Skip to content
AI Artist
GitHub repository
All benchmarks

Knight — pose and sword slash

A still 32×32 knight, then a sword slash animated from it. Tests whether a model can hold the brush — silhouette, palette discipline, a rig that holds its own contract, and motion that reads as an arc rather than a metronome.

  • validate
  • silhouette
  • palette
  • contract
  • tagged
  • poses
  • arc
  • timing
Canvas
32×32 RGB
Palette
pico8
Steps
2
Revision
1

The steps, verbatim

  1. Step 1:The still knight

    Draw me a 32×32 knight with a sword, standing still. One frame.

    new session

  2. Step 2:The slash

    Starting from this knight, animate a sword slash: windup, then follow-through. 4 to 6 frames.

Setup rules
  • Start from a new document created with sprite_manage op="new" width=32 height=32 colorMode="rgb", then palette op="preset" preset="pico8".
  • Hand the model exactly the step text — no reference image, no extra steering beyond the server's own instructions and skills.
  • One session per step. Step 2 runs in a fresh session against a copy of the step-1 result.
  • Do not intervene mid-run unless the model asks a genuine clarifying question, and record it under the step's interventions.

8 criteria

  1. Step 1: The still knight

    • validate passes with no unaddressed errors (warnings may stand if the report names and justifies them).
    • The preview reads as a knight with a sword at a glance — head, torso and sword distinguishable without the ascii grid.
    • Every colour is on the PICO-8 palette (validate check palette, or a clean palette op="analyze").

  2. Step 2: The slash

    • validate with expect set to the model's own rig (layerFrames, mustNotOverlap [["sword","body"]]) passes — the contract it built holds itself.
    • The cycle is tagged — tag op="list" shows a named tag covering all frames.
    • Windup reads as a distinct pose from follow-through on the filmstrip, not two frames either side of a single held pose.
    • The onion-skinned in-between shows the sword's arc moving, not just the arm translating in a straight line.
    • Frame timing is not uniform — frame op="list" durations vary, fast on the strike and longer on the hold.

Ranking

No ranked runs on revision 1 yet. The maintainers run the benchmarks; new models are added as they are tested.

Every run

No runs yet. The maintainers run the benchmarks; new models are added as they are tested.