SeedVR2TemporalChunk - ComfyUI Built-in Node Documentation

This node splits a SeedVR2 video latent into smaller overlapping temporal chunks that can be processed one at a time within available VRAM. The chunks are designed to be passed to both the Apply SeedVR2 Conditioning node and the sampler latent input, then later recombined using the Merge SeedVR2 Latents node.

Inputs

Parameter	Description	Data Type	Required	Range
`latent`	The VAE-encoded SeedVR2 latent to split	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 = use frames_per_chunk exactly; auto = predict the largest chunk that fits free VRAM	COMBO	Yes	”auto" "manual”

When chunking_mode is set to “manual”, the following parameter becomes available:

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

Notes on parameter constraints:

The input latent must be a 5-dimensional video latent with shape (B, C, T, H, W) and exactly 4 latent channels
When using “manual” mode, frames_per_chunk must be a 4n+1 value (1, 5, 9, 13, 17, 21, …)
temporal_overlap is automatically capped to be less than the chunk size
In “auto” mode, the node calculates the optimal chunk size based on available free VRAM

Outputs

Output Name	Description	Data Type
`latents`	The temporal chunks in sequence order	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