How it works
Plain text-to-image models hallucinate a plausible dent but never a precise one, in the right place, physically consistent with the bodywork. Masked inpaint fixes that: you generate on a real photo and keep the rest intact.
Each base image can then go through the deterministic stage (WhatsApp compression, angle, downscale) to mimic how a claimant actually sends a photo. One paid generation becomes a full test set.
Key points
- Masked inpaint: damage painted on the real photo, not invented from scratch.
- Labelled output: model, seed, ops and human rating stored with the image.
- Ideal for claim detectors, auto expertise and field robustness.