🚀 Apply Style Model (Simple)

Style transfer for Flux usually means one of two things: an IP-Adapter port with a dozen parameters to fight, or the official Redux adapter whose default workflow makes everything a variation of the input image whether you asked for it or not. This node is the "I just want to steer how strong the style is" middle ground. Give it a CLIP Vision model, a style model, and an image, and it returns a CONDITIONING that carries the image's style into the sampler - with a strength control that actually does something.

It sits in the same Redux family as the FLUX.1 Redux image-variation adapter (SigLIP-encoded image conditioning appended to your text prompt). The difference is you get a real dial: Redux fans have long noted that strength is everything with these adapters - go too high and the prompt gets overwhelmed and you just get variations of the input image. This node hands you the number to turn.

How it works

Mechanically it's three steps. Your image goes through CLIP Vision (clip_vision.encode_image), the resulting embeddings are turned into a style conditioning by the style model (style_model.get_cond), and that style sequence is appended to your text conditioning. The interesting part is the strength_type choice:

• multiply (default) - the style tokens are literally multiplied by strength before being appended. 1.0 is neutral, lower weakens the style, above 1.0 pushes it. Range goes to 10 if you ever need to obliterate your prompt.
• attn_bias - instead of scaling the tokens, it adds log(strength) as an attention bias between the text and style token groups via an attention mask. This changes how strongly the text attends to the style tokens rather than the style's magnitude - a different feel at low strengths, worth trying once you've got the multiply behavior memorized.

The crop option (center/none) controls whether the CLIP Vision encoder center-crops the image first. Center is the safe default; use none if the important style content lives at the edges.

Inputs and outputs

• clip_vision - a CLIP Vision model (the same one the Redux setup uses).
• style_model - a STYLE_MODEL, i.e. a Redux-style adapter you've loaded.
• image - the reference image whose style you're borrowing.
• strength - 0 to 10, default 1.0.
• strength_type - multiply or attn_bias.
• crop - center or none.
• conditioning - your existing text conditioning; it comes back modified.

Output is a single CONDITIONING you feed to the sampler. Note there's no negative-side handling - apply it to whichever side you want, usually positive.

Install

It's part of ComfyUI-YarvixPA: ComfyUI Manager (search "YarvixPA") or

cd ComfyUI/custom_nodes
git clone https://github.com/YarvixPA/ComfyUI-YarvixPA

restart, and it's under ComfyUI-YarvixPA/Flux/Redux. The pack has no extra model downloads of its own, but you need a Redux-style model and its CLIP Vision on disk - those come from the model repos you'd use with Redux anyway.

Where people get burned

Redux-style strength is not linear-feeling. At 1.0 with multiply you may already be past the point where the prompt matters - that's the known behavior, not a bug. Start at 0.3–0.6 and raise until the style reads without the subject morphing into a copy of your reference. And don't expect this node to work on models that don't consume style conditioning - it's Flux Redux-family territory. On SD1.5/SDXL you want an IP-Adapter, which is a different animal entirely.