YC_LG_Redux_Advance

Stock Flux Redux is a blunt instrument: reference image in, style out, and if it's overapplying you're mostly stuck swapping models. YC_LG_Redux (shown as YC_LG_Redux_Advance in the node menu) is the community attempt to give that pipeline a full toolbelt - sharpen the reference before it's encoded, crop it, tune how hard the style and the prompt each push, and mix them in three different ways. It's one of the two Redux-with-controls nodes in Comfyui_Redux_Advanced, and it's the one the README actually documents.

What it is

A replacement for the core StyleModelApply inside a Flux Redux workflow. Where the stock node takes a ready-encoded CLIP vision output, this one takes the raw clip_vision loader output and the image directly, does its own encoding, and gives you a pile of pre- and post-processing knobs on top. Output is one CONDITIONING that feeds your Flux sampler exactly like a stock Redux apply would.

How it works

Reading the source: first it pre-processes the reference image. Positive sharpen runs an OpenCV convolution kernel; negative values swap to a Gaussian blur instead, so you can actually scrub the reference clean before it's encoded. Then crop decides what to feed the vision model - center crops to the middle square, mask_area crops to the bounding box of your optional mask, none passes the image through. The image goes through clip_vision.encode_image(), then style_model.get_cond() produces the standard feature grid.

The interesting part: that grid starts at 27×27, and patch_res bilinearly upscales it - bigger values mean a finer-grained conditioning that keeps more spatial detail, at the cost of more tokens and VRAM. Then style_strength blends the style features against your prompt's text features, and blend_mode picks the recipe:

• lerp - linear mix, sticks close to the reference image.
• feature_boost - a tanh-normalized boost for extra realism.
• frequency - an FFT pass that lifts high-frequency detail.

prompt_strength scales the text side (it applies a cubed multiplier and doubles the sequence), and noise_level injects normalized random noise - the author's stated use is "repairing wrong details" by giving the sampler a bit of variance.

Inputs that matter

• conditioning - your encoded positive prompt.
• style_model - flux1-redux-dev from StyleModelLoader.
• clip_vision - the CLIPVisionLoader output (sigclip_vision_patch14_384), not the encoded output.
• image - the reference.
• style_strength / prompt_strength (0–2) - the two main dials.
• blend_mode - the three flavors above; lerp is the safe default.
• sharpen, patch_res, noise_level - the fiddly ones. Most people can leave these alone initially.
• Optional mask - enables crop = mask_area and, when set during sampling, blends text features back in outside the masked region.

Install

Same pack as its sibling - ComfyUI Manager → search "Comfyui_Redux_Advanced", or:

cd ComfyUI/custom_nodes
git clone https://github.com/yichengup/Comfyui_Redux_Advanced

then restart. No extra Python dependencies; it needs cv2, which ComfyUI already ships. Models are manual: the Redux style model into models/style_models/, the SigLIP vision into models/clip_vision/.

Common issues

The standard Redux failure mode applies: if you feed it a non-SigLIP CLIP vision you'll get mat-shape errors during encoding. And remember the difference from StyleAdvancedApply - this node wants the loader output, not a CLIPVisionEncode result. The README also shows Redux working alongside an openpose ControlNet, which is a nice combo if you want to control pose while borrowing style. The docs are Chinese-only, so if you're not a reader, the bundled example image (example/YC_LG_REDUX.png) is worth dragging into ComfyUI to see the wiring at a glance.