Experiments

The repository includes a sweep runner that records per-variant outputs, compare boards, and a canonical summary.

Commands

Run the full sweep:

uv run python .\scripts\run_edge_experiments.py

Resume only missing work:

uv run python .\scripts\run_edge_experiments.py --resume

Rebuild Markdown and image artifacts without rerunning inference:

uv run python .\scripts\run_edge_experiments.py --postprocess-only

Saved overview

Current recommendation

• balanced / 320 / 4 steps looked best on:
  • car-palette-team-style-2k.png
  • delpmaspu-2k.png
  • grok-image-square.png
• balanced / 256 / 4 steps is the safer fallback and was the best successful option for kling-generate-square.png

Per-input boards

car-palette-team-style-2k.png

A vehicle cutout with a long curved spoiler edge, trim decals, and narrow protrusions. Use this board to judge whether the alpha stays smooth along curved bodywork without turning into staircase pixels.

Start with balanced / 320 / 4 steps.

delpmaspu-2k.png

A flat illustration with a shallow cream-colored curve. This case is useful for separating soft halos from truly clean anti-aliased edges on rounded shapes.

Start with balanced / 320 / 4 steps.

grok-image-square.png

A repeated jet-like silhouette with diagonal fins and sharp corners. This board is the easiest way to compare crispness on angular edges without falling into the blockiness of hard.

Start with balanced / 320 / 4 steps.

kling-generate-square.png

A low-resolution pixel-art dog. This board shows how well each setting respects intentionally chunky outlines while still avoiding over-quantized hard-mask artifacts.

Start with balanced / 256 / 4 steps because the 320 / 4 steps variant failed on this machine.

Repository references

• Experiment gallery README
• Canonical summary CSV
• Per-input compare boards