ComfyUI Workflow: ComfyUI LoRA Training (In-Comfy)

LoRA training is the on-ramp skill every ComfyUI user eventually needs. Custom characters, product styles, specific art directions, the only way to teach a base model a new concept is to train a LoRA on it. The traditional approach is a separate Python environment, the Kohya_ss command line, or a standalone web UI. This workflow skips all of that by running the training as a ComfyUI custom node.

The LarryJane491 Lora-Training-in-Comfy custom node bundles Kohya's sd-scripts as a dependency, exposes the training config as ComfyUI node widgets, and runs the training in the same Python environment as ComfyUI itself. Point it at a folder of captioned images, set a name, queue the prompt, and come back to a LoRA in your loras folder.

Key benefits

• Trains SD 1.5 and SDXL LoRAs in ComfyUI, no separate environment.
• Output LoRAs drop into models/loras automatically, ready to test.
• Includes an Access Tensorboard node for monitoring loss curves.
• Pairs with the LJRE/LoRA Caption Load + Save nodes for in-Comfy captioning.
• 543 stars on GitHub, actively maintained.

How it works

1. Install the custom node: clone the repo into your custom_nodes folder, run pip install -r requirements_win.txt (or the Linux version).
2. Prepare your training data: put captioned images into a folder named like 5_my_concept (the number is repeat count, the name is for your reference). Put that folder inside a parent folder.
3. Set the data path in the Training node: the parent folder path goes in. The node scans it for subfolders matching the [number]_[name] pattern.
4. Configure the training: set a LoRA name, learning rate, steps, network rank. The defaults work for SD 1.5 character training.
5. Queue the prompt. Training runs in the same console where ComfyUI is running. Watch the loss output.
6. Test the LoRA: refresh your LoRA list in ComfyUI, add it to any generation. The LoRA is in models/loras ready to use.

The Advanced node adds more controls (network alpha, training scheduler, mixed precision) for users who need them. The basic Training node is enough to get a first LoRA working. Captions matter more than the training config for the first version: clean, consistent captions beat a fancy scheduler every time.