⭐ Star Tiled RelayVSR Upscaler

Upscales an image with RelayVSR by processing it in overlapping tiles to keep VRAM usage low. Tiles are enhanced one by one with the one-step RelayVSR pass and blended back together seamlessly.

By Starnodes2024·Created 2 years ago·Updated 6 days ago· 120

image
model_override

IMAGE

◄model_name▾►

◄vae_name▾►

◄seed0►

◄tiledtrue►

◄max_tile_size1024►

◄tile_overlap0.25►

◄detail0.60►

◄batch_size0►

Category⭐StarNodes/Image And Latent

Inputs (10)

Name	Type	Default	Description
image	IMAGE		The image to upscale. It goes in at its native size - the RelayVSR checkpoint decides the upscale factor (e.g. x4).
model_name	COMBO		The RelayVSR diffusion model (e.g. relayvsr_x4_int8_convrot.safetensors).
vae_name	COMBO		The RelayVSR Flash VAE (e.g. relayvsr_flash_vae_x4_fp32.safetensors).
seed	INT	00–18446744073709550000	Random seed for the one-step sampling pass.
tiled	BOOLEAN	true	On: split the image into overlapping tiles (low VRAM, for large images). Off: process the whole image in a single pass - faster and seam-free, but needs a lot of VRAM on large images. Ignores max_tile_size and tile_overlap.
max_tile_size	INT	102464–16384	Maximum tile width/height in pixels of the input image. The image is split into as many tiles as needed to stay under this size - smaller tiles = less VRAM, larger tiles = fewer seams.
tile_overlap	FLOAT	0.250.05–0.5	Overlap ratio between tiles. Higher values reduce seam artifacts but increase VRAM/time.
detail	FLOAT	0.600–1	Scales the high-frequency residual the model adds on top of the plain bicubic upscale. 0.0 = pure bicubic upscale, 1.0 = full model output.
batch_size	INT	00–10000	Frames processed together per tile. 0 = auto: the whole input runs as ONE video pass - the fastest setting for videos (the model only samples every 15th frame anyway). 1 = each image separately (use for unrelated images). >1 = fixed frames per pass (lower VRAM than auto on very long videos).
model_overrideopt	MODEL		Optional model override. When connected, this model will be used instead of the one selected in model_name dropdown. Useful for patched models or custom loaders.

Outputs (1)

Name	Type	Description
IMAGE	IMAGE	—