LoRA Dataset Studio
You need a ComfyUI install for local generation and checkpoint tests, but not its node editor: LoRA Dataset Studio (LDS) sends its own workflows for dataset shots, edits, upscaling and checkpoint tests, and Setup downloads the generation models. Importing, curating, captioning and ai-toolkit training on the official Hugging Face bases work without ComfyUI. With Docker, the GPU launcher can set up a fresh, isolated ComfyUI for you.
| Feature | Needs ComfyUI |
|---|---|
| Local Klein and Krea 2 Edit generation, reference edits | Yes |
| Test Studio, deploying a trained LoRA into ComfyUI | Yes |
| Klein watermark cleaning; the SeedVR2, Klein Improve and Camera angles plugins | Yes |
| SDXL training, which uses a base checkpoint from ComfyUI's model folders | Yes |
| Import, scraping, curation, manual captions, export, backup | No |
| Auto-captioning with Ollama, LM Studio or JoyCaption | No |
| API image engines (Gemini, ChatGPT, OpenRouter) | No |
| ai-toolkit training on official Hugging Face bases, cloud training | No |
In Settings → Local tools → ComfyUI, set the ComfyUI API URL (default http://127.0.0.1:8188) and the ComfyUI install directory, the folder that contains models/, input/ and output/. Test confirms the connection. LDS reads extra_model_paths.yaml the same way ComfyUI does, so models kept in other folders are found.
The Setup wizard checks that folder as you type. You can also continue without ComfyUI: Setup lists what turns off and what stays on, and remembers the choice until you enter a directory.
When a Test Studio family needs nodes your ComfyUI lacks, Fix missing ComfyUI nodes gives update steps for portable, Desktop and manual installs. LDS guides the update; it does not run ComfyUI's updater or restart it for you. Krea 2 Edit also needs its own custom node pack and model files, listed in the settings reference.
-
start-docker.batruns LDS in Docker and connects to the ComfyUI already on your host; it asks for its folder once. -
start-docker-gpu.batcreates a fresh, isolated ComfyUI with its own model and data folders beside the checkout, and never touches an existing ComfyUI by default. Budget about 20 GB of disk before model downloads. - The Docker lane without a GPU has no ComfyUI, so local generation, the Test Studio and local training are unavailable there.
- No Docker image includes local ai-toolkit training; connect ai-toolkit on the host or train in the cloud.
- If ComfyUI runs in another container, in WSL or on another machine, its
input/andoutput/folders must be visible to both sides at the same path, because LDS hands images to ComfyUI through the filesystem. - ComfyUI's port is not protected by the LDS access token. Keep it on loopback, or behind a firewall, VPN or authenticated reverse proxy.
- ComfyUI paths are tested extensively against a mocked API, but not every model and custom-node combination has run on live installations. A failed preflight names the missing asset.
- Local generation typically needs about 16 GB of NVIDIA VRAM.