ReChromaPatcher

ReChromaPatcher is a model patcher for Chroma. You pass a Chroma MODEL through it, it wires in the hooks RES4LYF needs, and you get a patched MODEL back to run through the sampler. It belongs to the author's little family of "Re-" patchers - ReFluxPatcher, ReWanPatcher - each one adapting a specific architecture so the pack's deeper features (style transfer, attention-level guidance, higher-precision style ops) have somewhere to attach.

Context on Chroma, since it's the whole point of this node. Chroma is the 8.9B community model built on Flux.1-schnell by LodestoneRock - fully open, uncensored, and the main reason people run Flux-architecture models outside of Flux Dev itself. RES4LYF explicitly lists Chroma among its best-supported models for style transfer, alongside HiDream. But Chroma isn't identical to Flux under the hood, so to get the pack's style and guide machinery working on it cleanly, the model needs a patch. That's this.

How it works

The node injects RES4LYF's model-side hooks into a Chroma checkpoint so the pack's guide and style nodes can reach inside the network's blocks during sampling. It also lets you set the precision those style operations run at - the default is float64, which is unusually high and is there because style-matching math (normalizing and transplanting feature statistics) is sensitive to numerical error, and Chroma's style work benefits from the headroom. There's a simple on/off so you can bypass the patch without unplugging the node.

The inputs and output that matter

• model (MODEL) - your loaded Chroma model. Required.
• style_dtype (enum: default, bfloat16, float16, float32, float64; default float64) - the precision for the pack's style operations. Leave it at float64 unless you're tight on memory or chasing speed, in which case step down. This is the only real tuning choice.
• enable (BOOLEAN, default true) - toggle the patch. Flip it off to compare against unpatched behavior.

Output is a single model (MODEL). Feed it into your ClownsharKSampler and your RES4LYF style/guide nodes; it slots in exactly where the unpatched model would.

How to install it

ComfyUI Manager: search RES4LYF, install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/ClownsharkBatwing/RES4LYF/, activate your venv, pip install -r requirements.txt (portable ComfyUI: embedded python's pip), restart, hard-refresh with F5. You'll separately need a Chroma checkpoint loaded (Chroma1-HD is the usual generation model) - this node patches it but doesn't download it.

Common issues

First: feed it Chroma, not something else. This patcher is shaped for the Chroma architecture, so handing it a plain Flux Dev or an SDXL model is asking for either a silent no-op or an error. If you're on Flux, the sibling ReFluxPatcher is the one you want; on Wan, ReWanPatcher.

Second: mind float64. Double precision on style ops is heavier on VRAM and can be slower, and not every GPU is happy running float64 kernels at speed. If you hit an out-of-memory wall or things crawl the moment style guides kick in, drop style_dtype to float32 or bfloat16 and see if the quality difference is one you can live with - often it is.

And the usual RES4LYF caveat applies: this node only earns its keep if you're actually using the pack's style-transfer or guide nodes on Chroma. If you just want to sample Chroma normally with a res solver and beta57, you don't need to patch anything - plug the model straight into the sampler and skip this entirely.