Consistent characters across hundreds of generations

Exploring whether modern image models can create production-ready game assets instead of isolated concept art. A ComfyUI graph combining reference conditioning, a trained LoRA and a strict prompt schema — the interesting part was making the failures predictable enough to fix.

The graph, in plain terms

1. Load reference sheet → face crop
2. IP-Adapter · weight 0.62
3. Character LoRA · 1400 steps
4. Prompt schema: [pose] [outfit] [lighting]
5. Seed sweep 8 → pick 1, upscale ×2
6. Manual pass: hands, silhouette, palette

Nothing here is exotic. The value is in the constraints: a fixed prompt schema, a small trained model, and a human pass that only fixes the four things the model reliably gets wrong.