Skip to main content

SeedVR2ProgressiveSampler

Sequential temporal chunking sampler for SeedVR2 native workflows. This node processes long video latents by splitting them into smaller temporal chunks, sampling each chunk sequentially, and blending the results together. It serves as a drop-in replacement for the standard KSampler when working with SeedVR2 models on sequences that would otherwise cause out-of-memory errors.

Inputs

ParameterDescriptionData TypeRequiredRange
modelThe model used for denoising the input latentMODELYes
seedThe random seed used for creating the noise (default: 0)INTYes0 to 0xffffffffffffffff
stepsThe number of steps used in the denoising process (default: 20)INTYes1 to 10000
cfgThe Classifier-Free Guidance scale balances creativity and adherence to the prompt. Higher values result in images more closely matching the prompt however too high values will negatively impact quality (default: 1.0)FLOATYes0.0 to 100.0
sampler_nameThe algorithm used when sampling, this can affect the quality, speed, and style of the generated outputCOMBOYesMultiple options available
schedulerThe scheduler controls how noise is gradually removed to form the imageCOMBOYesMultiple options available
positiveThe conditioning describing the attributes you want to include in the imageCONDITIONINGYes
negativeThe conditioning describing the attributes you want to exclude from the imageCONDITIONINGYes
latentThe latent image to denoiseLATENTYes
denoiseThe amount of denoising applied, lower values will maintain the structure of the initial image allowing for image to image sampling (default: 1.0)FLOATYes0.0 to 1.0
frames_per_chunkPixel frames per temporal chunk. Must be a 4n+1 value (1, 5, 9, 13, 17, 21, …) to match SeedVR2 constraints (default: 21)INTYes1 to 16384 (step of 4)
temporal_overlapLatent frames blended between adjacent chunks to hide the seam; 0 means no blend (default: 0)INTYes0 to 16384
chunking_modemanual = use frames_per_chunk exactly; auto = shrink the chunk until it fits in VRAM (default: “manual”)COMBOYes”manual"
"auto”
Note on frames_per_chunk: This parameter must be a 4n+1 pixel-frame count (1, 5, 9, 13, 17, 21, …). The node will raise an error if an invalid value is provided. Note on temporal_overlap: The overlap value is automatically capped to be at most one less than the latent chunk size to ensure valid chunk processing. Note on chunking_mode: When set to “auto”, the node will automatically try smaller chunk sizes if the current chunk causes an out-of-memory error. If all attempts fail, the node raises an error.

Outputs

Output NameDescriptionData Type
latentThe denoised latent output, concatenated from all temporal chunks back into a single collapsed SeedVR2 latent tensorLATENT
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): a4574c3e619954b5569551b5b2ba112ecbff918dcebb5ba718a14e77701144a9