Nodes/ComfyUI vsLinx Nodes/MultiDiffusion Tiled Hires Fix
ComfyUI Node

MultiDiffusion Tiled Hires Fix

Model-agnostic MultiDiffusion tiled hires-fix / refiner. Runs one sampling pass over the whole image, splitting the latent into overlapping tiles every denoising step and averaging the overlaps in latent space, so there are no tile seams and per-step VRAM stays tile-sized. Upscale the image first, then refine it here with a low denoise. Works on any model (SD/SDXL/Flux/etc.) - no extra node packs.

By vslinx·Created 12 months ago·Updated 27 days ago· 20
MultiDiffusion Tiled Hires Fix
  • image
  • model
  • positive
  • negative
  • vae
  • image
seed0
steps20
cfg7.0
sampler_name
scheduler
denoise0.50
rows2
columns2
overlap0.00
overlap_x64
overlap_y64
vae_decode_tiledfalse
vae_decode_tile_size512
CategoryvsLinx/sampling

Inputs (18)

NameTypeDefaultDescription
imageIMAGE
modelMODEL
positiveCONDITIONING
negativeCONDITIONING
vaeVAE
seedINT00–18446744073709550000
stepsINT201–10000
cfgFLOAT7.00–100
sampler_nameCOMBO44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38
schedulerCOMBO9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3
denoiseFLOAT0.500–1Denoise strength. For a tiled hires-fix on an already-upscaled image keep this low (e.g. 0.3-0.5).
rowsINT21–256
columnsINT21–256
overlapFLOAT0.000–0.5Overlap between tiles as a fraction of tile size, added on top of overlap_x/overlap_y.
overlap_xINT640–8192Extra horizontal overlap in pixels.
overlap_yINT640–8192Extra vertical overlap in pixels.
vae_decode_tiledBOOLEANfalseDecode the final full-image latent in tiles instead of one pass, to avoid a single huge VAE decode that can spike VRAM (or, on Windows, spill into slow shared system memory). The decode is a single full-image pass independent of rows/columns, so adding tiles won't shrink it - enable this instead.
vae_decode_tile_sizeINT51264–4096Tile size in pixels for the tiled VAE decode.

Outputs (1)

NameTypeDescription
imageIMAGE