Denoise

The node description calls this "ultimate professional-grade 32-bit float spatial-temporal denoising" and that's the marketing voice of the pack's author, not a spec sheet. Ignore the adjectives and look at what it actually does, because underneath the puffery this is a genuinely useful denoiser built the way a compositing denoiser is built: luma and chroma handled separately, three frequency bands, and optional temporal blending for sequences.

The mechanism

Noise almost never lives in the same place as detail. Grain is high-frequency and mostly luma; the ugly stuff - the chroma speckle in shadows after a low-light render - is what your eye actually notices. So the node splits luma from chroma and lets you set independent strengths (luma_strength, chroma_strength), then decomposes each into three frequency bands with separately weighted strengths: high_freq_denoise for fine grain, mid_freq_denoise for textures and compression artifacts, low_freq_denoise (default 0.5) for those big coarse gradient splotches that come out of an aggressive upscale.

That's a sensible architecture. "Denoise" as one number is what every other node in the ecosystem gives you, and it's why people end up with either visible noise or plastic faces.

Two extras worth knowing:

joint_chroma_guidance (on by default) uses the sharper luma channel's structural edges to guide chroma smoothing, which is how you avoid colour bleeding across a high-contrast edge - the classic magenta fringe around a silhouetted face.

filter_type picks the spatial core: Bilateral preserves edges, Guided runs faster, keeps finer detail and avoids halos. If you're on a deadline, Guided; if you're protecting skin texture, Bilateral.

For a sequence, temporal_blend (0 = off) blends neighbouring frames to kill the frame-to-frame fizz you get from per-frame denoising, with temporal_radius setting how far it searches. temporal_threshold is the one that saves you - it limits blending to static or slow-moving pixels, so fast motion doesn't smear into ghosting.

Inputs you'll actually touch

For a single still: d (base filter radius, 9), sigmaColor (0.15 - colour similarity, higher is smoother and softer), and luma_strength down around 0.2 if you want to keep natural grain.

Turn on auto_profiling if you've got a wildly noisy image and no patience: instead of guessing sigmaColor, it measures the noise floor (lowest luma standard deviation across 32px patches) and scales from that. profile_multiplier is then your only dial. For HDR input, hdr_auto_sigma (on by default) widens the colour threshold proportionally to the batch peak, so values above 1.0 don't get treated like 0–1 pixels.

And view_mode - set it to Noise Residual once. It shows you exactly what you're removing, and it will change your settings faster than any explanation.

Install

Part of Radiance (147 nodes). ComfyUI Manager → search Radiance → install, restart, refresh browser. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/fxtd-studios/radiance.git
cd radiance
python -m pip install -r requirements.txt

On Windows portable, pip with python_embeded\python.exe. The filter cores are OpenCV and SciPy, both in requirements.txt. Nothing downloads for this node.

Where people get burned

The filters are CPU-side and the bands make it slow. Three-band decomposition with per-band filtering on a 4K clip is real work. Denoise after a downscale, or accept the wait. The pack's own release notes list the heavy nodes; this one isn't in the "broke at production size" list, but it's not free either.

motion_compensation is narrower than it sounds. The author's own tooltip says it: it shifts each neighbour by the best of nine offsets within one pixel. That helps with sub-pixel drift and does nothing for actual motion. Don't expect it to rescue a tracking-heavy sequence - the known-issues list flags it explicitly as "a ±1 px search".

detail_recovery is not a blend back to the original. It restores the high-frequency part of what was removed, so texture comes back - and so does some fine noise. It's the right tool when the denoise ate skin pores; it won't give you the input image back.