LoRA Dataset Studio v2026.08.10.1

Two ways to install — and they do not update the same way

git clone is the one to prefer. Update & restart then runs git pull --ff-only,
so you get every fix the moment it lands — often days before it is packaged into a
release like this one.

This ZIP is a snapshot of this tag. Update & restart will only ever move you to
the next release, so a fix shipped today reaches you whenever the next one is cut.

git clone https://github.com/perfectgf/lora-dataset-studio.git
cd lora-dataset-studio
start.bat

Already running a clone? You do not need the ZIP below.

🎁 What's new in v2026.08.10.1

Run the watermark detector on a GPU Python you already have

The 🚩 Find scan quietly ran on the CPU even on machines with a perfectly good graphics card: its environment ships a CPU-only PyTorch on purpose, and nothing ever said so. The Bank’s Watermarks panel now tells you when the scan will run on the CPU and offers the same one-click picker ✨ Score already has — point the detector at the Python ComfyUI or ai-toolkit already uses and the scan runs about ten times faster, downloading nothing. The pass summary now also says which device it actually ran on.

Re-run passes clear their old red report line — and instant passes say they finished

Two silences fixed. A "Launch all" stopped early kept announcing "cancelled before it ran" over passes you had since re-run successfully — the re-run is now recorded for every pass, so the report marks the line as re-run since. And a pass that finishes in a couple of seconds (everything already cached) used to end without a progress bar, a toast, or any sign it had worked; its completion message now always arrives.

Group by person now recognises much smaller faces

Point the Bank at a general photo library and most of it used to come back ungrouped: the pass only kept a face filling at least a quarter of the frame, so full-body shots, group photos and anything taken a few steps back were skipped as "face too small" and never joined their person. The size test is now a real measurement of the face in pixels rather than a share of the image, which is what identification actually depends on — the same head is the same head whether the photo is 1 or 24 megapixels. Faces below 64 pixels are still left out, because that is where the recognition itself stops being trustworthy and starts merging different people. Banks you already scanned keep their existing groups; photos you add from now on, and any bank scanned from here, use the new floor.

Bokeh portraits are no longer flagged as blurry

The 🌫 blurry flag measured focus across the whole frame, so a shot with a razor-sharp subject on a creamy out-of-focus background averaged out as soft — and the pictures you most wanted to keep were the ones getting flagged. Sharpness is now read region by region and scored on the sharpest parts of the image, so the flag means what it always should have: even the crispest area of this photo is soft. It also stops the reverse mistake, where one blazing block — a watermark, a compression artefact, a burnt highlight — vouched for an otherwise soft frame. Scores are on a new, higher scale; run the Quality pass again on an already-scanned bank to pick them up. Thanks to the community for reporting it.

Full changelog: v2026.08.10...v2026.08.10.1