Release v2026.07.14 — Beta · perfectgf/lora-dataset-studio

🧽 Automatic watermark cleanup

Scraped images love to carry site logos, URLs and usernames — and a LoRA happily learns them. New in the Curate step:

Find watermarks — a local vision pass flags overlaid marks (scene text like shop signs is explicitly excluded) with the detected box stored per image.
Clean — each mark gets the least risky tool: a lossless crop for border marks, non-generative inpainting (LaMa) for fine text elsewhere; anything big or on the subject is left for manual review. Originals are preserved next to the files.
🔍 Review mode — a fullscreen viewer walks every flagged image with the detected box drawn on it: clean it and see the result in place, mark it not a watermark (never re-flagged again), or reject the image — keyboard shortcuts included.
⬇ One-click inpainting install — the LaMa extra installs right from the watermark tools, with live progress, no restart.

The capabilities list now includes Watermark inpainting (no more "all ready" while something is missing), and every ML capability has its own card with Install / ↻ Reinstall (scoped pip actions — no more re-running the whole extras step).
Machine-scan rows are clickable — each capability jumps to the step that installs it.
Ollama detection no longer requires the server to be running: installed-but-stopped shows as such, with a ▶ Start Ollama button (Settings and Setup).

Local Klein generation no longer needs any custom node packs — the shipped graph is now 100% ComfyUI core (it silently required two third-party packs; first generation on a fresh install crashed with a raw ComfyUI error). A node preflight also pre-checks the graph and names anything missing with install links. Thanks to thomas00no on Discord for the report.
Test Studio (previous fixes rounded out): per-family preflight for missing models/nodes, failed tiles show why, SDXL saves through a standard SaveImage node.

⎘ Share configuration — every run in the Runs hub exports a paste-safe text file with the full recipe (family, base, all ai-toolkit parameters, outcome) — no local paths, no keys. Perfect for sharing settings or asking for help.

Batch progress survives a page reload — captioning, watermark scans, face analysis and framing classification advertise their live progress server-side; the indicator comes back on its own.
Honest Clean toasts (one recap instead of contradictory messages), and the docs/README carry a full "Everything it does, at a glance" overview.