The model emits JSON defect proposals for dust, dirt, scratches, hair-like surface contamination, emulsion damage, chemical stains, and light leaks. The Project Halide runtime validates the JSON schema, removes low-confidence or duplicate boxes, and uses tiled inspection when large scans hide thin crack networks at full-frame scale.

Training Summary

Base model: openbmb/MiniCPM-V-4.6
Training method: LoRA fine-tuning with LLaMA-Factory, merged for inference
Curriculum: FilmDamageSimulator annotations, procedural film-defect positives, hard clean negatives, and a v7 crack curriculum
Held-out private negatives: used only for evaluation, not for training

Held-Out Smoke Result

Final v7 checkpoint with 960 px tiled fallback:

Sample	Expected surface condition	Result
negative1	Long scratches across portrait	8 defects
negative2	Abraded emulsion and dirt patches	9 defects
negative3	Severe emulsion damage and debris	6 defects
negative4	Near-clean hard negative	0 defects
negative5	Broad lifted crack network	45 defects

Runtime Notes

This model is intended to run inside Project Halide with GPU inference. The runtime refuses local CPU model inference and does not call cloud inference APIs.

The repo also includes a llama.cpp Q4_K_M GGUF artifact: minicpm-v-4.6-merged-v7-crack-curriculum-r1-ckpt625-q4_k_m.gguf.

Use the model as an inspection aid. It can over-box broad damage regions, and film metadata should be treated as context unless verified by notes or edge marks.

Lonelyguyse1/halide-vision

Halide Vision

Project Links

Training Summary

Held-Out Smoke Result

Runtime Notes

Lonelyguyse1/
halide-vision