Distilled ResShift SR

ComfyUI — Distilled ResShift SR (×4/×2, 1-step)

A 1-step ResShift super-resolution student as a pixel-space IMAGE → IMAGE upscaler for ComfyUI, at ×4 or ×2 (a scale dropdown on the loader). The student is distilled (paper 2603.22490, repo, One-Step Residual Shifting Diffusion via Distillation) from a 15-step ResShift teacher, so it does the whole super-resolution jump in a single stochastic network step — no multi-step sampler, no CFG.

It is model-agnostic: drop it after VAEDecode / before SaveImage and it upscales any RGB image, not just Anima / diffusion output.

… → VAEDecode → IMAGE ─┐
                       ├─► ResShiftUpscale ─► IMAGE (×scale) → SaveImage
ResShiftLoader ───┘

Nodes

| Node | In → Out | Notes | |---|---|---| | ResShift SR Loader (distilled ×4/×2) | → RESSHIFT_MODEL | Loads the student + vq-f4 VQGAN. scale: x4 (default) or x2 — picks the auto-download repo AND the diffusion schedule. Pick (auto-download) for student_name to fetch the scale's released student, or select a local checkpoint from ComfyUI/models/resshift/. dtype: bf16 (default, ~2× faster / ~half VRAM) or fp32. | | ResShift SR Upscale (1-step) | RESSHIFT_MODEL + IMAGE → IMAGE | seed (the step is stochastic), chop / overlap / tile_batch for tiling large images. |

The scale you pick on the loader must match the student it loads — the safetensors header carries the arch (noise_mode/noise_channels) but not the SR factor, so a ×4 checkpoint run under scale=x2 (or vice-versa) upscales at the wrong ratio. The (auto-download) path always fetches the matching student, so it can't mismatch.

Weights (auto-downloaded on first use)

Both land in ComfyUI/models/resshift/:

• Student (~478 MB, ema-only .safetensors) — per scale, from the public HF repos sorryhyun/Distilled-ResShift-4x (×4) and sorryhyun/Distilled-ResShift-2x (×2). Arch metadata (noise_mode / noise_channels) is carried in the safetensors header, so the loader rebuilds the network from the file alone.
• vq-f4 VQGAN (~211 MB) — from the upstream ResShift v1.0 GitHub release (zsyOAOA/ResShift/releases/download/v1.0/autoencoder_vq_f4.pth). If that download is blocked in your environment, place the file in models/resshift/ by hand.

To use your own student, drop a .safetensors (or a training .pth with an ema key) into models/resshift/ and select it in the loader.

How it works

The student is spatially same-resolution — the upscale lives in the residual-shift, not a spatial resize. Each image is bicubic-upsampled ×scale, encoded to the vq-f4 latent, run through one student step (z_T = z_y + κ·randn, plus injected noise ε), and decoded. Large images are tiled (ImageSpliterTh, overlap-averaged); each tile is reflect-padded to a multiple of the Swin alignment (256 px at ×4, 128 px at ×2) — chop is snapped to that automatically.

bf16 runs the student under bf16 autocast and the VQGAN in native bf16 (keeping the full-resolution GroupNorm activations — the VRAM peak — out of fp32).

Dependencies

torch / numpy / safetensors ship with ComfyUI. The vendored ResShift network also needs einops, timm, omegaconf, and huggingface_hub (auto-installed from pyproject.toml). No xformers is required — the vendored VQGAN mid-attention uses a query-chunked exact-SDPA path that runs on any GPU (including Blackwell / sm_120).

Licensing

Node code: see LICENSE. The vendored ResShift network under _vendor/resshift/ is S-Lab License 1.0 (non-commercial), from zsyOAOA/ResShift. The distilled student weights inherit the same non-commercial terms.