Nodes/RES4LYF/UltraSharkSampler Tiled
ComfyUI Node Runs on cloud

UltraSharkSampler Tiled

The tiled UltraCascade sampler in RES4LYF

By ClownsharkBatwing·Created 2 years ago·Updated 18 days ago· 1,222
UltraSharkSampler Tiled
  • model
  • positive
  • negative
  • sampler
  • sigmas
  • latent_image
  • latent_noise
  • guide
  • guide_weights
  • image_cv
  • LATENT
add_noisetrue
noise_is_latentfalse
noise_type
alpha1.0
k1
noise_seed0
cfg1.00
guide_type
guide_weight0.00
tile_width1024
tile_height1024
tiling_strategy
max_tile_batch_size64
clip_nameclip-vit-large-patch14.safetensors
strength1.00
noise_augment1.00

Niche node, clear job: it's a tiled sampler for UltraCascade, the Stable Cascade project by the same author as RES4LYF. The node's own description says it plainly - "For use with UltraCascade." The "Tiled" part is the reason it exists: it splits a large canvas into tiles and samples them in batches, so you can push resolution higher than a single pass would fit in VRAM, then stitch the result.

Some background, because this one sits off the beaten path. ClownsharkBatwing built UltraCascade (a fork line off ComfyUI-UltraPixel) alongside RES4LYF - Stable Cascade has an unusually good grasp of style and holds up at high resolution, which is what UltraPixel-style workflows chase. This sampler brings the pack's SharkSampler machinery - custom SAMPLER, custom SIGMAS, controllable noise - to that Cascade workflow, with tiling bolted on for big images. It lives under the pack's legacy tree, so treat it as stable-but-specialist: if you're not running UltraCascade, this isn't your node.

How it works

It's a sampler that tiles. You give it a Cascade model, conditioning, a SAMPLER and SIGMAS, plus a tile size and strategy; it denoises the canvas tile-by-tile in batches and returns a single LATENT.

The inputs and outputs that matter

  • model, positive, negative, sampler (SAMPLER), sigmas (SIGMAS), latent_image - standard sampler wiring, Cascade-side.
  • LATENT out - the assembled result.
  • tile_width / tile_height (default 1024) - the tile size. Smaller tiles fit smaller cards but mean more of them; this is your main VRAM-vs-speed lever.
  • tiling_strategy (padded / random / random strict / simple) - how tiles are laid out and blended. padded overlaps to hide seams; the random strategies jitter tile placement to avoid a visible grid.
  • max_tile_batch_size (default 64) - how many tiles run at once. Drop it if you OOM.
  • cfg (default 1) and the noise controls (noise_type, noise_seed, etc.) behave as in the pack's other samplers. Optional guide (LATENT) and image_cv (IMAGE) feed the Cascade-specific guidance path.

How to install it

ComfyUI Manager: search RES4LYF, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ClownsharkBatwing/RES4LYF
cd RES4LYF
pip install -r requirements.txt

then restart and hard-refresh (F5). You'll also need UltraCascade and its Stable Cascade models set up separately - this node samples them but doesn't fetch them.

Common issues & troubleshooting

Visible seams or a grid pattern. That's a tiling artifact. Switch tiling_strategy to padded for overlap-blended tiles, or one of the random modes to break up a regular grid. Larger tiles also mean fewer seams.

Out of memory despite tiling. Tiling caps the canvas cost, but each batch of tiles still uses VRAM. Lower max_tile_batch_size first, then tile_width/tile_height.

Wrong model. It's for UltraCascade specifically - point it at an ordinary SD/Flux checkpoint and it won't behave. If you just want a general tiled/high-res result on a mainstream model, this isn't the tool; it's a Cascade-workflow node.

CategoryRES4LYF/legacy/samplers/ultracascade

Inputs (26)

NameTypeDefaultDescription
add_noiseBOOLEANtrue
noise_is_latentBOOLEANfalse
noise_typeCOMBO17 options: fractal, gaussian, gaussian_backwards, uniform, pyramid-cascade_B, pyramid-interpolated, +11
alphaFLOAT1.0-10000–10000
kFLOAT1-10000–10000
noise_seedINT00–18446744073709550000
cfgFLOAT1.000–100
guide_typeCOMBO2 options: residual, weighted
guide_weightFLOAT0.00-100–100
tile_widthINT10242–16384
tile_heightINT10242–16384
tiling_strategyCOMBO4 options: padded, random, random strict, simple
max_tile_batch_sizeINT641–256
modelMODEL
positiveCONDITIONING
negativeCONDITIONING
samplerSAMPLER
sigmasSIGMAS
latent_imageLATENT
clip_nameCOMBOclip-vit-large-patch14.safetensors0 options:
strengthFLOAT1.00-10–10
noise_augmentFLOAT1.000–1
latent_noiseoptLATENT
guideoptLATENT
guide_weightsoptSIGMAS
image_cvoptIMAGE

Outputs (1)

NameTypeDescription
LATENTLATENT