Spaces:

---

build-small-hackathon

/

project-halide

like 3

Running on Zero

Project Halide is a GPU-backed Gradio workbench for analog film diagnostics.

It combines a fine-tuned MiniCPM-V 4.6 vision extractor with Nemotron-Mini-4B-Instruct reasoning, then shows the original scan, validated overlay, defect counts, lab-style diagnosis, full JSON, and diagnosis history in a custom light-table interface.

Useful links:

• Live app: https://huggingface.co/spaces/build-small-hackathon/project-halide
• Demo video: https://huggingface.co/spaces/build-small-hackathon/project-halide/blob/main/assets/demo_walkthrough.mp4
• Vision model: https://huggingface.co/Lonelyguyse1/halide-vision
• Source: https://github.com/LonelyGuy-SE1/Project-Halide

The current checkpoint was trained offline on Modal, published with a Q4_K_M GGUF artifact, and validated on private held-out negatives that were not added to training data. Runtime inference stays inside the GPU-backed Space and does not call cloud inference APIs.