⬡ SeedVR2 Upscale¶

AI-native upscaling using the SeedVR2 tiling and stitching engine. Tile/blending parameters are available via Show advanced inputs.

Inputs¶

Name	Type	Required	Advanced	Default	Description
`gen_pipe`	`UME_PIPELINE`	✅		—	Pipeline from KSampler
`enabled`	`BOOLEAN`	✅		ON	Toggle upscaling on/off
`model`	`COMBO`	✅		seedvr2_ema_3b_fp8	SeedVR2 model selection
`upscale_by`	`FLOAT`	✅		2.0	Scale factor (1–8x)
`tile_width`	`INT`	✅	✅	512	Tile width in pixels
`tile_height`	`INT`	✅	✅	512	Tile height in pixels
`mask_blur`	`INT`	✅	✅	0	Tile edge softening
`tile_padding`	`INT`	✅	✅	32	Tile overlap in pixels
`tile_upscale_resolution`	`INT`	✅	✅	1024	Per-tile processing resolution
`tiling_strategy`	`COMBO`	✅	✅	Chess	Tile layout strategy
`anti_aliasing_strength`	`FLOAT`	✅	✅	0.0	Edge smoothing between tiles
`blending_method`	`COMBO`	✅	✅	auto	Tile merging algorithm
`color_correction`	`COMBO`	✅	✅	lab	Color matching between tiles

Outputs¶

Name	Type	Description
`gen_pipe`	`UME_PIPELINE`	Pipeline with upscaled image

VRAM Requirements

SeedVR2 requires approximately 6 GB of free VRAM. The node includes automatic VRAM management — it will unload cached models if necessary to free space.