LTXVSpatioTemporalGuidance - ComfyUI Built-in Node Documentation - ComfyUI

This node improves the spatial detail and motion coherence of LTXV video generation by running one extra pass at each sampling step. During this pass, the self-attention of the selected transformer blocks is degraded to a value-passthrough, and the generation is guided away from that degraded result.

Inputs

Parameter	Description	Data Type	Required	Range
`model`	The base model to apply spatio-temporal guidance to. The model is cloned and a post-CFG guidance function is attached to the clone.	MODEL	Yes	—
`scale`	The strength of the guidance applied to the denoised result. When set to 0, the guidance has no effect. (default: 1.0)	FLOAT	Yes	0.0 to 100.0 (step 0.01)
`blocks`	Comma-separated transformer block indices to perturb. Only numeric values are used; any other characters are ignored. (default: “29”)	STRING	Yes	—
`start_percent`	The fraction of the sampling process at which guidance begins. This is an advanced parameter. (default: 0.0)	FLOAT	Yes	0.0 to 1.0 (step 0.001)
`end_percent`	The fraction of the sampling process at which guidance ends. This is an advanced parameter. (default: 1.0)	FLOAT	Yes	0.0 to 1.0 (step 0.001)

Note: Guidance is only applied during the sampling interval between start_percent and end_percent. Outside that interval, the original denoised result is returned unchanged. If scale is 0 or blocks contains no numeric values, the guided pass has no effect on the sampling process.

Outputs

Output Name	Description	Data Type
`MODEL`	The cloned model with the spatio-temporal guidance function attached.	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): 0e14137b3bf416d36005b6b4b6db46495b1523f88b2bf574e2dc582175422a48