Latent Upscale (WIP Exp.)

The display name is honest about this one: Latent Upscale (WIP Exp.). It resizes a latent in latent space - the "stay out of pixel space" trick that keeps big upscales fast - but unlike ComfyUI's plain latent resize, it actually tries not to smear your edges. It's experimental, so treat it as a toy with promise, not a production workhorse.

How it works

The trick is splitting the latent into a low-pass and high-pass component before resizing, then treating them differently:

• Split: blur the latent with a Gaussian of sigma hp_sigma (default 0.5) to get the low-pass; the high-pass is original minus blurred. hp_sigma 0 disables the split entirely.
• Resize: in hybrid mode (default), the low-pass goes up with bicubic + antialiasing (smooth, no shimmer) while the high-pass goes up with nearest-neighbor (crisp, no smearing), then the high-pass std is rescaled to match the original via variance matching.
• Optional renoise: renoise (0–1) adds high-frequency noise back, and if current_sigma > 0 it's scaled by that sigma - meaning you can feed this node mid-sampling and add noise proportional to how much denoising remains.

The other modes are plain bicubic, bilinear, or nearest for the whole thing, with antialias where supported.

Inputs and outputs

• latents - a LATENT dict. Handles the standard [B,C,H,W], and also squeezes a singleton frame dimension from [B,C,1,H,W] latents (e.g. Qwen-style), which is a nice compatibility touch.
• scale_by (default 2) - upscale factor in latent space, 1–8.
• Optional: mode, antialias, hp_sigma, renoise, current_sigma, seed.

Output: LATENT, same shape as input scaled by scale_by.

When to use it

If you're building a latent-space hi-res-fix pipeline - upscale the latent, run a second denoise pass at low denoise - this is the resize step. It fits the two-pass philosophy from the upscaling playbook: get the pixels bigger in latent space, then let a second sampler add detail back. The variance-matched hybrid mode is the reason to pick this over core's plain latent resize.

Installing

Part of WAS Affine by WASasquatch. ComfyUI Manager → Install Custom Nodes → search "WAS Affine" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was_affine

Restart ComfyUI. ComfyUI 1.0.0+, torch and numpy.

Common issues

It's marked WIP for a reason: latent-space resizing is inherently lossy vs. pixel-space upscalers, and you can get mushy structure or artifacts that a real ESRGAN pass wouldn't have. If the result looks soft, try renoise above 0 with a current_sigma set - the renoise path is exactly for re-seeding texture. And don't reach for this when you want a final high-quality upscale; that's the job of pixel-space upscalers, not a latent resize.