LoraExtractKJ

The LoraExtractKJ node supports multiple LoRA extraction modes, including standard fixed-rank (standard), full-rank (full), and various adaptive-rank methods (adaptive_*), to accommodate different precision and file size requirements.

The LoraExtractKJ node is a tool for extracting LoRA (Low-Rank Adaptation) weight files from the differences between a fine-tuned model and the original base model. Its core function is to analyze the parameter changes between the two models, compress these changes, and save them as independent, reusable LoRA files. This node does not output any data but directly saves the generated LoRA file to a specified directory.

Node Functionality

This node supports multiple LoRA extraction modes, including standard fixed-rank (standard), full-rank (full), and various adaptive-rank methods (adaptive_*), to accommodate different precision and file size requirements. Technically, it uses the Singular Value Decomposition (SVD) algorithm to compute a low-rank approximation of the weight matrices. Users can choose between the svd_linalg (precise but slower) or svd_lowrank (faster but approximate) algorithms. The extracted weights can be saved in fp16, bf16, or fp32 precision.

Node Parameter Description - LoraExtractKJ

Connection Inputs (Inputs)

Parameter Name	Data Type	Required	Default Value	Range/Options	Description
`finetuned`	MULTITYPE	Yes	-	-	The finetuned model or clip to extract LoRA from.
`original`	MULTITYPE	Yes	-	-	The original base model or clip to diff against.

Control Parameters (Parameters)

Parameter Name	Data Type	Required	Default Value	Range/Options	Description
`lora_type`	COMBO	Yes	standard	"standard", "full", "adaptive_ratio", "adaptive_quantile", "adaptive_energy", "adaptive_fro"	Full-rank (full) and four adaptive-rank methods (adaptive_ratio, adaptive_quantile, adaptive_energy, adaptive_fro).
`algorithm`	COMBO	Yes	svd_lowrank	"svd_linalg", "svd_lowrank"	Select the algorithm used for Singular Value Decomposition (SVD). `svd_linalg` is more precise but slower, `svd_lowrank` is faster but approximate.
`output_dtype`	COMBO	Yes	fp16	"fp16", "bf16", "fp32"	The precision for saving the extracted LoRA weights: fp16, bf16, or fp32.
`filename_prefix`	STRING	Yes	loras/ComfyUI_extracted_lora	-	Sets the save path and filename prefix for the output LoRA file. By default, it is saved in the `loras/` directory.
`rank`	INT	Yes	64	1 - 4096 (step: 1)	The rank to use for standard LoRA, or maximum rank limit for adaptive methods.
`lowrank_iters`	INT	Yes	7	1 - 100 (step: 1)	The number of subspace iterations for the lowrank SVD algorithm.
`bias_diff`	BOOLEAN	Yes	True	-	Whether to include the differences in the model's bias parameters when extracting LoRA.
`adaptive_param`	FLOAT	Yes	0.15	0.0 - 1.0 (step: 0.01)	For ratio mode, this is the ratio of the maximum singular value. For quantile mode, this is the quantile of the singular values. For fro mode, this is the Frobenius norm retention ratio.
`clamp_quantile`	BOOLEAN	Yes	false	-	Controls whether to clamp singular values in adaptive quantile mode to improve numerical stability.

Usage Scenario

A typical usage scenario is when a user has performed personalized fine-tuning on a Stable Diffusion model using methods like Dreambooth. After fine-tuning, the user can connect the complete fine-tuned model and the original, unmodified base model to the node's two input ports. By configuring the extraction parameters, a lightweight LoRA file can be generated for quickly applying the same style or concept in other workflows.

LoraExtractKJ Node Source Code Link

The LoraExtractKJ node is from the ComfyUI-KJNodes node package.

Lora Extract KJ