Normal2RGB

Normal-map conditioned generation helper. Lora trained to preserve source geometry / normals while generating RGB output.

Checkpoints in this repo:

Normal2RGB_2000.safetensors
Normal2RGB_3000.safetensors
Normal2RGB_4000.safetensors

Each comparison below is a 3-panel strip: Source (Normal Map) | with Normal2RGB | without Normal2RGB.

Comparisons

1. Product Shot - Wireless Mouse

Left: Source normal map. Middle: with Normal2RGB - preserves mouse shape, scroll wheel, side buttons and viewpoint. Right: without Normal2RGB - geometry and details drift.

2. Illustration - Boy Studying with Dog

Left: Source normal map. Middle: with Normal2RGB - preserves pose, table layout, and dog position. Right: without Normal2RGB - composition changes, extra objects appear.

3. Studio Portrait - Swimmer with Racket

Left: Source normal map. Middle: with Normal2RGB - preserves profile pose, racket strings, and ball stack. Right: without Normal2RGB - face angle, racket and props shift.

4. Landscape - Abandoned Industrial Railway

Left: Source normal map. Middle: with Normal2RGB - preserves tracks, structures, and foreground debris layout. Right: without Normal2RGB - scene gets overgrown / re-imagined, layout drifts.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support