Quick follow-up to my last post: JoyCaption beta one in HybridScorer no longer uses the GGUF / llama.cpp path. It's now pure Transformers on CUDA, yessss!
Why I dumped llama-cpp:
- One less runtime to babysit on Windows
- Cleaner install — no separate GGUF builds, no CUDA/CPU wheel roulette
- It's the worst on installing, building the wheel took up to 15 !!! minutes. GONE.
- Even less VRAM needed with the nvfp4 version and slightly better quality.
Please give it a try: setup_update-windows.bat and you're done
For anyone who missed the first post ( https://www.reddit.com/r/StableDiffusion/comments/1sg5paj/built_a_tool_for_anyone_drowning_in_huge_image/ ) :
HybridScorer is a 100% local tool for cutting huge image folders down to the keepers. PromptMatch, TagMatch (anatomy errors), ImageReward, SamePerson (face), ObjectSearch (DINOv2), Similarity — point it at a folder, let it score, review the SELECTED / REJECTED split, export.
Since then, HybridScorer got a faster local FastAPI + Tabler UI, live WebSocket progress, media serving without raw file paths, recursive subfolder loading, drag-and-drop multi-select sorting, a seamless resizable image grid, full-size preview overlay, better PromptMatch/TagMatch score pills, 1.5-2x faster TagMatch on large folders, JoyCaption NF4 prompt generation, and LM Search running through Hugging Face Transformers instead of llama.cpp.
GitHub: https://github.com/vangel76/HybridScorer (GPL-3.0, Windows + Linux)