LC Detail Daemon (BETA)

Detail Daemon is one of the better cheap tricks in this hobby: you lie to the model about how noisy the latent is, it removes slightly less noise than it planned, and the fine texture that usually gets smoothed away survives. It showed up as a ComfyUI node in late 2024, got a real following (a HiDream comparison post hit +211), and then mostly settled into being a checkbox people forget they left on. LC's version keeps the original's knobs - so your old numbers carry over - and drops one thing that annoyed people.

What it actually does

The original Detail Daemon, and this node, follow the same recipe. Inside a window of the sampling steps, the sigma handed to the model is scaled down by a small amount. amount 0.25 means the model is told there's 2.5% less noise than there really is. It then denoises like the image is already cleaner, so it holds onto pores, fabric weave and hair instead of sanding them flat. Negative amounts go the other way and smooth.

The window is set by start (0 = first step, 1 = last, default 0.15) and end (default 0.85 - leave the final steps alone or you're just leaving noise in the picture). peak says where in that window the effect is strongest: 0 puts it right at the start on the noisy steps, which is where it shapes folds and fabric. exponent bends the ramp - 0 is flat, full strength across the whole window, which is what the author recommends.

Three modes, and this is the part worth understanding:

• look only (default) - the model is told less noise, but its answer is rescaled back to the real sigma. You get crispness without the composition reshuffling. This is the mode you want on modern models.
• classic - the original behaviour. More detail, but small things move. Worth knowing that this is real: in the HiDream thread people noticed dresses gaining floral prints that were never prompted. That's the technique working and also a prompt-adherence bill.
• keep structure - classic, then only the fine part of the change is kept so the layout stays. Costs two model calls per step inside the window, roughly 45% slower.

One difference from the original is buried in the tooltip and matters now that half the good models run at CFG 1: the change is not multiplied by CFG. Same amount, same effect, whether you're at 1 or 7.

Inputs and outputs

Wire sampler - from KSamplerSelect, from LC Sigmas, or any SAMPLER including ClownSampler if you run RES4LYF - into this node, then this node's SAMPLER output into SamplerCustomAdvanced. The optional sigmas input does nothing to the result; it's there so the node can draw where your steps land.

That's the whole surface. The slider you'll actually touch is amount. Defaults are tastefully set for Krea 2 / Krealism (look only, 1.0, start 0.15, end 0.85, peak 0, exponent 0). Two pieces of advice from the author's own tooltip are worth repeating verbatim:

• At 16+ steps, use less.
• Samplers that add their own noise - dpmpp_2m_sde, euler_ancestral, er_sde, any res_* or ClownSampler with eta - can leave pepper specks. Lower amount or pick a non-SDE sampler.
• On 4-step distilled models there's not much for the daemon to bite on. That's not a bug; the window is tiny.

Install

It's part of the LC123 pack, which is one of the cleaner installs in ComfyUI: no requirements.txt, it only needs the torch and numpy ComfyUI already ships.

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

Or search ComfyUI LC123 Nodes in ComfyUI Manager. Restart and watch for [LC123] + lines per module in the console. If you see nothing, __init__.py isn't sitting directly inside ComfyUI/custom_nodes/ComfyUI_LC123_nodes/ - zips sometimes unpack one folder too deep, and that's the usual cause.

Where people get burned

The most common complaint about Detail Daemon in general is that it "does nothing" - and usually it's a 4-step turbo model, an amount left at a flat 1.0 on a sampler that already injects noise, or an end value of 1.0 shuffling late-step noise back in. The second most common is that it changed something you liked: drop amount and switch from classic to look only before you blame the model.

If you're on a plain KSampler rather than SamplerCustomAdvanced, this node can't help you - the SAMPLER output only plugs into custom-sampling nodes. Use LC Detail Daemon (model) instead; same knobs, different socket.