SeedVR2TemporalChunk - ComfyUI Built-in Node Documentation

This node splits a SeedVR2 video latent into smaller temporal chunks that can be processed one at a time within available VRAM. It automatically calculates the optimal chunk size based on your GPU memory or lets you specify the chunk size manually, and outputs the chunks in sequence order for processing.

Inputs

Parameter	Description	Data Type	Required	Range
`latent`	The VAE-encoded SeedVR2 latent to split. Must be a 5‑D tensor with the correct channel count.	LATENT	Yes	-
`temporal_overlap`	Latent frames shared between adjacent chunks and crossfaded at merge; 0 means no overlap (default: 0).	INT	No	0 to 16384
`chunking_mode`	Manual uses `frames_per_chunk` exactly; auto predicts the largest chunk that fits free VRAM.	COMBO	Yes	`"auto"` `"manual"`

When chunking_mode is set to "manual", an additional parameter becomes available:

Parameter	Description	Data Type	Required	Range
`frames_per_chunk`	Pixel frames per temporal chunk. Must be a 4n+1 value (1, 5, 9, 13, 17, 21, …) (default: 21).	INT	Yes	1 to 16384

Note: The frames_per_chunk parameter only appears when chunking_mode is set to "manual". The value must satisfy the formula (frames_per_chunk - 1) % 4 == 0, meaning it must be one of: 1, 5, 9, 13, 17, 21, etc. If the total pixel frames in the latent are equal to or less than the chosen chunk size, the node returns the original latent as a single chunk with zero overlap.

Outputs

Output Name	Description	Data Type
`latents`	The temporal chunks in sequence order. Each chunk is a LATENT with the same batch, channel, height, and width but a reduced temporal dimension.	LATENT
`temporal_overlap`	The effective latent-frame overlap between adjacent chunks, for Merge SeedVR2 Latents.	INT

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): 40af2b690f74555efbe38f4cf76825417f52735ce90296a9ad662f78cfe4c6bf

​Inputs

​Outputs

Inputs

Outputs