ComfyUI Node: Realtime LoRA Trainer (Qwen Image Edit - Musubi Tuner)

Realtime LoRA Trainer (Qwen Image Edit - Musubi Tuner):

The MusubiQwenImageEditLoraTrainer is a specialized node designed for training Qwen Image Edit LoRAs using the Musubi Tuner. This node is particularly useful for AI artists who wish to develop image editing behaviors by leveraging source and target image pairs. It operates by utilizing folder paths for images, rather than direct image inputs, which simplifies the process of organizing and managing training data. The node supports both Qwen-Image-Edit and Qwen-Image-Edit-2509 models, providing flexibility in the types of image editing tasks it can handle. By using this node, you can create customized image editing models that can be integrated into various creative workflows, enhancing the ability to automate and refine image editing tasks.

Realtime LoRA Trainer (Qwen Image Edit - Musubi Tuner) Input Parameters:

images_path

This parameter specifies the folder path where the source images are stored. These images serve as the input data for training the LoRA model. The quality and relevance of these images directly impact the effectiveness of the training process.

control_path

This parameter indicates the folder path for the control images, which are used as target outputs during training. These images guide the model in learning the desired editing transformations.

musubi_path

This parameter defines the path to the Musubi Tuner, which is the tool used for training the LoRA model. It is essential for the execution of the training process.

model_mode

This parameter determines the mode of the model being trained, affecting how the model processes the input data and learns from it. Different modes may be available depending on the specific requirements of the task.

dit_model

This parameter specifies the DIT (Diffusion Image Transformer) model to be used during training. The choice of model can influence the quality and style of the image edits produced by the trained LoRA.

vae_model

This parameter sets the VAE (Variational Autoencoder) model, which is crucial for encoding and decoding images during the training process. The VAE model helps in capturing the essential features of the images.

text_encoder

This parameter involves the text encoder used in conjunction with the image data, allowing for the integration of textual information into the training process. This can enhance the model's ability to understand and apply complex editing instructions.

training_steps

This parameter defines the number of training steps to be executed. More steps generally lead to better model performance but require more computational resources and time.

learning_rate

This parameter sets the learning rate for the training process, which controls how quickly the model updates its parameters. A suitable learning rate is crucial for effective training and avoiding issues like overfitting or underfitting.

lora_rank

This parameter specifies the rank of the LoRA, which affects the model's capacity and complexity. A higher rank can capture more intricate patterns but may require more resources.

vram_mode

This parameter determines the VRAM (Video RAM) usage mode, which can be adjusted based on the available hardware resources to optimize performance and prevent memory-related issues.

blocks_to_swap

This parameter indicates which blocks of the model should be swapped during training, allowing for customization of the model architecture to better suit specific tasks.

keep_lora

This boolean parameter decides whether to retain the trained LoRA after the training process is complete. Keeping the LoRA can be useful for further refinement or reuse.

output_name

This parameter sets the name for the output LoRA file, allowing you to easily identify and manage the trained models.

custom_python_exe

This parameter allows you to specify a custom Python executable, which can be useful if you need to run the training process in a specific Python environment.

Realtime LoRA Trainer (Qwen Image Edit - Musubi Tuner) Output Parameters:

lora_path

This output parameter provides the path to the trained Qwen Image Edit LoRA file. This path is essential for accessing and utilizing the trained model in subsequent image editing tasks. The output file can be integrated into various workflows, enabling automated and refined image editing capabilities.

Realtime LoRA Trainer (Qwen Image Edit - Musubi Tuner) Usage Tips:

• Ensure that the images in the images_path and control_path are well-organized and relevant to the editing task to improve training outcomes.
• Adjust the training_steps and learning_rate parameters to balance between training time and model performance, starting with default values and fine-tuning as needed.
• Use the vram_mode setting to optimize performance based on your hardware capabilities, especially if you encounter memory limitations.

Realtime LoRA Trainer (Qwen Image Edit - Musubi Tuner) Common Errors and Solutions:

"FileNotFoundError: [Errno 2] No such file or directory"

• Explanation: This error occurs when the specified path for images, control images, or Musubi Tuner is incorrect or the directory does not exist.
• Solution: Double-check the paths provided for images_path, control_path, and musubi_path to ensure they are correct and the directories exist.

"ValueError: Invalid model mode"

• Explanation: This error indicates that the model_mode parameter has been set to an unsupported value.
• Solution: Verify the available model modes and ensure that the model_mode parameter is set to a valid option.

"MemoryError: Unable to allocate VRAM"

• Explanation: This error suggests that the VRAM required for the training process exceeds the available memory.
• Solution: Adjust the vram_mode to a lower setting or reduce the lora_rank to decrease memory usage.