LatentDetailer

You know the complaint: SDXL output looks clean, smooth, plastic. The classic fixes all cost something - cranking CFG gets you saturation and burn, a full hi-res fix is a second generation, and a generative upscaler like SeedVR2 or SUPIR rewrites your faces along the way. LatentDetailer is the cheap middle option: a detail pass that runs entirely in latent space, right before the VAE decode, at the price of one extra UNet evaluation.

Its workflow position tells you what it does: wire it between your KSampler's LATENT output and the VAE Decode, and it adjusts the latent before it ever becomes pixels. It's a "detailer" in the Impact-Pack sense of the word, but where FaceDetailer detects a region and re-generates it at higher resolution, this one never detects anything and never re-generates.

How it works

The node runs your UNet again on the finished latent - once per entry in sigmas - at a chosen point in the denoise schedule, and computes what the model thinks the clean image would be there. Same denoised-estimate math your sampler uses every step, just a one-off probe instead of a full walk.

sigmas is a comma-delimited string of normalized positions from 0.0 (start of denoising, noisiest) to 1.0 (end, cleanest). Each entry triggers one UNet evaluation and the estimates are averaged, so 0.40 (the default) means one probe, and 0.25,0.55 means two. Then the interesting part: the difference between that clean estimate and your latent is split into low and high frequency, and the high-frequency part is reinjected as detail - scaled by detail_strength, tamed in flat areas by protect_lows, separated by hf_radius.

There's also a CFG-delta branch, a nice piece of reuse: it amplifies den_pos - den_neg, the same signal classifier-free guidance is built on. cfg_hf_boost pushes that high-frequency guidance into the result. The node's own cfg input is just the base for that shaping - set it to match your sampler's CFG. On a distilled model at CFG 1, this branch quietly shuts off, which is expected, not a bug.

On top of that come the finishing touches, all optional: noise_scale/noise_radius for photographic micrograin, bloom_strength/bloom_threshold for halation, color_drift for granular color variation, and chromatic_aberration for lens-like fringing. Defaults are subtle by design - the author aimed at "real photography," not a dramatic filter.

The inputs that matter

You won't set forty knobs on your first run. Start here:

• model, latent, positive, negative - you wire your checkpoint and conditioning in again. It's a MODEL node that runs the UNet, not a free post-filter, so it needs the same ingredients as your KSampler.
• sigmas - leave at 0.40 until you understand the effect.
• detail_strength - the master detail dial (default 0.65). This is the one you'll actually turn.
• hires_scale / hires_strength - a cheap built-in hi-res trick: if >1, one UNet pass at the upscaled latent, then downsampled and blended back. Not a substitute for a real hi-res fix, but a way to feed the detail pass more resolution.
• noise_scale - micrograin; the author pairs a little bloom with it for convincing rolloff.

The output is a single LATENT, fed straight into your VAE Decode (or another KSampler, if you're feeling ambitious).

Install

ComfyUI Manager, search "LatentDetailer" (it's on the Comfy Registry), or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/SparknightLLC/ComfyUI-SpectralVAEDetailer

Restart ComfyUI. That's it - no requirements.txt, no model downloads, no heavy dependencies; the only cost is what those extra UNet evaluations use.

Gotchas

• Images turn "splotchy"? The author hit this too and shipped a preset_subtle example workflow. If the effect is too strong, load that preset instead of fighting the sliders.
• Noisy flat backgrounds were the original worst case - that's what the cfg_radius_adaptive mode (with cfg_radius_flat and cfg_adapt_feather) was built to fix. It uses a cheap detail mask so flat regions get gentler CFG-detail treatment.
• It's subtle. You're micro-adjusting a latent, not running a restoration model. Compare side by side or you'll think it did nothing.
• Speed scales with sigmas and hires_scale - one sigma at 1x is fast, three sigmas at 2x is a second mini-sampler.

It's tuned for SDXL photoreal workflows, though the author says it's model-agnostic - ZIT and friends work, just budget time to re-tune the starting parameters. For killing the AI-look on portraits without a face-rewriting upscale pass, it earns its slot.

---

The alias, and a name change

If you also see SpectralVAEDetailer in your node menu, that's this same node under its old name - the pack renamed itself from ComfyUI-SpectralVAEDetailer to ComfyUI-LatentDetailer in v0.3.0 and kept the old name registered as an alias. Both map to the same class, same inputs, same output; use either. See the SpectralVAEDetailer page for the full rename story.