LC Natural Look 🍃

You know the look. Saturated to the point of candy, skin with an orange cast and no pores, a soft glossy bloom over everything that reads as "generated" even to people who can't say why. The standard advice for years has been to grade it out by hand: pull saturation down 5–10%, add a cast, add grain. LC Natural Look 🍃 is that checklist in one node, with a before/after wipe on its face.

The point is that it isn't a saturation slider

A flat desaturate makes a render dull, and dull is its own tell. What this node does instead, from its own source, is six targeted passes, all scaled by strength:

1. Only the loudest colours get pulled back. It measures each pixel's chroma spread (max minus min channel) and moves it toward its own luminance only where that spread is high - a gate that opens around 0.4 and hits full strength at 1.0. Skin, wood, foliage and anything already restrained stay put, which is why it reads as a camera rather than a filter.
2. Skin gets less orange. A soft YCbCr skin mask (blurred, so it doesn't fringe), then a push down on red-over-green and a light desaturate. skin multiplies this pass on top of strength - 2 is a double dose, 0 turns the correction off while everything else still runs, which is what you want when the skin was fixed upstream already.
3. Highlights roll off instead of clipping. Above a knee at 0.82 the curve bends through a tanh shoulder, so blown windows and speculars stop flattening into pure white.
4. The glossy glow comes down, fine texture goes up. It subtracts a wide blur (the large-scale "airbrush" component) and adds back a 1-pixel high-pass - pores, fabric weave, hair. This is the pass that actually kills the AI look, and it's the one a saturation slider can't imitate.
5. Your black point, contrast and brightness are handed back. It measures how far the 1% luma quantile lifted, lifts shadows back, pushes the mean 75% of the way toward the original and regains contrast up to a 20% cap. That's the difference between "graded" and "washed", per image, so a batch doesn't get one frame's grade applied to all of them.
6. A whisper of grain (0.004 × strength) from the seed - flat areas are their own tell.

Inputs and outputs

• image - the render.
• strength - default 0.5, and the author's own guidance is 0.3–0.6 for most renders. This is the dial you'll touch. At 0 it's a pass-through, so the node is safe to leave parked in a workflow.
• skin - default 1.0, range 0–2: how hard the orange/plastic correction hits.
• seed - seeds the grain only. It will not change your image.

image out goes to Save Image or the next node in your post chain. Because it previews on the canvas, it's also an output node - hover it and you get the wipe against the original.

Where it sits in a chain

LC123's own post order is: denoise or upscale → Skin Beauty → Sharpen Pro → Depth FX → colour → Bloom → Lens Profile → Film Stock → Film Grain last. Natural Look belongs in that colour slot, and it does two of those steps itself, so watch what you're doubling - with Vibrance or a LUT, back one off; LC Film Grain after it means two noise layers that read as JPEG. If LC Skin Beauty already fixed the skin, drop skin toward 0.

Install

ComfyUI Manager → search ComfyUI_LC123_nodes → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC123_nodes.git

Restart. The pack deliberately has no requirements.txt - it uses only what ComfyUI already ships, so there's no pip step. That matters in one situation: if you're loading a CivitAI workflow built on LC123 nodes on a hosted cloud ComfyUI, the platform has to install the pack, and you can't always do that yourself. Missing-node walls on rented runners are a known headache with this ecosystem; local install is where this pack is frictionless.

Things people get wrong

• Cranking strength to 1.0. 0.3–0.6 covers nearly everything; past that you're sanding off the render's intent and the shadow lift plus contrast regain starts to look like HDR.
• Expecting it to fix a bad generation. It's a deterministic pixel operation - no model runs, nothing is re-drawn. It makes a decent image photographically believable; it won't make a bad one good.
• Forgetting it costs CPU time. The math runs on the CPU: zero VRAM cost, a moment of wall-clock on a big batch or a 4K upscale.
• Wanting the wipe in Nodes 2.0. The hover compare is classic-canvas only.