MiniMax H3 Skin Finish Frequency Split / 肤色纹理解耦 (Advanced EXP)

The classic complaint about AI skin isn't that it's too smooth - it's that smoothing flattens the detail that makes skin read as real. Too-smooth skin is the tell that's survived every generation of models. The Frequency Split node is T8's answer to that specific failure: instead of smearing a skin-finish candidate over the whole face, it rebuilds the result from two layers - the candidate's low-frequency tone correction, plus the source frame's existing high-frequency detail. Smooth the color, keep the pores.

How it works

It's deliberately non-generative, and the node description says so plainly: it cannot deblur, cannot create pores, cannot repair identity, and cannot decide a result is better. The math is the boring, reliable kind:

• A low-pass separation splits the candidate and source into tone/brightness (low-frequency) and detail (high-frequency) layers.
• low_frequency_strength (default 1.0) blends the candidate's low-frequency correction in; 0 keeps the source's tone entirely.
• source_detail_gain (default 1.0) controls how much of the source's existing high-frequency detail comes back. 1.0 is neutral; it doesn't invent detail that isn't there.
• separation_radius_percent (default 1.0) is the low-pass radius as a percentage of the shorter image side, so the same node behaves sensibly across resolutions - capped by maximum_radius_px (default 32) as a CPU-cost guard.

Then a bundle of conservative gates decides whether each frame is allowed to change at all: minimum/maximum mask area, a maximum_new_clipped_fraction that bounces frames back to source if too many pixels clip, and chunk_frames (default 4) for CPU batching.

The input/output shape

Required: source_frames (IMAGE), candidate_frames (IMAGE - the skin-finish result you're trying to rescue), and used_skin_mask (MASK - semantic skin, e.g. from the pack's ParseNet semantic-mask nodes). Optional audio passes through untouched.

Outputs: frequency_split_candidate (the recombined result), source and selected (the exact source, because by default source stays selected), audio, effective_mask, rejected_mask (which frames/regions the gates refused), difference, and report_json. The accept_candidate toggle (default false) is the entire point of this architecture: nothing is applied until a human flips it after review.

Why you'd reach for it

If you've got a skin-finish candidate that evened out the complexion but wiped out the texture - the "influencer default" look the photorealism lore keeps warning about - this node is the texture rescue. It's also a good citizen in a longer chain: it only touches pixels inside the mask, preserves everything outside it and all auxiliary channels, and keeps audio bound to the same object.

Installing it and gotchas

Part of the comfyui-minimax-h3-audio-T8 pack: ComfyUI Manager search "MiniMax H3 Audio T8", or clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git into custom_nodes, update ComfyUI core first, restart. No extra model files - it's pure CPU image math.

Where people get burned: treating it as a beauty filter and cranking low_frequency_strength expecting more smoothing. It can't exceed what the candidate gives it, and over-aggressive settings trip its own clipping gates, silently sending frames back to source. And remember the contract: accept_candidate=false means the selected output is bit-exact source no matter how good the candidate looks. That's the pack's whole ethos - review first, ship second.