MiniMax H3 Temporal Detail Enhance / 时序保护细节增强 (Advanced)

Plain sharpening on video is a trap: it makes every frame crisper and then the motion stroboscopically flickers, because the sharpening fires on moving regions frame after frame. MiniMaxH3TemporalDetailEnhanceT8Advanced is a smarter sharpener - it gates its detail enhancement on motion, so it beefs up luma detail in quiet areas and backs off exactly where things are moving, which is where per-frame sharpening turns into shimmer. It also throws in optional bicubic upscaling on a 32-pixel grid, so it doubles as a mild size-up node.

What it actually does

It works on decoded frames (an IMAGE batch), post-decode - the description is explicit that it runs where Temporal Detail stays: after AV Decode, and it never touches audio. The mechanism is a luma detail pass with a motion gate: motion_threshold decides what counts as "moving," and temporal_guard (default 0.85) controls how hard the gate suppresses sharpening in those regions. The blur_radius/blur_sigma pair shapes the detail-extraction blur that separates "detail to enhance" from "noise to ignore."

The settings a beginner actually touches

• strength - 0.35 default, the overall amount of detail boost. Start here, nudge up if it's soft, back off the instant you see edge halos.
• upscale_factor - 1 by default (no upscale). Crank it and it bicubic-upscales on multiples of 32, which matters because H3's latents and decoders are grid-aligned.
• motion_threshold - 0.04 default. Lower = more regions treated as "moving" = more flicker protection but less sharpening overall.
• frame_chunk_size - 8 default. This is the memory governor: it processes frames in chunks with a one-frame temporal halo, so long clips don't blow up VRAM.
• maximum_output_megapixels - 2.1 default, and the tooltip calls it a "fail-closed output pixel budget": if the requested output exceeds it, the node refuses rather than OOMing. H3's official high-res reference is 1920×1088, which this tracks.

Outputs:

frames (the enhanced batch - same shape in, same out) and report_json. Drop it between AV Decode and your save/preview, and it slots into a workflow with zero other changes.

Installing:

part of MiniMax H3 Audio T8. Manager → "MiniMax H3 Audio T8", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart. No mandatory pip deps.

The honest limits

This is a sharpen, not a repair: it "cannot reconstruct missing faces or geometry," so a mushy face stays a mushy face - you want a face-refine pass for that. It's also motion-gated, not motion-aware in any semantic sense; heavy camera shake may still gate awkwardly. And it's experimental in the pack's careful way: default modest, fail-closed on the pixel budget, and the author isn't claiming it's a universal quality win.