ComfyUI Node: FL VoxCPM LoRA Trainer

FL VoxCPM LoRA Trainer:

The FL VoxCPM LoRA Trainer is a specialized node designed to facilitate the training of a LoRA (Low-Rank Adaptation) adapter for the VoxCPM models, specifically versions V1 and V2. This node automatically detects the version of the VoxCPM model you are working with and routes the training process to the appropriate trainer, ensuring a seamless and efficient workflow. The primary goal of this node is to enhance the capabilities of the VoxCPM models by fine-tuning them with a LoRA adapter, which can significantly improve the model's performance in specific tasks without the need for extensive computational resources. By leveraging this node, you can achieve a more tailored and optimized model that meets your specific needs, making it an invaluable tool for AI artists looking to push the boundaries of their creative projects.

FL VoxCPM LoRA Trainer Input Parameters:

base_model_name

This parameter specifies the base VoxCPM model you wish to fine-tune, either V1 or V2. Selecting the correct model version is crucial as it determines the training path and the specific configurations that will be applied. The default option is the first model in the list of available models.

train_config

This input requires a configuration object from either the VoxCPM Train Config (V1) or V2 Train Config node. It contains all the necessary parameters and settings for the training process, ensuring that the model is fine-tuned according to your specific requirements.

dataset_path

The path to the train.jsonl file generated by the VoxCPM Dataset Maker. This file contains the training data, which is essential for the training process. Providing the correct path ensures that the model is trained on the intended dataset.

output_name

This parameter defines the name of the subfolder within models/loras/VoxCPM/ where the training results will be saved. The default value is "my_lora_v1". Naming the output folder appropriately helps in organizing and retrieving the trained models efficiently.

max_steps

Specifies the total number of training steps to be executed. The minimum value is 100, and the maximum is 100,000, with a default of 1,000 steps. This parameter directly impacts the duration and thoroughness of the training process.

save_every_steps

Determines how frequently checkpoints are saved during the training process. The minimum value is 50 steps, and the maximum is 5,000 steps, with a default of 200 steps. Regular checkpoints allow you to monitor progress and recover from interruptions.

num_workers

Indicates the number of dataloader workers to be used, with a range from 0 to 8 and a default of 0, which means the main thread will handle data loading. Adjusting this parameter can optimize data loading efficiency based on your system's capabilities.

validation_text

An optional parameter where you can input text to synthesize at each checkpoint for audio validation. If left empty, validation is skipped. This feature allows you to assess the model's performance at various stages of training.

validation_steps

Specifies the number of inference timesteps for generating validation audio. The range is from 1 to 100, with a default of 10 steps. This parameter helps in evaluating the model's progress and quality of output during training.

FL VoxCPM LoRA Trainer Output Parameters:

LoRA Output Path

This output provides the path to the directory where the trained LoRA adapter is saved. It is crucial for locating and utilizing the trained model in subsequent tasks or deployments. The output path ensures that you can easily access and manage your trained models.

FL VoxCPM LoRA Trainer Usage Tips:

• Ensure that the base_model_name matches the version of the VoxCPM model you intend to fine-tune to avoid compatibility issues.
• Regularly save checkpoints using the save_every_steps parameter to prevent data loss and facilitate progress tracking.
• Utilize the validation_text and validation_steps parameters to monitor the model's performance and make necessary adjustments during training.

FL VoxCPM LoRA Trainer Common Errors and Solutions:

"Model version not detected"

• Explanation: The node could not determine whether the model is V1 or V2. - Solution: Verify that the base_model_name is correctly set to a valid VoxCPM model version.

"Invalid dataset path"

• Explanation: The specified dataset_path does not point to a valid train.jsonl file.
• Solution: Check the path for typos and ensure the file exists at the specified location.

"Training configuration missing"

• Explanation: The train_config input is not provided or is incorrect.
• Solution: Ensure that a valid configuration object from the appropriate Train Config node is connected to the train_config input.