chx_YC_LG_Redux

A normal Flux Redux setup in ComfyUI is a small chain: load a CLIP vision encoder, encode your reference image, load the Redux style model, apply it to your conditioning. Four nodes, easy to wire wrong, easy to lose track of which model file goes where. chx_YC_LG_Redux collapses that whole chain into one node - and then adds controls Flux's stock Redux path doesn't expose at all.

Redux is BFL's own image-variation/style adapter: it encodes a reference image with SigLIP and folds that embedding into your conditioning, which is why Flux workflows reach for it (and for PuLID or InstantX's IP-Adapter port) instead of the classic IP-Adapter - that line never got official Flux weights. This node sits in the pack's chx_IPA bucket, which tells you the author is treating Redux as this pack's answer to "the IP-Adapter slot" for Flux.

How it works

It SigLIP-encodes your reference image via clip_vision (default sigclip_vision_patch14_384.safetensors, matching BFL's own recipe), hands that to style_model (default flux1-redux-dev.safetensors), and patches the result into a CONDITIONING output you plug straight into your sampler as the positive. Redux is strong on style and vibe but - like most style adapters - happy to overpower your prompt if you push it too hard; this node's whole design is giving you separate levers to fight that instead of one blunt strength slider.

The inputs and outputs that matter

• style_strength (0-2, default 1) and prompt_strength (0-2, default 1) - two independent dials for "how much reference" vs. "how much text prompt," which is exactly the tension stock Redux forces you to fight with one knob.
• crop (center / mask_area / none, author's own tooltip) - crop the reference before encoding; mask_area pairs with the optional mask input to encode only part of your reference image.
• sharpen (-5 to 5, negative blurs, positive sharpens) and patch_res (1-64, larger = finer patches) - pre-processing on the reference before it hits SigLIP.
• blend_mode (lerp = faithful linear blend, feature_boost = pushes realism, frequency = pushes high-frequency detail - again, the author's own tooltip) and noise_level (0-1, adds noise, per the tooltip useful for "fixing erroneous details").
• Optional image, mask, and guidance (0-100, default 30 - Flux's usual guidance scale, exposed here too).
• Outputs: context (RUN_CONTEXT, pass-through) and positive (CONDITIONING) - wire the latter into your sampler.

How to install it

Base pack via ComfyUI Manager (search "ComfyUI-Apt_Preset") or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset.git

then run install.bat (Windows) or pip install -r requirements.txt (Linux/Mac), restart. This node specifically needs two model files that the pack does not ship: flux1-redux-dev.safetensors into ComfyUI/models/style_models/, and sigclip_vision_patch14_384.safetensors into ComfyUI/models/clip_vision/. Both are the standard BFL Redux files - grab them from black-forest-labs/FLUX.1-Redux-dev on HuggingFace if you don't already have them from a prior Redux workflow.

Common issues & troubleshooting

Dropdowns for style_model/clip_vision are empty. The schema reads your local models/style_models and models/clip_vision folders - an empty list means the files above aren't downloaded yet, not a bug.

Output looks like it ignores your prompt. Drop style_strength before touching guidance - that's the more direct fix for Redux overpowering text, and it's exactly why this node splits the dial out.

Result is noisy or muddy. Check noise_level is at 0 unless you deliberately want that "fix erroneous details" effect - it's easy to leave it bumped from a previous test.

IMPORT FAILED on the whole pack. ComfyUI-Apt_Preset is a big, fast-moving pack (real users have reported import failures on it after updates); reinstall via Manager and restart before assuming it's this specific node.