StyleModelApplySimple

Flux Redux has a well-known problem: give it a reference image and a prompt, and the prompt gets steamrolled. There's no strength setting anywhere in BFL's stock workflow to dial that back. StyleModelApplySimple is kaibioinfo's low-effort fix - a drop-in swap for ComfyUI's built-in StyleModelApply node that adds exactly one control: a strength dropdown. If you just want Redux to stop ignoring your prompt without learning what "downsampling factor" means, this is the node.

What it's doing under the hood

Redux crops your conditioning image to a square, resizes it to 384×384, and CLIP-vision-encodes it into 729 tokens that get tacked onto the end of your T5 prompt. Because that's a lot more tokens than most prompts, the image wins by sheer volume. This pack's sibling node, ReduxAdvanced, fixes that by exposing a raw downsampling_factor (1 to 9) that shrinks the token count before it reaches your prompt. This node is the same trick with the knob pre-labeled: the image_strength dropdown maps onto that same scale, so you're picking a point on it by name instead of a number.

The one setting you actually touch

Beyond the standard conditioning, style_model, and clip_vision_output inputs, there's a single parameter: image_strength, a five-way choice of highest, high, medium, low, lowest (default medium). medium corresponds to a downsampling factor of 3 - it's what every example image in the pack's own README was generated with, so it's a genuinely sane default, not just a placeholder. Push toward highest and you get closer to raw, undownsampled Redux (image dominates); push toward lowest and the image's pull weakens further, leaving more room for your prompt.

One wiring detail worth knowing up front: this node wants clip_vision_output, not a raw image. That means you need a CLIP Vision Encode node upstream doing the encoding first, then feeding its output in here - unlike ReduxAdvanced, which takes the raw image and clip vision model directly and does that encoding step itself. If you're used to ReduxAdvanced's wiring, this is the one difference that'll trip you up.

The single output is CONDITIONING - wire it straight into your KSampler's positive conditioning input, exactly where you'd plug in a normal text-prompt conditioning.

Installing it

It ships in the same repo as ReduxAdvanced, so one install gets you both nodes. In ComfyUI Manager, search "Advanced Reflux control." Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/kaibioinfo/ComfyUI_AdvancedRefluxControl

Restart ComfyUI afterward. No extra Python dependencies to chase down - it's built entirely on ComfyUI's stock conditioning and CLIP-vision types, so you won't hit the usual custom-node dependency-conflict headache. You do need the standard Flux Redux prerequisites already in your workflow: a Redux style-model checkpoint and a CLIP vision model, loaded through ComfyUI's normal loader nodes.

Where people get stuck

If your output still reads as a straight Redux variation and your prompt seems invisible, check two things: that you're actually using this node and not the stock StyleModelApply, and that image_strength isn't sitting on highest. Drop it to medium first - that's the setting behind every example in the README.

A type-mismatch error on the clip_vision_output input almost always means you wired a raw IMAGE in directly instead of routing it through a CLIP Vision Encode node first - this node expects the already-encoded output, not the picture itself.

And if you need masking, or your reference photo is badly non-square and center-cropping is cutting off the part you care about, this node simply can't do either - those are ReduxAdvanced-only features. That's the right moment to graduate to the advanced node rather than fighting this one; the dropdown here is deliberately limited to keep the simple case simple.