HaDeS-Booru v1.1.2 — the Booru-to-LoRA pipeline updated with new features.

If you've ever built a LoRA dataset the "manual" way — half a dozen tabs in Grabber, a gallery-dl config you can't remember setting up, and a Notion doc of prompts you copied by hand — you'll know why this exists.

HaDeS-Booru is a small local web app that does one job end to end: search a booru, pull the images, clean the prompts, and hand you a dataset ready for Kohya. It runs on your machine, talks to the booru APIs, and the parts that should be offline (tag dictionary, fonts) actually are.

Since I shipped v1.0.0, a lot landed. Here's what you actually get in v1.1.2:

• 28 booru sources, 7 API families. Danbooru, Gelbooru/DAPI, Moebooru, E621/E926, Philomena (Derpibooru, Furbooru), Zerochan, and a stack of art platforms. Pick one, search, done.

• SFW by default, and it filters on the server. Where a board supports rating in its API, the G/S/Q/E filter goes out with the query instead of pulling everything and hoping. Where it doesn't, it filters locally as a backstop.

• Offline Tag Dictionary. A local SQLite catalog of the ~1,000 most popular Danbooru tags, with definitions and alias resolution (miku → hatsune_miku). English definitions are stored as-is; Spanish shows up only when there's a genuine translation — I'm not going to paste a machine-translation and call it official.

• Your own Blacklist & Whitelist. Add tags you never want, or force-keep ones the defaults would drop (e.g. commission). Wildcards work. There's even a one-click button to blacklist every filtered tag on the current post at once.

• Image to Prompt (my favorite). Drag in a generated PNG/JPEG/WebP and it pulls the prompt straight out of the file, 100% locally. It reads A1111 parameters, ComfyUI JSON graphs, NovelAI chunks, Civitai XMP, and booru EXIF/XMP comments — then gives you three clean, copyable boxes. You can drop a new image on top of an existing one without clearing first.

• Recent searches + Top tags. Your last queries are one click away, and the tags you reuse most sit right there to build a query faster.

• Light / Dark / System theme, with the fonts bundled locally so nothing reaches out to Google.

• Sensible defaults. It starts on port 7863 (not 7860, so it won't fight your SD WebUI), and if that's taken it walks to the next free port and just launches.

• Export that's ready for training. TXT + a ZIP trainer archive with your normalized prompts, plus an errors.txt manifest if any image download failed — nothing silently vanishes.

How to install? Grab the last update from the GitHub releases: one zip, intall Python 3.11+, a single pip install -r requirements.txt, and it's up. I've been using it on my own datasets, and it's the part of the LoRA workflow I used to hate least.

And a thank you — to everyone who's tried it, and especially to the ones who filed bugs and sent feature ideas. A surprising chunk of what's landed since v1.0.0 came straight from your reports. If something's broken or a source is missing, the fastest way to fix it is to point it out. The roadmap is basically a to-do list of your DMs.

Github repo: https://github.com/Devildonia/HaDeS-Booru

v1.0.0 article: https://civitai.red/articles/36901/i-got-tired-of-fighting-with-grabber-and-gallery-dl-so-i-built-hades-booru

HaDeS.