LOOP — the sequence's end must flow back into its start. Don't force a count: if the motion reads best with 8, 10, or 12 frames, deliver that many (more frames = smoother in-engine). For a cycle, select one full period using motion landmarks (foot contacts, wing extremes, flame peaks).
- Package. Deliver frames in play order (zero-padded names) and/or a
sheet per game-asset-core rules (uniform cells, no dividers) — or the
transparent strips/grids emitted by
video2dsprite/generate2dsprite. State the intended fps.
Fall back to keyframe-by-keyframe imagine_text_to_image (still on #FF00FF when
postprocessing with the sprite scripts) only when video fails the motion
(rare: very stylized poses, single dramatic keyframes) — and then plan
phases yourself and obey the laws below. Prefer generate2dsprite for
crisp production multi-frame grids.
Motion laws (verify against these, whatever the pipeline)
- Cycles loop; alternating gaits spend half the period mirrored.
- Continuity: limbs, props, anatomy, effects move on continuous paths — nothing teleports, vanishes, or duplicates between adjacent frames.
- Physics reads in stills: airborne shows air, anticipation compresses, follow-through overshoots; effects stay anchored to their origin unless the request moves them.
- Energy matches the ask: idle/subtle means barely-different frames.
Verify — the flip test
View the final frames strictly in order and narrate the motion; check loop closure explicitly (last→first). A hedge in your narration is a failed frame. The video pipeline usually passes this on the first try — that's why it's the default.