Osmium Workshop
A feature-heavy, ComfyUI-compatible dataset manager built around Anima (Circlestone Labs'
diffusion model) and the booru-tag models in its family. Other models work too, but the tag
conventions, categories and generation workflow all follow Anima. Curate a folder of images, tag
them, grow the dataset with generated images, and train a LoRA on it, all in one portable app.
Built with Electron + TypeScript, and with a shitload of Claude and Opencode Go.
Disclaimer: as obviously stated, nearly all of this is AI-built. If for whatever reason that
somehow makes you grossed out when the whole point of a dataset curator is to train genAI, then
I don't know go eat a rock or something.
Docs: the in-app ❓ Help covers every tab · User guide ·
Patch notes
Download
Everything is on the releases page.
| File | What it is |
|---|
Osmium.Workshop-<version>-win.zip | The dataset manager (Windows, portable) |
Anima-TrainFlow-Osmium-<version>.zip | The trainer for the Trainflow tab (optional) |
Comfy.Bridge-<version>-win.zip | Comfy Bridge: a simple generation UI for ComfyUI (Windows) |
ComfyBridge-<version>-android.apk | Comfy Bridge for Android |
ComfyUI-DataSetManagerNodes.zip | The ComfyUI node pack SynthDat and Comfy Bridge need |
Install
Osmium Workshop. Unzip anywhere and run Osmium Workshop.exe. Nothing is installed; your
settings live in the data/ folder beside it. Open a dataset with File → Open dataset folder
(images plus matching .txt captions).
Trainflow (training LoRAs, NVIDIA GPU).
- Unzip
Anima-TrainFlow-Osmium-<version>.zip somewhere with about 8 GB free.
- Run
Install_Requirements.bat in it once. It downloads PyTorch and the trainer's packages
(about 5 GB) at the exact versions Trainflow is tested with.
- In Osmium's Trainflow tab, choose that folder, then pick your Anima (or Janima) DiT, Qwen3
text encoder and VAE.
This is Anima-TrainFlow by ThetaCursed with a newer
sd-scripts (v0.12.0) and the patches Trainflow needs; it's the version Trainflow is tested with.
ComfyUI features (SynthDat, Comfy Bridge). Unzip ComfyUI-DataSetManagerNodes.zip into your
ComfyUI's custom_nodes folder. Both can use a running ComfyUI, or (on Windows) start your install
themselves as Osmium Comfy, without a server. WD14 tagging runs on-device and doesn't need ComfyUI.
Comfy Bridge. Unzip and run, or sideload the APK on your phone. The phone can use a ComfyUI server
directly, or go through the desktop Bridge (Let the phone app use it). For extra samplers
(res_2m, res_2s and friends), install the separate RES4LYF
node pack in your ComfyUI (for example through ComfyUI Manager); Bridge lists them once it's there.
Updating. There's no auto-updater: unzip the new version and move your old data/ folder into
it. Comfy Bridge's Android app updates install over the old APK.
Major features
- Gallery & tagging. Grid, Single and Sequential views; tags as editable chips, sorted into
categories; batch add/remove/rename, find & replace, Master Tags across the dataset, and
Merge/Void with standing rules that keep catching stray tags.
- WD14 autotagger. Runs on your PC (no ComfyUI needed), with a review window to accept or reject
each tag before it lands.
- Curate tab. Go through a folder of new images first: grid or one at a time, duplicate finder,
low-resolution flags, sort them into artstyle folders or junk, then add the keepers to a dataset.
- Image editing. Crop, rotate, Isolate and Isolate (Split) for cutting up sheets and panels; every
edit can be undone.
- Bucket Images. Crops every image to its training bucket around the subject, keeping the
originals. Sizes are per dataset, with presets for 512/1024-style training.
- Trainflow tab. Trains a LoRA, LoKr or LoHa for Anima/Janima on the loaded dataset in the
background. Continue a run, send a checkpoint to your LoRA folder, and every retrain gets its own
folder.
- SynthDat Overseer. Generates new dataset images from your existing ones through ComfyUI, for you
to accept or reject.
- Edit Log. Every change is logged with undo/redo, per image or for the whole edit.
- Comfy Bridge. A no-node-graph generation UI for the same workflow, on Windows and Android: LoRA
stack, 2-pass, USDU upscaling, a gallery, image/LoRA metadata, and a phone link to your PC.
- Themes. 26 full themes (5 free, the rest bought with Edibits you earn from achievements), Theme
Studio for your own, and night mode.
License
MIT, see LICENSE. Third-party code, fonts and models are listed in
THIRD_PARTY_NOTICES.md.
