ModelMergeMOSSImage - ComfyUI Built-in Node Documentation

The ModelMergeMOSSImage node merges two MOSS image models by blending their matching weights block by block, so one part of the result can come from a base model and the rest from a fine-tune or a second model. Every major component of the architecture has its own ratio: the video embedder, the text embedder, the time embedding, the shared block modulation, all 42 transformer layers, and the output heads.

Inputs

Parameter	Description	Data Type	Required	Range
`model1`	The first model to merge. Everything the second model does not contribute is kept from this model	MODEL	Yes	-
`model2`	The second model to merge. Its weights are blended in according to the ratios below	MODEL	Yes	-
`video_embedder.`	Blend ratio for the video embedder block (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step: 0.01)
`text_embedder.proj.`	Blend ratio for the projection layer of the text embedder (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step: 0.01)
`text_embedder.blocks.0.` and `text_embedder.blocks.1.`	Blend ratio for each of the two text embedder blocks (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step: 0.01)
`time_embedding.`	Blend ratio for the time embedding block (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step: 0.01)
`shared_block_modulation.`	Blend ratio for the shared modulation layer used by the transformer blocks (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step: 0.01)
`layers.0.` through `layers.41.`	Blend ratio for each of the 42 transformer layers (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step: 0.01)
`final_norm_video.`	Blend ratio for the final normalization of the video stream (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step: 0.01)
`final_linear_video.`	Blend ratio for the final linear layer of the video stream (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step: 0.01)
`video_head_modulation.`	Blend ratio for the modulation layer of the video output head (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step: 0.01)

A ratio of 0.0 keeps that component as it is in model1, a ratio of 1.0 takes it entirely from model2, and values in between produce a weighted blend. Both inputs must be MOSS image models with the same architecture; weights that do not match any of the prefixes above fall back to the ratio of video_embedder..

Outputs

Output Name	Description	Data Type
`MODEL`	The merged model, combining both inputs according to the specified ratios	MODEL

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): 3b7d9c806a8d53ad00044bf30f5e63431fea7ee565414acb70a56e3059164288

​Inputs

​Outputs

Inputs

Outputs