A Visual Training Workbench for Diffusion Models
Dataset Management · AI Auto-Captioning · Visual Training Configuration · Live Monitoring · Post-Training Sampling
Built on DiffSynth-Studio
Run the entire diffusion LoRA fine-tuning workflow from your browser: no training scripts, no YAML to edit, no command-line flags to memorize. Every run's configuration is saved in full, so results are reproducible, comparable, and can be re-run at any time.
- Dataset management — Create image / video / audio datasets, import files, and maintain
metadata.jsonlalong with extra fields - AI auto-captioning — Connect any OpenAI-compatible multimodal model to generate or rewrite sample captions in bulk
- Visual training configuration — Pick a training recipe by model architecture, then configure LoRA, resolution, epochs, optimizer, and more
- Post-training sampling — Automatically generate samples with the latest checkpoint when training finishes, so you can verify results right away
- Live monitoring — Task status, launch command, streaming training logs, and loss curves
- Artifact management — Preview and download checkpoints, samples, and logs in the browser
Screenshots
Datasets and auto-captioning
Sample extra fields
Training configuration
Task monitoring
Prerequisite: Node.js ≥ 20 (used to build the web frontend; check with
node -v)
git clone --recurse-submodules https://github.com/modelscope/DiffSynth-WebUI.git
cd DiffSynth-WebUI
pip install -e DiffSynth-Studio/
pip install -e .Linux:
bash training_ui/launch.shWindows:
powershell -ExecutionPolicy Bypass -File .\training_ui\launch_windows.ps1On first launch, the script installs the frontend dependencies if node_modules is missing, then builds the Next.js frontend.
Default URL:
http://127.0.0.1:8100/dashboard
To use different ports:
NEXT_PORT=9000 BACKEND_PORT=9001 bash training_ui/launch.shTraining runs in a separate process, so closing the browser or exiting the launch script will not interrupt training. Just run the launch script again to resume monitoring task status.
- Launch and open
http://127.0.0.1:8100/dashboard - Training Datasets → Create Dataset, choose Image, and import a few dozen images
- Click Auto-Caption to generate captions in bulk (configure a captioning model in Settings first), or write prompts by hand
- New Task: pick a model → pick a GPU → pick the dataset you just created → leave everything else at its default
- Click Create Task, then open the task detail page to watch the logs and loss curve
- When training finishes, download
*.safetensorsat the bottom of the page and review the sampling results
See the Usage Guide below for a full description of every option.
For your first training run, we recommend going through the steps in order.
Open the Settings page and review the following:
| Setting | Description | Default |
|---|---|---|
| Dataset Directory | Where datasets and their metadata are stored | training_ui/data/datasets |
| Training Output Directory | Where task logs, checkpoints, and samples are stored | training_ui/data/outputs |
| Model Storage Directory | Base directory for downloaded models | models |
| Model Download Source | Where models are downloaded from: ModelScope or Hugging Face | ModelScope |
| Attention Implementation | Attention implementation; leave blank to select automatically | Blank |
To generate captions automatically, add a multimodal model under Captioning Models.
Open the Datasets page:
- Enter a dataset name
- Choose the Image, Video, or Audio type
- Click Create Dataset
- Open the dataset and import your sample files. Supported import methods:
- Single file
- Multiple files
- Archive (a
.txtfile with the same name as a media file is imported as its prompt)
- Write or edit the prompt for each sample, or use Auto-Caption to generate captions
- Add extra fields such as
edit_imageto each sample as required by the model you are training
See Dataset Format for the on-disk directory layout.
Open Tasks and click New Task:
- Task Information — Enter a task name and select a GPU
- Model — Choose the Model Architecture and Target Model
- Model Paths — Use the default model repositories, or enter local model paths; FP8 can be enabled for some models
- LoRA Settings — Use the default LoRA target modules, or enable custom targets and set the rank
- Dataset — Choose the dataset, repeat count, and resolution
- Training — Set epochs, learning rate, and optimizer; enter a checkpoint if you want to resume training
- Sampling — For models that support post-training sampling, add prompts and any other input conditions
- Review the Launch Command, and create the task once the final command looks right
Resolution supports two modes:
max_pixels: caps the total pixel count as the product of input height and width, e.g.1024 * 1024 = 1048576 px.height × width: fixes the exact height and width of the training input.
You can start the task immediately when creating it, or save it and start it manually from the task list later. A task can be run multiple times, and each run creates its own output directory.
The task detail page shows:
- Current status and elapsed time
- The actual launch command and the full task configuration
- Streaming training logs
- The loss curve
- Checkpoints, final samples, and other files produced by the current run
You can stop a task manually while it is running; after stopping, you can adjust its configuration and start it again.
The default output layout:
training_ui/data/outputs/
└── <task-id>_<task-name>/
└── <run-timestamp>/
├── training_config.json
├── train.log
├── loss.csv
├── step-*.safetensors
└── final_samples/
training_config.json: the user configuration, resolved settings, and actual command for this runtrain.log: training stdout and stderrloss.csv: training metrics stored asstep,key,value*.safetensors: the trained model filesfinal_samples/: validation samples generated after training finishes
Each dataset is a self-contained directory under Dataset Directory. Files belonging to extra fields are all kept under _fields/,
which the UI manages for you:
training_ui/data/datasets/<dataset-name>/
├── metadata.jsonl
├── 001.jpg
├── 002.jpg
└── _fields/
└── 002/ # sample name (without extension)
└── edit_image/ # extra field name
└── input_image.jpg
One sample per line in metadata.jsonl:
{"file": "001.jpg", "prompt": "a cat sitting on a sofa"}
{"file": "002.jpg", "prompt": "make the sofa red", "edit_image": "_fields/002/edit_image/input_image.jpg"}
No manual cropping or resizing required — samples are processed automatically at the configured resolution during training.
- Anima
- Boogu-Image
- ERNIE-Image
- FLUX.1
- FLUX.2
- HiDream-O1-Image
- Ideogram 4
- JoyAI-Image
- Krea 2
- Qwen-Image
- Stable Diffusion
- Stable Diffusion XL
- Z-Image
- LingBot-Video
- LTX-2
- MiniMax-H3
- MOVA
- Wan
- ACE-Step
For each model's capabilities, VRAM requirements, and training tips, see the DiffSynth-Studio documentation.
DiffSynth-WebUI and DiffSynth-Studio are maintained by the same team: Studio provides the model implementations and the training framework, while the WebUI brings the whole training workflow into the browser. If you are more comfortable with the command line and Python scripts, use DiffSynth-Studio directly.
This repository already contains the full DiffSynth-Studio codebase and keeps merging upstream updates, so installing this repository is all you need — there is no need to install DiffSynth-Studio separately.