Pixel-Curator: Local CLI That Turns 1,000s of Raw Images Into LoRA-Ready Datasets
Impossible-Car-2809 · reddit · 2026-10-14
pixel-curator is an open-source, local-first CLI that turns a folder of raw images or video frames into a clean, captioned dataset ready for DreamBooth or Kohya LoRA training, optimized for identity/subject/style similarity training.
The pipeline: prefilter unusable frames → perceptual-hash dedup (removes most of video/burst-sourced sets) → CLIP embedding + k-means clustering with visualization of angle/lighting imbalance → optical quality scoring → CLIP semantic filtering (e.g., watermark removal) → WD14 tagging + Florence-2 captioning → HTML-report human review round-trip → one-command export to Colab DreamBooth or multi-concept LoRA directory structures.
Fully local, dry-run by default, resumable via SQLite manifest, runs on modest hardware. The author notes a well-curated 200–500 image subset beats a messy 2,000.
Related event: pixel-curator: open-source CLI for LoRA dataset prep(4 posts)→
More from Multimodal
- Open-source Fish Audio Drama 3 studio lets you design character voices in any language — camenduru · 2026-10-14
- WOVEN benchmark finds all 38 frontier MLLMs fail visual transition reasoning far below humans — MLL-Lab · 2026-10-14
- Low-poly previs plus Fal agent rendering with Seedream 5 and Seedance 2.5 — adamho · 2026-10-14
- Hands-on with PPT.AI: paste a link, get a 40-second concept video from your website — yihui_indie · 2026-10-14
- Astra: surreal AI-generated short film wows with gothic cathedral visuals — Extraaltodeus · 2026-10-14
- Video Made by Argon Shows Fourier Series Drawing a Curve With 7 Spinning Circles — Jaguar_2454 · 2026-10-14