Ultimate SD Upscale node in ComfyUI

Ultimate SD Upscale enlarges a picture with an ESRGAN model, then re-samples it tile by tile through the diffusion model with your prompt at low denoise, adding real detail at 2x-4x without running out of VRAM.

Node pack: ComfyUI_UltimateSDUpscale

Category: Upscaling

What Ultimate SD Upscale does

The node from the ComfyUI_UltimateSDUpscale pack (a port of the A1111 script) combines two ideas: a pixel upscaler enlarges the image, then the result is split into tiles of tile_width x tile_height, and each tile is re-rendered by the sampler with the prompt at a low denoise (0.2-0.4), so the model paints in texture, skin, fabric and foliage detail. Overlapping tiles and an optional seam fix pass hide the joins.

Because each tile is a normal sampler pass at model resolution, the node works on any GPU that can render a single picture, regardless of the final size. It takes the same MODEL, positive, negative and VAE as your main KSampler plus an upscale model, and is often paired with a Tile ControlNet to keep each tile faithful to the original.

Inputs

Name	Type	What it is
image	IMAGE	The picture to upscale.
model / positive / negative / vae	MODEL, CONDITIONING, VAE	Same connections as KSampler.
upscale_model	UPSCALE_MODEL	ESRGAN model for the first pixel upscale.
upscale_by	FLOAT	Final factor, typically 2.0.
seed / steps / cfg / sampler_name / scheduler / denoise	sampler settings	Denoise 0.2-0.4 keeps the picture; higher values redraw it.
mode_type	COMBO	Linear, Chess (recommended) or None (upscale only).
tile_width / tile_height / mask_blur / tile_padding	INT	Tile size (1024 for SDXL, 512-768 for SD1.5), blur and overlap.
seam_fix_mode	COMBO	None, Band Pass, Half Tile, Half Tile + Intersections.
force_uniform_tiles / tiled_decode	BOOLEAN	Equal tiles; tiled VAE decode for low VRAM.

Outputs

Name	Type	What it is
IMAGE	IMAGE	The upscaled, re-detailed picture.

Skip the setup: ComfyUI workflows preinstalled on the cloud

BitVector runs ready-made ComfyUI workflows on its own GPUs. No install, no missing nodes, no red boxes. Open it on your phone or laptop and generate in a minute.

Try BitVector

How to use Ultimate SD Upscale

Install the pack through ComfyUI Manager (search "Ultimate SD Upscale").
Connect the decoded IMAGE, the same model/positive/negative/VAE as your sampler, and an upscale model.
Set upscale_by 2.0, denoise 0.3, steps 20, mode Chess, tile 1024 for SDXL.
Optionally add Apply ControlNet with a Tile ControlNet to the positive before this node.
Queue, then compare with the input at 100% zoom; raise denoise for more detail, lower it for fidelity.

Settings and tips

Keep the prompt generic ("highly detailed, sharp focus") or reuse the original; detailed subject prompts can repeat the subject in every tile.
Denoise above 0.45 without a Tile ControlNet invents faces and objects in tiles.
Chess mode plus Half Tile seam fix hides nearly all joins.
For 4x, run the node twice at 2x rather than once at 4x.
tiled_decode saves VRAM at the final decode on 8 GB cards.

Troubleshooting Ultimate SD Upscale

Faces or objects appear in textures (grass, walls)

Why it happens

Denoise too high for the prompt and no Tile ControlNet to anchor the content.

How to fix it

Lower denoise to 0.25-0.35, shorten the prompt, or add a Tile ControlNet at strength 0.5-0.8.

Visible grid of tiles

Why it happens

seam_fix_mode None, small tile_padding, or Linear mode with a strong denoise.

How to fix it

Use Chess mode, tile_padding 32-64, mask_blur 8-16 and seam fix Half Tile.

Very slow: hundreds of tiles

Why it happens

A large upscale_by with small tiles multiplies the number of sampler passes.

How to fix it

Use the largest tile the model supports (1024 for SDXL) and upscale in two 2x stages.

The node is missing after import (red node)

Why it happens

The pack is not installed or failed to import because of a missing dependency.

How to fix it

Install "ComfyUI_UltimateSDUpscale" in ComfyUI Manager, restart, and check the console for import errors.

Run this workflow from your phone

Every workflow on this page is preinstalled on BitVector with the models and custom nodes already in place. Pick one, type a prompt, done. Zero setup, nothing to download.

Try BitVector

Questions about Ultimate SD Upscale

Do I need the Tile ControlNet?

Not required, but it allows a higher denoise without content drift. For SDXL use a Tile or TTPlanet tile model.

Can I use it without an upscale model?

Yes: set upscale_model to None (or use Ultimate SD Upscale (No Upscale) on an already enlarged picture).

Flux support?

Yes, with cfg 1.0 and Flux Guidance in the positive conditioning; tiles of 1024 and denoise around 0.3.

Related nodes

Upscale Image (using Model)

ComfyUI core

Upscale Image (using Model) runs an ESRGAN-style upscaler (4x-UltraSharp, RealESRGAN, 4x_NMKD-Siax, 4x_foolhardy_Remacri) over an IMAGE and returns it enlarged by the model's fixed factor with added detail.

Apply ControlNet

ComfyUI core

Apply ControlNet attaches a guide image and a loaded ControlNet to the positive and negative CONDITIONING with strength and start/end percentages, so the sampler follows edges, depth or pose from the guide.

KSampler

ComfyUI core

KSampler runs the denoising loop: it takes the model, prompts and a latent and produces the finished latent image, controlled by seed, steps, cfg, sampler, scheduler and denoise.

Upscale Latent and Upscale Latent By

ComfyUI core

Upscale Latent resizes a latent to a new size (or by a factor) so a second KSampler can add detail at higher resolution: the classic hires fix.

FaceDetailer (Impact Pack)

ComfyUI-Impact-Pack

FaceDetailer finds faces (or hands, eyes) with a YOLO detector, crops each one, re-renders it at full resolution with your model and prompt, and pastes it back, fixing the small, mushy faces in wide shots.

Tired of fixing nodes? Let the cloud do it

BitVector keeps hundreds of ComfyUI workflows installed, updated and tested on fast cloud GPUs. No Python, no CUDA errors, no VRAM limits. Works from any browser.

Try BitVector