ComfyUI Extension: ComfyUI-TrainTools-MZ

ComfyUI-TrainTools-MZ Introduction

ComfyUI-TrainTools-MZ is an extension designed to facilitate the fine-tuning of LoRA (Low-Rank Adaptation) models within the ComfyUI environment. This extension leverages powerful training tools such as kohya-ss/sd-scripts to streamline the process of model training and customization. For AI artists, this means you can easily enhance and adapt your AI models to better suit your creative needs, without needing deep technical expertise.

How ComfyUI-TrainTools-MZ Works

At its core, ComfyUI-TrainTools-MZ provides a set of nodes that integrate with ComfyUI, allowing you to set up, configure, and execute training workflows for LoRA models. Think of these nodes as building blocks that you can connect to define how your model should be trained. Each node has specific functions, such as initializing the training environment, copying images, configuring training parameters, and running the training process. By connecting these nodes in a sequence, you can create a customized training pipeline that fits your specific requirements.

ComfyUI-TrainTools-MZ Features

MZ_KohyaSS_KohakuBlueleaf_HYHiDInitWorkspace | MZ_KohyaSS_KohakuBlueleaf_HYHiDLoraTrain

These nodes are used to initialize and train LoRA models using the HunYuanDiT branch of the sd-scripts. They set up the necessary workspace and execute the training process.

MZ_KohyaSSInitWorkspace

This node initializes the training folder within the output directory. You can specify:

• lora_name: The name of the LoRA model, which will be used to create the training folder.
• branch: The branch of the sd-scripts to use, defaulting to the current branch used during code debugging.
• source: The source of the sd-scripts, defaulting to GitHub. If there are download issues, you can switch to an accelerated source.

MZ_ImagesCopyWorkspace

This node copies images to the training folder and configures the dataset. Key options include:

• images: A list of images to be copied. It's recommended to use the upload folder node from the .
• force_clear: Whether to clear the original folder content before copying images.
• force_clear_only_images: Whether to clear only the image folder content.
• same_caption_generate: Whether to generate the same annotation file for all images.
• same_caption: The content of the same label to be generated.

MZ_KohyaSSUseConfig

This node provides basic training configurations. For detailed field references, see the .

MZ_KohyaSSAdvConfig

This node offers more advanced training configurations. Again, refer to the for detailed field references.

MZ_KohyaSSTrain

This is the main training node. It includes options such as:

• base_lora: Load a base LoRA model for training, consistent with the network_weights parameter in sd-scripts.
• sample_generate: Enable example image generation each time the model is saved.
• sample_prompt: The prompt used for generating example images.

Troubleshooting ComfyUI-TrainTools-MZ

Here are some common issues and their solutions:

Issue: Training folder not initializing

Solution: Ensure that the lora_name, branch, and source parameters are correctly set in the MZ_KohyaSSInitWorkspace node. Verify that the output directory has the necessary write permissions.

Issue: Images not copying to the training folder

Solution: Check the images parameter in the MZ_ImagesCopyWorkspace node. Make sure the image paths are correct and that the source folder is accessible. If using the ComfyUI-VideoHelperSuite, ensure it is properly installed and configured.

Issue: Training process fails

Solution: Review the configurations in the MZ_KohyaSSUseConfig and MZ_KohyaSSAdvConfig nodes. Ensure all required fields are correctly filled. Refer to the for detailed configuration options.

Learn More about ComfyUI-TrainTools-MZ

For additional resources, tutorials, and community support, consider the following:

• for video-related workflows and image handling. These resources provide comprehensive guides and community forums where you can ask questions and share your experiences with other AI artists.