Fooocus Upscale

Every ComfyUI user eventually builds the same graph by hand: generate at native resolution, then run a second pass at higher resolution with low denoise to add detail that doesn't fall apart. That's hi-res fix, and it's exactly what this node is - pre-wired, so you don't have to remember the settings. If you've been sampling at 1024x1024 and upscaling the PNG afterwards with an ESRGAN model, stop. This is the step you're missing.

How it works

The node takes the pipe from your completed generation, upscales the latent (or the optional image you feed in), and runs a fresh low-denoise sampling pass at the bigger size. The first pass locked in composition; the second pass invents coherent detail instead of just stretching pixels. That's the difference between "bigger" and "better" - the whole reason hi-res fix exists and why generation above native resolution produces tiling artifacts without it.

The inputs

• pipe - the pipeline from the KSampler (or anywhere upstream). Required.
• image (optional) - feed an image instead of re-sampling from the pipe, if you just want a Fooocus-style second pass on something you already have.
• upscale - 1.5 or 2. The sweet spot from the upscaling playbook is 1.5–2x; pick 1.5 for faces you want to keep recognisable.
• steps - default 18 for the second pass. You don't need as many as the first pass; this is refinement, not composition.
• denoise - default 0.382, and that number is a Fooocus fingerprint. The general rule is 0.3–0.5 for hi-res fix: high enough to add real detail, low enough to keep the composition. 0.382 sits right in that band.
• fast - default off. On runs a cheaper variant when you're iterating and don't care about the last few percent.
• image_output - Hide, Preview, Save, or Hide/Save. The pack's standard output switch.
• save_prefix - filename prefix when saving.

Outputs: a pipe (PIPE_LINE) and an image (IMAGE), so you can either continue the graph or hang a Save node off it.

Installing

Same pack, same path: search "Fooocus Nodes" in ComfyUI Manager, or clone the repo into custom_nodes. No extra model download for this node - it reuses whatever checkpoint is in your pipe. Note the README's clone URL currently 404s because the author's GitHub account was renamed or removed; Manager is the reliable route.

The honest caveat

This is a generative upscaler, which means faces get subtly rewritten on the way up - every diffusion rung changes identity a little. If the subject matters, run the pack's Fooocus detailerFix after the upscale to re-fix the face at native detail, or drop to 1.5x and accept the drift. And if your source is already sharp and you just need pixels, you don't need this node at all - that's what Lanczos is for.