Context-Anchored Tile Refine (VL)

A ComfyUI node in image/upscaling with 14 inputs and 1 output.

By Blakeem·Created about a month ago·Updated 3 days ago· 17

image
guider
sampler
sigmas
vae
noise
clip
mask

IMAGE

◄max_tile_width1024►

◄max_tile_height1024►

◄context_anchor32►

◄context_overlap32►

◄anchor_sourcesource image►

◄vlm_methodvision tokens and captions►

Categoryimage/upscaling

Inputs (14)

Name	Type	Default	Description
image	IMAGE		The image to refine. Upscale it before this node.
guider	GUIDER		The guider that denoises each tile.
sampler	SAMPLER		The sampler used to denoise each tile.
sigmas	SIGMAS		The sigma schedule used when sampling each tile.
vae	VAE		The VAE that encodes and decodes each tile.
noise	NOISE		Noise is drawn once for the entire image and then sliced for each tile.
max_tile_width	INT	1024256–16384	Hard cap on the width the model ever sees per sampled crop, including the context_overlap and context_anchor rings. Set to the largest width the model supports.
max_tile_height	INT	1024256–16384	Hard cap on the height the model ever sees per sampled crop, including the context_overlap and context_anchor rings. Set to the largest height the model supports.
context_anchor	INT	320–512	Pixels around each tile that are frozen and shown to the model as context, then cropped away. With a mask it is also the frozen background the region is refined against, so keep it above 0.
context_overlap	INT	320–512	Overlapped context that is diffused from both sides and then blended. It anchors the tiles to each other, like context_anchor anchors each tile to its surroundings.
anchor_source	COMBO	source image	What fills context_anchor. source image keeps the result true to the input. live canvas adds more detail and drifts further from the input.
vlm_method	COMBO	vision tokens and captions	Whether each tile is conditioned on a caption of itself, on its slice of the entire image's vision encode, or on both. The name in parentheses is the caption preset it asks. Copy settings.toml to settings.user.toml to write your own tile prompts.
clip	CLIP		Must be a vision-language text encoder (Krea 2 family). The guider's positive prompt is ignored and its negative still applies.
maskopt	MASK		Only the masked region is refined and the rest is left untouched. Feed an inverted mask for a second pass.

Outputs (1)

Name	Type	Description
IMAGE	IMAGE	—