CosmosPredict2ImageToVideoLatent - ComfyUI Built-in Node Documentation

Creates video latents for the Cosmos Predict2 image-to-video workflow. The node can produce an empty video latent of a given size and length, or insert encoded start and/or end images into the sequence so those frames are kept during generation. Any supplied image is resized to the requested width and height and encoded with the provided VAE before being placed at the beginning and/or end of the latent sequence.

Inputs

Parameter	Description	Data Type	Required	Range
`vae`	The VAE model used to encode the start and end images into latent space	VAE	Yes	-
`width`	Width of the output video in pixels (default: 848, must be a multiple of 16)	INT	Yes	16 to MAX_RESOLUTION
`height`	Height of the output video in pixels (default: 480, must be a multiple of 16)	INT	Yes	16 to MAX_RESOLUTION
`length`	Number of frames in the video sequence (default: 93)	INT	Yes	1 to MAX_RESOLUTION
`batch_size`	Number of video sequences to generate (default: 1)	INT	Yes	1 to 4096
`start_image`	Optional starting image for the video sequence	IMAGE	No	-
`end_image`	Optional ending image for the video sequence	IMAGE	No	-

Note: When neither start_image nor end_image is provided, the node simply returns an empty latent of the requested size and length. When one or both images are provided, they are resized to width and height, encoded with the vae, and placed at the beginning and/or end of the latent sequence. The corresponding regions are marked in the noise mask so they are preserved during generation. The encoded latents are converted with the Wan 2.1 latent format, and the resulting latent and mask are repeated batch_size times.

Outputs

Output Name	Description	Data Type
`LATENT`	The generated video latent containing `samples` (the video latent sequence) and, when at least one of `start_image` or `end_image` is provided, a `noise_mask` marking the frames that should be preserved during generation	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): 842bd2b8cda438e7b938439d4eba280478939e3302dc1846d52595d40082ff05

​Inputs

​Outputs

Inputs

Outputs