LTXVAddLatentGuide - ComfyUI Built-in Node Documentation

The LTXV Add Latent Guide node pins an already-encoded latent as a guide, for when the guide comes out of an earlier stage rather than an image. It has the same effect as LTXV Add Guide without the VAE decode/encode round trip. A guide that is spatially smaller than the target (an IC-LoRA or detailing reference) is dilated onto a sparse grid, and its RoPE end positions are expanded by the same ratio so it covers the target canvas instead of addressing only the top-left corner of it.

Inputs

Parameter	Description	Data Type	Required	Range
`positive`	Positive conditioning input.	CONDITIONING	Yes	N/A
`negative`	Negative conditioning input.	CONDITIONING	Yes	N/A
`vae`	The VAE model used to read the downscale index formula for frame placement.	VAE	Yes	N/A
`latent`	Target video latent the guide is pinned onto.	LATENT	Yes	N/A
`guiding_latent`	Guide latent. Its spatial size must divide the target’s by the same whole number on both axes; equal size pins it as-is, half size is treated as an x2 IC-LoRA reference.	LATENT	Yes	N/A
`latent_idx`	Latent frame index to start the guide at, counted in latent frames rather than pixel frames. Negative values place the guide on frames before the start of the latent, not counted back from its end. Default: 0.	INT	Yes	-9999 to 9999
`strength`	Capped at 1.0. A dilated guide marks its padding positions with a negative denoise mask so the model drops them; above 1.0 the kept positions would go negative too and the whole guide would be dropped. Amplify beyond 1.0 with attention_mask instead. Default: 1.0.	FLOAT	Yes	0.0 to 1.0, step 0.01
`attention_mask`	Optional pixel-space spatial mask. Controls per-region conditioning influence via self-attention, multiplied by strength.	MASK	No	N/A

Notes

Both latent and guiding_latent must be 5D video latents shaped (batch, channels, frames, height, width).
The guide must fit inside the target latent: the guide’s frame count added to latent_idx must not run past the end of the target latent. Negative latent_idx values are allowed and place the guide before the start of the latent.
The guide’s spatial size must divide the target’s spatial size by a whole number on both the height and width axes.
The height ratio and the width ratio must be the same value (a square ratio). A non-square ratio raises an error, because the dilation and RoPE placement use a single downscale factor for both axes.

Outputs

Output Name	Description	Data Type
`positive`	Positive conditioning with the guide attached.	CONDITIONING
`negative`	Negative conditioning with the guide attached.	CONDITIONING
`latent`	Latent output with the guide applied, including the updated `noise_mask`.	LATENT

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): 19542500484dbc57fdbeeab8ba05bc2978246b3be5decc413825f616cab46f73

​Inputs

​Notes

​Outputs

Inputs

Notes

Outputs