LoraLoader (dirty)

Documentation

• Class name: LoraLoader (dirty)
• Category: Bmad/api/dirty loaders
• Output node: False

The LoraLoader (dirty) node is designed to dynamically load LoRA (Low-Rank Adaptation) parameters into models and CLIP instances, allowing for the customization and fine-tuning of these components based on specified LoRA files and strength parameters. It serves as a bridge between static model states and adaptable, performance-enhanced versions tailored to specific tasks or datasets.

Input types

Required

• model
  • The model parameter represents the pre-trained model into which LoRA parameters will be loaded. It is crucial for defining the base architecture that will be enhanced with LoRA adaptations.
  • Comfy dtype: MODEL
  • Python dtype: torch.nn.Module
• clip
  • The clip parameter signifies the CLIP model instance that will be adjusted using LoRA parameters. It plays a key role in adapting the CLIP model for enhanced performance in specific tasks.
  • Comfy dtype: CLIP
  • Python dtype: torch.nn.Module
• lora_name
  • The lora_name parameter specifies the filename of the LoRA parameters to be loaded. It determines which specific LoRA adaptations will be applied to the model and CLIP instances.
  • Comfy dtype: STRING
  • Python dtype: str
• strength_model
  • This parameter controls the intensity of the LoRA adaptation applied to the model. It allows for fine-tuning the impact of LoRA parameters on the model's performance.
  • Comfy dtype: FLOAT
  • Python dtype: float
• strength_clip
  • Similar to strength_model, this parameter adjusts the intensity of the LoRA adaptation on the CLIP model, enabling precise control over its influence.
  • Comfy dtype: FLOAT
  • Python dtype: float

Output types

• model
  • Comfy dtype: MODEL
  • The modified model with LoRA parameters applied, reflecting the adaptations made for enhanced performance or customization.
  • Python dtype: torch.nn.Module
• clip
  • Comfy dtype: CLIP
  • The adjusted CLIP model with LoRA parameters incorporated, showcasing the enhancements made for improved task-specific performance.
  • Python dtype: torch.nn.Module

Usage tips

• Infra type: GPU
• Common nodes: unknown

Source code

class DirtyLoraLoader:
    @classmethod
    def INPUT_TYPES(cls):
        return {"required": {"model": ("MODEL",),
                             "clip": ("CLIP",),
                             "lora_name": ("STRING", {"default": ""}),
                             "strength_model": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
                             "strength_clip": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
                             }}

    RETURN_TYPES = ("MODEL", "CLIP")
    FUNCTION = "load_lora"

    CATEGORY = f"{api_category_path}/dirty loaders"

    def load_lora(self, model, clip, lora_name, strength_model, strength_clip):
        lora_name = DirtyLoaderUtils.find_matching_filename(
            lora_name, folder_paths.get_filename_list("loras"))

        loader = nodes.LoraLoader()
        return loader.load_lora(model, clip, lora_name, strength_model, strength_clip)