Apply Style Model (Adjusted)

If you've run FLUX.1 Redux through ComfyUI's built-in Apply Style Model node, you know the problem: the reference image wins. Redux is BFL's official image-variation adapter, and it is aggressive - the community's standing advice is that anything above a whisper of style strength and "Redux will overwhelm the prompt and just produce a variation." The stock ApplyStyleModel node gives you zero control over that balance. It just slaps the style conditioning onto your prompt conditioning at full strength and hopes for the best.

Apply Style Model (Adjusted) is a drop-in replacement with one extra input: strength. Same inputs, same single CONDITIONING output, same wiring - but now you actually get a dial between "text prompt dominates" and "style reference dominates." If you've been fighting Redux for the ability to add new elements to a scene instead of just re-rolling the same image, this is the node you reach for.

How it works

Peek at the source and it's a ~15-line wrapper around the same style_model.get_cond() call the core node makes. The trick is what it does to the two halves before concatenating them:

• The prompt conditioning gets scaled by (3.0 - 2.0 * strength) - at strength = 0 your prompt is boosted up to 3×; at strength = 1.0 it's untouched.
• The style conditioning gets scaled by strength * 0.7 - so even at maximum, style only reaches 70% of what the stock node applies.

Both halves then get concatenated exactly like the core node does. That's the whole trick: it turns the style down while turning the prompt up to fill the gap, which is why the transitions feel smoother than just sliding one side.

The inputs that matter

Only four, and all required:

• conditioning - your text-prompt conditioning. In the included Flux workflows this usually comes after a FluxGuidance node.
• style_model - the loaded FLUX style model, from a StyleModelLoader pointed at flux1-redux-dev.safetensors.
• clip_vision_output - the CLIP Vision encoding of your reference image, from a CLIPVisionLoader (the siglip-so400m-patch14-384 model) feeding a CLIPVisionEncode. Wire this wrong or leave it unplugged and the node errors out immediately.
• strength - float, 0.0–1.0, default 1.0. This is the dial. Low = prompt wins, high = style wins.

The output is a single CONDITIONING that goes straight into your sampler's positive input.

One gotcha worth internalizing: the default of 1.0 is not "balanced." It's maximum style. The author's own README says very low values (under 0.2) can start hurting image quality because the prompt gets multiplied into weird territory. Treat 1.0 as "style max" and dial down from there - most people end up in the 0.3–0.7 band.

Install

No extra dependencies, no requirements.txt, nothing to download beyond what you already need for Redux. Either ComfyUI Manager (search "Apply Style Model Adjust") or the manual route:

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

Then restart ComfyUI. You'll find it under conditioning/style_model, right next to the node it replaces.

Reality check

Keep your expectations honest. This node only fixes the balance between prompt and style - it can't fix what Redux itself can't do. The input image still leans on the output heavily, and Redux's inherent weakness at adding genuinely new elements is a model limitation, not a settings problem. What this node genuinely fixes is the abrupt style-overwhelms-prompt cliff you get with the stock ApplyStyleModel, and it gives you a repeatable number to tune instead of fighting workflows. For style-transfer jobs where you want the reference's look without losing your prompt's instructions, it's the version I'd actually use.