ClownGuide Style

This is RES4LYF's style-transfer node - hand it a reference image and it pulls the look of that image (its palette, lighting, texture, the painterly-vs-photo feel) into whatever you're generating, without dragging the reference's actual content along. No LoRA, no IPAdapter model to download. It's one of the pack's genuinely standout features, and the author spent a chunk of mid-2025 showing it off across ten different model families.

The node's own description is refreshingly honest about what it does: it transfers "some visual aspects of style from a guide image," and if your prompt says nothing about style, "it may just transfer the lighting and color scheme." That's the right expectation. Style transfer here works with your prompt, not instead of it - the README's own advice is to mention the style you want in the prompt too, and to run once with the guide off to check the prompt isn't already doing the heavy lifting. Supported models are broad: SD1.5, SDXL, Stable Cascade, SD3.5, AuraFlow, Flux, HiDream, Wan and LTXV, with HiDream and Chroma giving the best results in practice (plain Flux Dev is weak on style unless you add a style LoRA).

How it works

You encode your reference image to a latent and feed it as the guide. During sampling the node nudges the denoising toward the statistical "style" of that guide using a chosen method. It outputs a GUIDES bundle that plugs into a RES4LYF sampler, so the transfer happens inside the sampling loop rather than as a post-process.

The inputs and outputs that matter

• guide (LATENT) - your style reference, VAE-encoded. This is the image whose look you're borrowing.
• method (default WCT) - how style gets matched. WCT/WCT2 and AdaIN are classic feature-statistics transfers; scattersort is ClownsharkBatwing's own mode, which he wrote specifically to stop transfers from dragging color and lighting off the source. If WCT bleeds too much of the reference in, try scattersort.
• weight (default 1) - strength of the transfer. Dial down if the style overwhelms your content.
• apply_to (positive / negative / denoised) - which side of guidance the style rides on. Default positive is what you want almost always.
• guides (GUIDES) in / guides (GUIDES) out - chain this through other guide nodes; the output goes to your sampler's guides input.
• Optional mask - restrict the style to part of the image.

How to install it

ComfyUI Manager: search RES4LYF, install, restart. Manual:

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

then restart and hard-refresh (F5). No extra models for the node itself - you'll just need a VAE to encode your reference.

Common issues & troubleshooting

The output burns, or the preview goes very dark/bright then produces garbage. This is the documented failure mode, and the fix is in the node's own description: duplicate this node and chain it so the guide is applied to both positive and negative conditioning (set the second copy's apply_to to negative). It usually settles the burn caused by CFG.

It transferred the content, not just the style. Lower weight, and switch method to scattersort - that mode exists precisely to keep the reference's colors and lighting from crossing over.

Nothing changed. Confirm the guide actually did something by running once with the style off; if the "styled" and "unstyled" results are identical, your prompt is probably already dictating the look, or the guide latent isn't wired in. Also check you're on a supported model - style is weakest on plain Flux Dev.