Download

Open Releases Download Page

SD-reScripts

✨ Enjoy Stable Diffusion Train！ ✨

v1.5.4

Fork from 秋葉 aaaki/lora-scripts
Modify By Lulynx

Download · Documents · 中文README · AGENTS · FRONTEND

SD-reScripts is a maintained fork / continuation of LoRA-scripts (a.k.a. SD-Trainer).

This is an experimental project currently in beta, and there are tons of bugs.

LoRA & Dreambooth training GUI & scripts preset & one key training environment for kohya-ss/sd-scripts

• improved Launcher graceful stop flow, duplicate-stop protection, and clearer Stopping... task state feedback
• fixed failed trainer startup leaving task slots occupied and improved Windows stop / kill behavior
• fixed multiple training issues across Anima, Aesthetic Scorer, and Newbie resume / checkpoint flows
• hardened community UI installation with safer GitHub validation, manifest handling, and zip extraction checks

A dedicated desktop launcher is now included for runtime setup, launch control, runtime diagnostics, managed preset import, and safer day-to-day startup flow.

The REAL Stable Diffusion Training Studio. Everything in one WebUI.

Follow the installation guide below to install the GUI, then run run_gui.ps1 (Windows) or run_gui.sh (Linux) to start it.

Tensorboard	WD 1.4 Tagger	Tag Editor

A redesigned community UI is also supported through the frontend profile system. You can switch to it from the launcher or your configured frontend profile workflow.

If you use the bundled portable Python runtime in the repo root, system Python is optional.

Otherwise you need:

• Python 3.10+
• Git

git clone --recurse-submodules https://github.com/WhitecrowAurora/lora-rescripts.git

Run run_For_≤RTX40series.bat or run_For_SageAttention_Experimental.bat.

• If a ready-to-run python folder already exists in the repo root, the installer will use it first
• Otherwise it falls back to creating a virtual environment
• setup_embeddable_python.bat is now mainly a repair helper for broken raw embeddable Python installs, not a normal first-run requirement

run run_gui.ps1, then program will open http://127.0.0.1:28000 automanticlly

If you want to try sageattn, there are now dedicated experimental startup scripts on Windows:

• run_For_SageAttention_Experimental.bat: general SageAttention runtime for NVIDIA GPUs
• run_For_NVIDIA_SageAttention_Experimental.bat: compatibility alias for the same general SageAttention runtime
• run_For_Only_Blackwell_SageAttention_Experimental.bat: recommended experimental path for RTX 50 / RTX PRO Blackwell users when xformers is unreliable

Notes:

• the first run will automatically prepare a dedicated runtime and keep the main python / python_blackwell / xformers environments untouched
• SageAttention only affects routes and configs that explicitly enable sageattn; launching with a SageAttention script does not force every trainer to stop using sdpa or xformers
• you can verify the runtime with check_sageattention_env.bat or check_sageattention_env.bat --blackwell
• if you want to provide a prebuilt local wheel, place it in sageattention-wheels or sageattention_wheels
• for the Blackwell runtime, wheel names containing blackwell or sm120 are preferred automatically

Current validated experimental base stack:

• Python 3.11.9
• Torch 2.10.0+cu128
• TorchVision 0.25.0+cu128
• Triton Windows 3.5.1.post24
• SageAttention 1.0.6

Run install.bash.

• if python/bin/python already exists, the installer will use it first
• otherwise it will use venv/bin/python if present
• otherwise it will create venv automatically unless you explicitly pass --disable-venv
• it now installs the same base PyTorch / dependency stack as the current Windows installer

Run bash run_gui.sh, then program will open http://127.0.0.1:28000 automatically.

• run_gui.sh now auto-detects python/bin/python, venv/bin/python, or system python
• if base dependencies are missing, it will run install.bash for you
• if tag editor dependencies are missing and the current Python is compatible, it will run install_tageditor.sh
• for mainland China mirror settings, use bash run_gui_cn.sh
• on Windows, use run_gui_cn.bat, run_auto_cn.bat, or run_manual_cn.bat
• dedicated experimental routes also provide matching _cn.bat launchers
• the first CN startup will let you choose a PyPI mirror; pressing Enter keeps the default Tsinghua preset and saves it to config/china_mirror.json

TensorBoard is already integrated into the GUI startup path.

The launcher Managed tab can connect to a hosted preset site for one-click preset import, rollback, and 24-hour local cache sync.

Reference repository:

• WhitecrowAurora/lulynx-lora-share

Recommended Linux prerequisites for the hosted preset site:

• git
• Node.js 20+
• npm 10+
• if native modules need local compilation: build-essential, python3, pkg-config, libvips-dev

git clone https://github.com/WhitecrowAurora/lulynx-lora-share.git
cd lulynx-lora-share

Install backend dependencies:

cd backend
npm install

Install frontend dependencies:

cd ../frontend
npm install

Run the backend locally:

cd ../backend
PORT=3000 CORS_ORIGIN=http://127.0.0.1:5173 npm run start

Run the frontend in dev mode:

cd ../frontend
VITE_API_URL=http://127.0.0.1:3000/api npm run dev -- --host 0.0.0.0 --port 5173

Create a production frontend build:

cd frontend
VITE_API_URL=https://your-domain.example/api npm run build

Then configure your reverse proxy to serve frontend/dist and forward /api to the backend server.

After the site is online, create an API key in LORA Share and paste the server URL + API key into the launcher Managed tab.

This project stands on the work of multiple open-source communities. Respect and thanks to:

• aaaki/lora-scripts: direct upstream fork base for this repository.
• Akegarasu/lora-scripts: earlier foundation of the script-first and GUI workflow.
• kohya-ss/sd-scripts: core training backend used by this project.
• kozistr/pytorch_optimizer: optimizer and scheduler collection used for the extended optimizer/scheduler options in this project.

Special thanks to

DrRelax599