Stack_Redux

Flux Redux has a well-known problem: run it with the default stock ComfyUI workflow and you basically just get variations of your input image. One community member testing it early on put it exactly right - "running the default workflow in Comfy felt like I was just getting variations of the same image" - and had to build a custom setup just to get the strength dial fine enough to be useful, landing on something like 0.04-0.07 out of a 0-1 range before Redux started drowning out the text prompt entirely. Stack_Redux is Apt_Preset's answer to that: it wraps the same style-model + CLIP-vision mechanism but exposes a lot more surface area for actually controlling the balance between "look like the reference" and "listen to the prompt."

How it works. Redux takes a reference image, runs it through a CLIP vision encoder, and feeds that as style conditioning alongside your text prompt - no LoRA training, no fine-tune, just image-prompt conditioning bolted onto Flux. This node's style_strength and prompt_strength (both 0-2, default 1) are the direct fix for the "overwhelms the prompt" problem above: turn style_strength down and prompt_strength up if you're getting a near-copy of the reference instead of something that actually follows your text. blend_mode picks the algorithm doing that mixing - lerp is a straight linear blend that stays closest to the reference, feature_boost pushes toward more realistic detail, and frequency specifically emphasizes high-frequency texture over broad composition. crop controls what part of the reference actually gets used: center crops to the middle, mask_area restricts the style source to wherever your mask input is set (so you can pull style from just a piece of clothing or a background element instead of the whole image), and none uses it uncropped.

The finer controls are patch_res (how finely the reference gets divided into patches before encoding - higher means finer, more localized style transfer), sharpen (negative blurs the reference before use, positive sharpens it, 0 leaves it alone - useful if your reference photo is a bit soft and you want cleaner style signal), and noise_level, which the tooltip describes as a repair tool: adding a bit of random noise can smooth over erroneous details the style transfer introduced. guidance is Flux's usual guidance scale, exposed here at 0-100 (default 30).

Chaining and output. redux_stack as an optional input lets you feed the output of one Stack_Redux into another - layering multiple reference images with different settings before it all hits the sampler. The output is that same redux_stack type, consumed downstream.

Installing it. No special dependency for this node beyond the base pack - git clone https://github.com/cardenluo/ComfyUI-Apt_Preset into custom_nodes, restart, or search the pack title in ComfyUI Manager. What you do need is the actual Redux model files, same ones any stock Redux setup uses: flux1-redux-dev.safetensors in models/style_models, and sigclip_vision_patch14_384.safetensors in models/clip_vision. Neither ships with the pack - grab them from Black Forest Labs' HuggingFace repo if you don't already have a Flux setup with Redux working.

Troubleshooting. If your output looks like a near-identical copy of the reference image regardless of your prompt, that's the classic Redux failure mode described above - bring style_strength down and prompt_strength up, and try feature_boost or frequency instead of lerp if you want more of your prompt's own detail to show through. If style transfer is bleeding in stuff you didn't want from the reference (a background, an outfit you weren't targeting), set crop to mask_area and wire in a mask covering just the part you want. And if the output looks artifact-y or oddly textured, try backing sharpen toward zero before reaching for noise_level - sharpening a soft reference too aggressively before encoding is a more common cause than the model itself misbehaving.