☁️BizyAir Apply Redux model (advanced)

Redux is Black Forest Labs' own answer to "make this image look like that reference," built directly into Flux Tools rather than bolted on by the community the way most IP-Adapter-style tools are for other architectures. It reads a reference image through SigLIP encoding and feeds that as style/content guidance into your conditioning - good for image variation and carrying a look (palette, texture, lighting mood, composition feel) from one image into a new generation without literally copying it.

This "advanced" version is the one with the actual controls exposed, as opposed to a simpler apply node that just takes an image and a weight. If you've used Redux at all, this is where you go when the default behavior isn't giving you what you want.

The inputs that matter

• conditioning, style_model, clip_vision - the pieces Redux needs: your existing prompt conditioning to merge into, the Redux weights, and the CLIP vision encoder that actually reads the reference image. All required, all upstream loader outputs.
• image - the style reference itself.
• weight (0–1, default 1) - how strongly the reference influences the result. This is the first dial to reach for if the style is either barely showing up or completely dominating your prompt.
• downsampling_factor (1–9, default 3) and downsampling_function (nearest, bilinear, bicubic, area [default], nearest-exact) - together these control how much the reference image gets compressed before it's injected. A higher downsampling factor pools more aggressively, which loosens the match - you get the general feel of the reference without it dictating fine detail. Lower values keep more of the reference's specific structure. area pooling is the default and a reasonable starting point; the others are worth trying if area gives you artifacts on a particular image.
• mode - center crop (square) (default), keep aspect ratio, or autocrop with mask. This decides how the reference gets framed before encoding. Center crop is simplest and works fine for most references; autocrop with mask is for when you only want a specific region of the reference image treated as the style source.
• mask (optional) and autocrop_margin (optional, 0–1, default 0.1) - only meaningful when mode is autocrop with mask. The mask defines the region of interest, and the margin pads around it before cropping.

Outputs: BIZYAIR_CONDITIONING (your merged conditioning, ready for the sampler), plus IMAGE and MASK - likely the cropped reference region actually used, useful for confirming the autocrop did what you expected rather than guessing.

Installing it

ComfyUI Manager: search BizyAir, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir.git

Set your BizyAir API key on first use.

Where people get burned

The most common disappointment is treating weight as the only knob when a result looks off - if the style reference is dominating in a way that flattens detail everywhere in the image, try lowering the downsampling factor's compression less (i.e., raising the factor) before just cranking weight down, since those two settings interact rather than acting independently. And if you're using autocrop with mask and the result doesn't look like it's using the region you expected, check the MASK output first - it'll show you exactly what got cropped, which is a faster diagnosis than re-running with different settings blind.