LoraLoaderBypass - ComfyUI Built-in Node Documentation

The LoraLoaderBypass node applies a LoRA (Low-Rank Adaptation) to a diffusion model and a CLIP model in a special “bypass” mode. Unlike a standard LoRA loader, this method does not permanently modify the base model’s weights. Instead, it computes the result by adding the LoRA’s contribution to the model’s normal forward pass, which is useful for training or when working with models that have their weights offloaded.

Inputs

Parameter	Description	Data Type	Required	Range
`model`	The diffusion model the LoRA will be applied to.	MODEL	Yes	-
`clip`	The CLIP model the LoRA will be applied to.	CLIP	Yes	-
`lora_name`	The name of the LoRA. Available LoRA files are loaded from the `loras` folder.	COMBO	Yes	List of available LoRA files
`strength_model`	How strongly to modify the diffusion model. This value can be negative (default: 1.0).	FLOAT	Yes	-100.0 to 100.0 (step: 0.01)
`strength_clip`	How strongly to modify the CLIP model. This value can be negative (default: 1.0).	FLOAT	Yes	-100.0 to 100.0 (step: 0.01)

Note: If both strength_model and strength_clip are set to 0, the node returns the original, unmodified model and clip inputs without processing. Note: The selected LoRA file is cached after it is first loaded. It is only re-read from disk when a different lora_name is chosen.

Outputs

Output Name	Description	Data Type
`MODEL`	The modified diffusion model.	MODEL
`CLIP`	The modified CLIP model.	CLIP

Note: This node is marked as experimental.

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 025f0638a6690a53b1a6c4548dac24fb7e7f26e04ff4b1c88d29b061430037a8

​Inputs

​Outputs

Inputs

Outputs