image-forge

Quality-first image generation skill for OpenClaw. It runs a round-1 generator tournament, critiques candidates with three specialist reviewers, merges their scores into a weighted rubric, and either accepts, regenerates, or routes the image into targeted editing.

What Ships

• Round-1 tournament generation across fal.ai backends
• Three parallel critics: structural, material/color, photography
• Weighted composite scoring and explicit pass/refine routing
• Targeted follow-up editing via Nano Banana edit when the failure is local
• Per-run artifacts in outputs/

Requirements

• python3
• FAL_KEY
• OPENROUTER_API_KEY

Install dependencies:

python3 -m pip install -r requirements.txt

Usage

Balanced quality-first run:

python3 scripts/generate.py \
  --prompt "Luxury flatlay product mockup of a cropped flannel overshirt on a clean warm-white background" \
  --reference-image ./refs/garment.png \
  --reference-image ./refs/swatch.png

Fast mode:

python3 scripts/generate.py --preset fast --prompt "..."

Premium mode:

python3 scripts/generate.py --preset premium --prompt "..."

Outputs

Each run writes:

• outputs/<timestamp>-<slug>/run.json full manifest
• outputs/<timestamp>-<slug>/summary.json condensed status
• outputs/<timestamp>-<slug>/images/ downloaded candidates/finals when enabled

Notes

• Default preset is balanced
• fast skips the tournament and runs a single FLUX generation
• balanced runs FLUX + Nano Banana tournament, then refines for up to 2 rounds
• premium increases depth and resolution
• Critics run through OpenRouter; generation/editing runs through fal.ai