Skip to content

About

A windows/linux app for training with sd-scripts.

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Repository files navigation

LoRA Studio icon

LoRA Studio

A friendly Windows & Linux app for training LoRAs with kohya-ss/sd-scripts:
set up a run, watch the loss, tidy your dataset, and send the result to SwarmUI or ComfyUI, all in one place.

LoRA Studio home screen


What it does

Training

  • Any sd-scripts training script: SD 1.x/2.x, SDXL, SD3, FLUX, Anima, Lumina and more. The form is built from the script's own options, so new options appear automatically when sd-scripts updates.
  • Simple and full views: Simple shows the settings that matter; All options lists every one, with search.
  • Queue: line up several runs and walk away. Watch live loss, speed, time left and the log.
  • Run history: every run keeps its exact config, sd-scripts version, loss curve and log. Re-run it, tweak it, or resume from a saved state.
  • Sample previews: your prompts are rendered every epoch and shown as a prompt × epoch grid. Compare any two with a slider or fade, or turn a prompt's samples into an animated GIF.
  • Presets: save your favourite setups, or import existing training.toml / dataset_config.toml files. Pick a default preset that every New session starts from, and share presets as small files (your own paths are left out).
  • Time estimates: see roughly how long a run will take before you start; LoRA Studio learns your GPU's speed from past runs.
  • Best epoch: star the epoch you like in the sample grid; it's saved as …-best.safetensors and can be copied to SwarmUI or ComfyUI in one click.
  • GPU monitor: VRAM, load, temperature and power while you train.
  • Notifications: a desktop notification when a run finishes or fails.
  • Check before training: wrong values, missing models or empty image folders are caught in seconds, before the model loads.
  • Overnight batches: let the PC sleep or shut down when the queue is done (with a 1-minute countdown you can cancel).

Dataset tools

  • New dataset in one click: download with gallery-dl (Danbooru, Gelbooru, e621, pixiv and hundreds more), skip tiny images, number everything, turn the site's tags into captions, and check for duplicates. Site logins and API keys are stored on your PC only.
  • Images & tags: caption editor with tag chips, autocomplete, filters, keyboard shortcuts and bulk add/remove/replace with undo. The tag overview shows every tag by how often it's used (sort by rarest to spot typos), and renames or removes it everywhere.
  • Auto-tagging: with the WD EVA02 tagger.
  • Crop & resize: a Photos-style crop window, plus batch resize to a long side, short side or exact size (e.g. 1024×1024, 832×1216). Originals are always kept.
  • Duplicates: finds look-alike images with CLIP; you review before anything moves, then delete them for good with one click.
  • Health check: catches missing/empty captions, unreadable images, caption clashes, missing trigger words and odd sizes, and estimates buckets and steps.
  • Rename: numbered names with a preview first.

SwarmUI & ComfyUI

  • Start them from the app: use your existing SwarmUI and/or ComfyUI install, and open either one right inside LoRA Studio.
  • ComfyUI two ways: the ComfyUI that runs inside SwarmUI (no extra setup), or a separate ComfyUI install (portable or git).
  • Copy LoRAs over: finished LoRAs go straight into SwarmUI or ComfyUI. Title, trigger word and thumbnail are stored in the file, where SwarmUI's model browser shows them.
  • Test grids: render epoch × strength grids through SwarmUI.

Looks

  • Modes and accents: light and dark modes, plus 10 pastel accent colours.
  • Custom header: a patterned header (checker, gingham, sparkles…) or your own art.
  • Your own styles: a custom.css file for anything else.
Train page Queue with live loss chart
Images and tags editor Crop window
Sample grid on a run's page Appearance settings

Requirements

  • Windows 10 or 11, or Linux (see Linux below). On Windows the app window uses Microsoft Edge WebView2, which is built in.
  • Python 3.10 or newer, from python.org. Tick "Add python.exe to PATH" when installing.
  • Git, from git-scm.com.
  • sd-scripts. If you don't have it yet, LoRA Studio can install it for you.

LoRA Studio itself is small. All the heavy machine-learning parts (PyTorch etc.) run inside your sd-scripts install, so nothing big is downloaded twice.

Install

On Linux? Skip to Linux.

  1. Download this project: the green Code button → Download ZIP, or grab the latest from Releases. Unzip it somewhere permanent, e.g. C:\AI\lora_studio.
  2. Double-click Install.bat. It creates a private Python environment in the folder and installs two small packages (pywebview and Pillow).
  3. Double-click LoRA Studio.bat to start. Optionally run Create desktop shortcut.bat once for a desktop icon.
  4. On first launch, point it at your sd-scripts folder (the one with train_network.py), or let it install a fresh copy.

Updating

The easy way: when a new version is out, a small download button with a dot appears in LoRA Studio's sidebar (and in Settings → About). Click Update now: LoRA Studio downloads the new version, keeps your .venv, your custom.css and all your settings, runs and presets, installs anything new, and restarts with one click.

By hand: download the new zip and unzip it over your LoRA Studio folder (keep the .venv folder), or Pull in GitHub Desktop / git pull if you cloned it. Then start LoRA Studio as usual.

Using it

Page What it's for
Home What's training, what's queued, recent results, sd-scripts status
Train Pick a script on the left, fill in the form, then Train now or Add to queue
Queue Live progress, loss chart and log; reorder or remove waiting runs
Runs Past runs: loss curve, samples, saved LoRAs and their info, resume, config and log
Dataset New dataset (gallery-dl), images & tags, health check, duplicates, crop & resize, rename
SwarmUI & ComfyUI Start/open either one, use it inside the app, and see the LoRAs you've sent it
sd-scripts Update, switch branch, roll back
Settings Appearance, launcher options, notifications, folders, backup & restore

Folders LoRA Studio creates in your dataset

  • _originals/ keeps the original of any image you crop or resize.
  • _deleted/ holds images deleted in the viewer, with their captions. You can restore them.
  • _duplicates/ holds images moved by the duplicate finder.

sd-scripts ignores all three folders.

Downloading with gallery-dl

Dataset → New dataset → Install gallery-dl (one time). Pick a site and type tags, or paste any link gallery-dl supports.

  • Site logins: sites that need a login or API key are set up under Site logins.
  • pixiv: press Get token and follow the window that opens.
  • Other logged-in sites: gallery-dl can use your browser's cookies; Firefox works best.

Downloads go to the downloads folder in LoRA Studio's data folder (%APPDATA%\LoRA Studio on Windows, ~/.local/share/lora-studio on Linux) unless you pick a folder.

Backup & restore

Settings → Folders & backup → Create backup saves your settings, presets, themes and run history in one zip. Your datasets, trained LoRAs and site logins aren't included.

Resuming a run

Turn on Save state (Train → Output) before training. If a run is stopped or crashes, open it on the Runs page and press Resume.

SwarmUI & ComfyUI

  • SwarmUI: choose your SwarmUI folder on the SwarmUI tab. LoRAs are copied to <SwarmUI LoRA folder>/LoRA Studio/<name>/. Test grids need SwarmUI running; the Start SwarmUI button launches it.
  • ComfyUI: on the ComfyUI tab, pick The one inside SwarmUI (nothing else to set up) or A separate ComfyUI install. For a separate install, choose its folder: the portable folder with run_nvidia_gpu.bat, or a git install with main.py. LoRAs are copied to models/loras/LoRA Studio/<name>/; refresh ComfyUI's LoRA list to see them.

Making it yours

  • Settings → Appearance: pick a mode, an accent and a header, then Save as theme.
  • web/css/custom.css: loaded last, so anything you put there wins. You can add accents, change roundness or fonts, and more.

For developers

The app is plain Python plus HTML/CSS/JS with no build step.

main.py                 starts the window (or --browser for a normal browser tab)
install.sh, lora-studio.sh, create-desktop-entry.sh   the Linux versions of the .bat files
tests/                  quick checks (python -m pytest tests) + a browser smoke test, run on every pull request
studio/                 Python backend
  api.py                every function the UI can call
  sdscripts.py          find / install / update / switch / roll back sd-scripts; read a script's options
  _schema_dump.py       runs inside the sd-scripts venv to read argparse options
  trainer.py            queue, launching, live log + loss parsing
  runs.py               run history, resume points, outputs
  dataset_tools.py      captions, bulk edits, rename, crop/resize, delete/restore, autocomplete
  health.py             dataset health check + bucket/step estimate
  samples.py            sample prompt files and sample-image discovery
  metadata.py           read/write .safetensors metadata (no torch needed)
  swarm.py              SwarmUI: detect, launch, copy LoRAs, test grids
  gallerydl.py          gallery-dl: install, site logins, downloads
  importer.py           downloaded folder → clean numbered, captioned dataset
  estimate.py           training-time estimates from past runs
  gpu.py / notify.py    nvidia-smi stats / desktop notifications
  backup.py             backup & restore of LoRA Studio's own data
  comfy.py              ComfyUI: inside SwarmUI or a separate install
  workers/              tagger + duplicate finder; these run inside the sd-scripts venv
web/                    the UI
  css/tokens.css        theme colours (light/dark + accents)
  css/app.css           components
  css/custom.css        user overrides
  js/pages/*.js         one file per page
  js/tools/*.js         dataset tools

Run python main.py --browser to develop in a normal browser with dev tools.


Credits

  • kohya-ss/sd-scripts does all the actual training.
  • SwarmUI by mcmonkey, and ComfyUI.
  • The WD EVA02 tagger and OpenAI CLIP models are used by the dataset tools and downloaded from Hugging Face on first use.
  • pywebview provides the app window.
  • gallery-dl handles downloads (installed on request).

LoRA Studio is an independent project and isn't affiliated with kohya-ss, SwarmUI, ComfyUI or Stability AI.

License

MIT

About

A windows/linux app for training with sd-scripts.

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Contributors

Languages