SeedVR2 Video Path Upscaler (v2.5.24)

Upscales a native ComfyUI VIDEO in bounded chunks. The complete frame sequence never becomes an IMAGE tensor in the workflow; the output remains file-backed. Source audio is preserved in the returned VIDEO and is also returned once as AUDIO.

By ethanfel·Created 5 days ago·Updated 3 days ago· 0

video
dit
vae
auto_settings

video
audio

◄seed42►

◄resolution1080►

◄max_resolution0►

◄batch_size5►

◄chunk_size21►

◄color_correctionlab►

◄chunk_overlap2►

◄uniform_batch_sizefalse►

◄temporal_overlap0►

◄prepend_frames0►

◄input_noise_scale0.000►

◄latent_noise_scale0.000►

◄offload_devicecpu►

◄temporary_video_crf18►

◄enable_debugfalse►

CategorySeedVR2 Video Path

Inputs (19)

Name	Type	Default	Description
video	VIDEO		Native file-backed VIDEO, normally from Load Video or MiniMax H3 Full-Chain Latent Video Adapter. Frames are decoded only as needed, so the full clip does not become a ComfyUI IMAGE tensor.
dit	SEEDVR2_DIT		DiT model configuration from a SeedVR2 model loader or Auto Configurator.
vae	SEEDVR2_VAE		VAE model configuration from a SeedVR2 model loader or Auto Configurator.
seed	INT	420–4294967295	Random seed for reproducible generation (default: 42). The same seed and inputs produce the same output.
resolution	INT	108016–16384	Target resolution for the shortest edge in pixels (default: 1080). The input aspect ratio is preserved automatically.
max_resolution	INT	00–16384	Maximum resolution of either output dimension (default: 0, no limit). Useful for limiting VRAM use with extreme aspect ratios.
batch_size	INT	51–16384	Frames processed together by SeedVR2 (default: 5). Use the 4n+1 pattern: 1, 5, 9, 13, 17, 21, ... Higher values generally improve temporal consistency and speed but use more VRAM.
chunk_size	INT	210–4096	Maximum source frames decoded and held in RAM at once. Set to 0 to disable file chunking and process the complete active video at once.
color_correction	COMBO	lab	Correct output color shifts to resemble the source (default: lab). LAB is recommended; choose none to preserve the model output unchanged.
chunk_overlapopt	INT	20–32	Raw context frames carried from the previous file chunk (default: 2). These context frames are removed from the output, so duration is unchanged.
uniform_batch_sizeopt	BOOLEAN	false	Pad the final model batch to batch_size (default: False). This adds compute but can prevent artifacts from a very small final batch.
temporal_overlapopt	INT	00–16	Overlapping frames between consecutive model batches (default: 0). Values from 1 to 4 can improve consistency across batch boundaries.
prepend_framesopt	INT	00–32	Frames mirrored before the beginning of the video (default: 0). This can reduce start-of-video artifacts; added frames are removed from the result.
input_noise_scaleopt	FLOAT	0.0000–1	Noise added to source frames before encoding (default: 0, disabled). Small values can help with some input artifacts.
latent_noise_scaleopt	FLOAT	0.0000–1	Noise added in latent space during diffusion (default: 0, disabled). This can soften details when input noise is not sufficient.
offload_deviceopt	COMBO	cpu	Device for intermediate tensors between processing phases (default: cpu). Use none to retain tensors on the inference GPU for maximum speed and VRAM use.
temporary_video_crfopt	INT	180–51	H.264 quality of the file-backed intermediate; lower is higher quality.
enable_debugopt	BOOLEAN	false	Enable detailed memory, timing, and processing logs. Useful when diagnosing errors or tuning performance.
auto_settingsopt	SEEDVR2_AUTO_SETTINGS		Optional runtime settings from SeedVR2 Video Path Auto Configurator. When connected, the configurator overrides the corresponding controls above.

Outputs (2)

Name	Type	Description
video	VIDEO	File-backed upscaled video with source audio preserved when present.
audio	AUDIO	Full unchunked audio from the active source video window.