Nodes/RES4LYF/ReFluxPatcher
ComfyUI Node Runs on cloud

ReFluxPatcher

Turn Flux into a style/reference engine for RES4LYF

By ClownsharkBatwing·Created 2 years ago·Updated 18 days ago· 1,222
ReFluxPatcher
  • model
  • model
style_dtypefloat64
enabletrue

If you're trying to do zero-shot style transfer or a reference-guided Flux generation with RES4LYF and the workflow keeps erroring on the model, this is the node you were missing. ReFluxPatcher patches a plain Flux model so the pack's guide and style-transfer machinery can actually hook into it. Without the patch, the fancy stuff (composition guides, "scattersort" style mode, faceswap-by-inpainting) has nothing to attach to.

There's a nice bit of backstory here. RES4LYF's author, ClownsharkBatwing, wrote ReFluxPatcher because the existing Flux PuLID/Redux path was broken and he fixed it from the outside rather than waiting on an upstream repo. That's the whole vibe of this pack: capabilities first, ship the patch, explain it in a comment thread later.

What it actually does

It's a model patcher, same family as the ModelSampling nodes: model in, model out, nothing destructive. It rewires the Flux transformer so RES4LYF can inject reference/style information at sample time. You drop it between your Flux checkpoint and your sampler, and everything downstream (a ClownGuide node, the Shark/ClownsharK sampler) can now do image-guided work on Flux that stock Flux can't.

The inputs and outputs that matter

There are only three, and you'll leave two of them alone:

  • model (MODEL) - your Flux model. The single output is also model (MODEL), which you wire onward into your sampler or the next patcher.
  • style_dtype (default float64) - the precision the style computation runs at. float64 is the default for a reason: RES4LYF leans on high precision for accuracy, and this is where it matters. Drop it to float32 or bfloat16 only if you're tight on VRAM or something is choking; expect a small quality/consistency cost.
  • enable (default true) - an on/off toggle so you can A/B the patch without unplugging the node. Flip it off and you get plain Flux back, which is genuinely useful for the sanity check the README begs you to run: turn the guide off and confirm your prompt isn't secretly doing the work.

That's it. No hidden parameters, nothing to tune to death.

How to install it

Easiest path is ComfyUI Manager: open it, search RES4LYF, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ClownsharkBatwing/RES4LYF
cd RES4LYF
pip install -r requirements.txt

On portable ComfyUI, swap pip for your embedded pip (e.g. python_embedded\Scripts\pip.exe). Restart ComfyUI and hard-refresh the page with F5, not just the reload button. There are no models to download for the node itself - it patches the Flux you already loaded.

Common issues & troubleshooting

The node doesn't show up. You skipped pip install -r requirements.txt, or you soft-reloaded instead of restarting. Do both, then F5.

float64 is crawling or blowing up memory. Not every GPU is happy running float64, and on some cards it's dramatically slower than float32. If sampling times cratered after you added this node, set style_dtype to float32 and see if the quality difference is even visible.

It does nothing. ReFluxPatcher is an enabler, not an effect - on its own it just hands back a patched model. The actual style/guide result comes from the guide node and sampler you feed it into. If your output looks identical with enable on and off, check that a ClownGuide (or the sampler's guide input) is actually wired up and weighted above zero.

Wrong model family. This is the Flux patcher. Running Wan? Use ReWanPatcher instead; the pack ships a per-architecture patcher rather than one universal node.

CategoryRES4LYF/model_patches

Inputs (3)

NameTypeDefaultDescription
modelMODEL
style_dtypeCOMBOfloat645 options: default, bfloat16, float16, float32, float64
enableBOOLEANtrue

Outputs (1)

NameTypeDescription
modelMODEL