Native LoRA Training

Train LoRA models directly in ComfyUI using built-in training nodes

ComfyUI includes native support for training LoRA (Low-Rank Adaptation) models without requiring external tools or custom nodes. This page provides an overview of the training workflow. For detailed parameter documentation, refer to the individual node pages linked below.

The training nodes are marked as experimental. Features and behavior may change in future releases.

Node Overview

The native LoRA training system is organized into dataset nodes and training nodes.

Dataset Nodes

Used to prepare and manage training data:

Node	Purpose
Load Image Dataset from Folder	Batch-load images from an input subfolder
Load Image and Text Dataset from Folder	Load images with paired captions (supports kohya-ss folder structure)
Make Training Dataset	Encode images with VAE and text with CLIP to produce training data
Resolution Bucket	Group latents by resolution for efficient batched training
Save Training Dataset	Save encoded dataset to disk to avoid re-encoding on future runs
Load Training Dataset	Load a previously saved encoded dataset from disk

Training Nodes

Used to run training, save results, and apply the LoRA:

Node	Purpose
Train LoRA	Train a LoRA from latents and conditioning data
Save LoRA Weights	Export trained LoRA weights to a safetensors file
Load LoRA Model	Apply trained LoRA weights to a model for inference
Plot Loss Graph	Visualize training loss over time

Requirements

• A GPU with sufficient VRAM (training typically requires more memory than inference)
• Training images placed in a subfolder under ComfyUI/input/
• A base model (checkpoint)

Typical Training Workflow

1. Load training images

   Place your training images in a subfolder under ComfyUI/input/.

   • Use Load Image Dataset from Folder for images only
   • Use Load Image and Text Dataset from Folder for image–caption pairs (each image needs a matching .txt file with the same base name)

   Start with 10–20 high-quality images. Quality matters more than quantity.

2. Encode the dataset

   Connect images and text to Make Training Dataset along with a VAE and CLIP model. This produces latents and conditioning outputs.

   To reuse the same dataset across multiple training runs, save it with Save Training Dataset and load it later with Load Training Dataset — no re-encoding needed.

3. (Optional) Resolution bucketing

   If your images have varying dimensions, pass the encoded data through Resolution Bucket to group them by resolution, then enable bucket_mode in the Train LoRA node for efficient batched training.

4. Configure and run Train LoRA

   Connect the model, latents, and conditioning to Train LoRA and adjust parameters as needed.

   Recommended starting values:

   Parameter	Starting value
   steps	100–500
   rank	8–16
   learning_rate	0.0001–0.0005
   optimizer	AdamW
   loss_function	MSE

   The node outputs trained lora weights, a loss_map, and the completed steps count.

5. Monitor training progress

   Connect loss_map to Plot Loss Graph to visualize the loss curve. Training can be stopped once the loss plateaus.

6. Save and test your LoRA

   Connect lora to Save LoRA Weights to export a .safetensors file to ComfyUI/output/loras/.

   In your inference workflow, use Load LoRA Model to apply the trained LoRA to the base model and test the results.

VRAM Optimization

Technique	Notes
gradient_checkpointing (on by default)	Reduces VRAM by recomputing activations during backward pass
Lower batch_size	Most direct way to reduce VRAM
Higher grad_accumulation_steps	Equivalent to a larger batch size with no extra VRAM cost
offloading	Moves model weights to CPU; requires gradient_checkpointing to be enabled
bypass_mode	Applies adapters via forward hooks instead of weight modification; required for quantized models (FP8/FP4)

Quantized Model Training

To train a LoRA on a quantized model (FP8/FP4), use these settings in Train LoRA:

• training_dtype: none
• quantized_backward: enabled
• bypass_mode: enabled

Also enable bypass in Load LoRA Model when using the resulting LoRA for inference.

Continuing Training

Set existing_lora in Train LoRA to an existing saved LoRA file to resume from a checkpoint. The total step count accumulates automatically.

Supported Algorithms

The algorithm parameter in Train LoRA selects the weight adapter type:

Algorithm	Notes
LoRA	Standard low-rank adaptation — recommended for most use cases
LoHa	Hadamard-product low-rank adaptation
LoKr	Kronecker-product low-rank adaptation — more parameter-efficient
OFT	Orthogonal Fine-Tuning (experimental)