Icy StyleModelApply

IcyStyleModelApply is the IcyHider clone of StyleModelApply - and unlike most of the pack's clones, this one lives in a corner of ComfyUI that even experienced users trip over, because it needs three things that don't usually sit together: a conditioning, a STYLE_MODEL, and a CLIP_VISION_OUTPUT.

What a "style model" is

Style models are small companion networks (the lineage runs through StyleAligned and similar research) that transfer how something looks - brushwork, palette, texture, composition mood - from a reference image into a generation, without copying the subject. The style signal isn't text, so it can't ride along in your prompt; it comes from a CLIP vision model looking at an actual image. That's the part people miss.

So the node has three inputs:

• conditioning - your text conditioning (usually positive).
• style_model - loaded by a StyleModelLoader from a .safetensors file in your models folder. Not a checkpoint, not a LoRA - its own file type.
• clip_vision_output - the output of CLIPVisionEncode, which ran a CLIP vision model over your reference image. If this is empty or wired wrong, the node quietly produces a no-op.

Then there are the two knobs: strength (how much style, default 1.0, and on most setups you'll tune below 1.0 so the style doesn't swamp the subject) and strength_type, which is the dropdown that confuses everyone: multiply vs attn_bias. It selects how the style features get injected into the attention layers - multiply scales the features into the attention weights, attn_bias adds them as a bias term. Community experience: attn_bias tends to give a more pronounced, sometimes more stable transfer on SDXL-style models, while multiply is subtler and blends better at low strength. It's worth A/B-ing on your model because there's no universal winner.

Output is a single CONDITIONING that feeds the sampler.

The realistic picture

This is a niche, somewhat fiddly node. Style models exist for SD 1.5 and SDXL, and for the model families that dominate 2026 the file you need may simply not exist - there's no universal "style model for Flux." If you want reference-image style transfer on a modern base, IP-Adapter-style approaches are usually the more traveled road. Reach for this when you have an actual style model file in hand and a CLIP vision model loaded; that's the workflow it was built for.

Install and the "Icy" part

Pack install: ComfyUI Manager → search "IcyHider" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/icekiub-ai/ComfyUI-IcyHider.git

No Python deps, no bundled models (the style model and CLIP vision files are yours to source). The node appears under IcyHider Comfy Core. As with every non-image output in this pack, the wrapper adds no hiding value here - the underlying StyleModelApply is exactly what runs. If it produces nothing, the wrapper isn't your problem; check your clip_vision_output wiring first, then the strength.