ResShift SR Upscale (1-step)

This is where the Distilled ResShift pack does its party trick: it upscales any RGB image ×4 or ×2 in one diffusion step. Not fifteen, not eight - one. There's no sampler dropdown, no CFG, no scheduler anywhere in the node. You feed it a RESSHIFT_MODEL (from ResShiftLoader) and an IMAGE, and it hands back the upscaled IMAGE. That's the whole interface.

What it actually is

The student is distilled from a 15-step ResShift teacher into a single stochastic step. That's the trade you should know going in: distillation collapses the whole denoising trajectory into one jump, and distilled models always give up some refinement for the speed. So this won't rebuild a face the way a heavy restoration model will - which, for most upscaling jobs, is a feature, not a bug. In the upscaling taxonomy (more pixels vs. more detail vs. video), this is a detail model that's fast enough to feel like an ESRGAN-class upscaler: seconds, not the multi-minute grind of a SeedVR2 pass.

It's also genuinely model-agnostic. Because it's a pixel-space IMAGE → IMAGE node, it doesn't care what generated the image. Drop it after VAEDecode and before SaveImage, and it'll happily upscale anime, photos, or a random PNG you loaded from disk.

How the one step works

The student runs at the same spatial resolution - the upscale lives in the residual-shift, not in a spatial resize. The image gets bicubic-upsampled ×scale, encoded to the vq-f4 latent, pushed through one student step (z_T = z_y + κ·randn, plus injected noise ε), and decoded. Large images are split into tiles with overlap-averaged blending, and each tile is reflect-padded to a multiple of the Swin alignment (256 px at ×4, 128 px at ×2).

The inputs that matter

Most of the defaults are fine to leave alone. The ones you'll actually touch:

• seed (default 0) - the step is stochastic, so the same seed plus the same input gives the same output. Change it if you want a different rendering of detail; keep it if you want reproducibility.
• chop (default 512) - tile size for large images, snapped to the 256 alignment. An image at or under chop runs as a single tile. This is your VRAM lever - if you're OOM, lower it. Bigger tiles mean more context per tile and weaker seams.
• overlap (default 64) - the seam overlap between tiles (tile stride = chop - overlap), and also the feather ramp the tile weights blend across. Larger = fewer seams but more redundant compute; 0 gives hard-abutted tiles.
• tile_batch (default 4) - how many tiles go into one forward pass. Small tiles underfill the GPU one at a time, so batching amortizes launch overhead. VRAM scales roughly linearly with it; tune to your card.

The one optional input, shared_noise (default true), draws a single full-image noise field and shares it across overlapping tiles so the seam-average is discontinuity-free - it kills the faint tile-block seams you'd otherwise see in flat regions. Leave it on; it only matters when the image is actually tiled.

Output is IMAGE at ×scale, ready for SaveImage or further processing.

Installing it

Same pack, same install as the loader:

# ComfyUI Manager: search "Distilled ResShift SR"
# ...or by hand:
cd ComfyUI/custom_nodes
git clone https://github.com/sorryhyun/ComfyUI-Distilled-ResShift

Restart, and the first load auto-downloads the student (~478 MB) and the vq-f4 VQGAN (~211 MB) into ComfyUI/models/resshift/. Dependencies (einops, timm, omegaconf, huggingface_hub) auto-install from pyproject.toml, and there's no xformers requirement. Keep the loader's dtype on bf16 - it's ~2x faster and half the VRAM, and it matches how the student was trained.

Common issues

• Out of VRAM on big images - lower chop, then tile_batch, before touching anything else. That's the whole story.
• Faint seams in flat skies/gradients - make sure shared_noise is on and raise overlap a bit.
• Output keeps changing - that's the stochastic step. Same seed + same input = same output; different seed = different detail.
• Wrong scale factor - that's a loader-side scale/student mismatch, not this node. Check ResShiftLoader.

One step, no sampler, no CFG, no key. It's the rare distilled model that just works out of the box.