ComfyUI Node: MinusZone - HYDiTTrain

MinusZone - HYDiTTrain:

MZ_HYDiTTrain is a specialized node designed to facilitate the training of AI models using the HunyuanDiT framework. This node is particularly useful for AI artists who want to fine-tune or train models for text-to-image generation tasks. It integrates seamlessly with the ComfyUI environment, ensuring that the necessary tools and dependencies are automatically managed and configured. The primary goal of MZ_HYDiTTrain is to provide a streamlined and efficient training process, allowing you to focus on the creative aspects of model development without worrying about the underlying technical complexities. By leveraging advanced features such as distributed training, automatic model downloading, and comprehensive configuration options, MZ_HYDiTTrain ensures that you can achieve high-quality results with minimal effort.

MinusZone - HYDiTTrain Input Parameters:

model

This parameter specifies the model architecture to be used for training. It determines the structure and capabilities of the neural network. The default value is "DiT-g/2".

task_flag

This parameter indicates the specific task or objective for the training process. It helps in customizing the training pipeline according to the desired outcome.

resume_split

This boolean parameter determines whether to resume training from a previous split. The default value is True.

ema_to_module

This boolean parameter specifies whether to apply Exponential Moving Average (EMA) to the module. The default value is True.

deepspeed

This boolean parameter indicates whether to use DeepSpeed for optimizing the training process. The default value is False.

predict_type

This parameter defines the type of prediction to be used during training. The default value is "v_prediction".

training_parts

This parameter specifies the parts of the model to be trained. The default value is "lora".

batch_size

This parameter sets the number of samples per batch during training. The default value is 1.

grad_accu_steps

This parameter defines the number of gradient accumulation steps. The default value is 1.

global_seed

This parameter sets the global seed for random number generation, ensuring reproducibility. The default value is 0.

use_flash_attn

This boolean parameter indicates whether to use Flash Attention for faster training. The default value is False.

use_fp16

This boolean parameter specifies whether to use 16-bit floating-point precision for training. The default value is True.

qk_norm

This boolean parameter determines whether to apply QK normalization. The default value is True.

ema_dtype

This parameter sets the data type for EMA. The default value is "fp32".

async_ema

This boolean parameter indicates whether to use asynchronous EMA. The default value is False.

ckpt_latest_every

This parameter specifies the frequency of saving the latest checkpoint. The default value is 0x7fffffff.

multireso

This boolean parameter indicates whether to use multiple resolutions during training. The default value is True.

epochs

This parameter sets the number of training epochs. The default value is 50.

target_ratios

This parameter defines the target aspect ratios for the generated images. The default values are ['1:1', '3:4', '4:3', '16:9', '9:16'].

rope_img

This parameter specifies the base resolution for the RoPE (Rotary Position Embedding) image. The default value is "base1024".

image_size

This parameter sets the size of the images to be generated. The default value is 1024.

rope_real

This boolean parameter indicates whether to use real RoPE. The default value is True.

index_file

This parameter specifies the path to the index file used during training. The default value is None.

lr

This parameter sets the learning rate for the training process. The default value is 1e-5.

MinusZone - HYDiTTrain Output Parameters:

is_running

This boolean parameter indicates whether the training process is currently running. It helps in monitoring the status of the training.

samples

This parameter provides the generated samples during the training process. It allows you to visualize the progress and quality of the model.

MinusZone - HYDiTTrain Usage Tips:

• Ensure that your training configuration file is correctly set up to match your desired training parameters and objectives.
• Regularly monitor the training process using the is_running output parameter to ensure that everything is proceeding as expected.
• Utilize the samples output parameter to periodically check the quality of the generated images and make adjustments to the training parameters if necessary.

MinusZone - HYDiTTrain Common Errors and Solutions:

训练失败!!! 具体报错信息请查看控制台...

• Explanation: This error indicates that the training process has failed due to an unspecified issue.
• Solution: Check the console for detailed error messages and stack traces to identify the root cause of the failure. Ensure that all dependencies are correctly installed and that the training configuration file is properly set up.

LOG: {log} e: {e}

• Explanation: This error message logs an exception that occurred during the training process.
• Solution: Review the logged exception and stack trace to diagnose the issue. Common causes include incorrect parameter values or missing files. Make necessary corrections and restart the training process.