沐阳 H3 · 画质与分镜增强（内部）

Long AI video has two failure modes that show up in the same place - the face. Across a chain of segment generations a character's identity drifts, and on a second pass at higher resolution the face is the first thing to go soft, because the low-denoise redraw has the least signal to work with exactly where the detail matters most. The wider ecosystem's answer to this has always been an extra pass: detect the face, crop, redraw at higher resolution, stitch it back. ADetailer's descendants, basically.

H3DirectorEnhancementSettings is where you configure that for H3 - and, more practically, where you find out whether you're set up to do it at all.

What it is

A settings bundle. It has no pixels in it. It gathers ten widgets into a MYANG_H3_ENHANCE object that the Director reads when it expands a segmented job. Two blocks of options ride on it:

Face refine. face_enabled (off), face_detector (a path - default bbox\face_yolov8m.pt, note the Windows-style backslash), face_steps (4), face_denoise (0.45), face_crop_factor (2.5), and face_identity_ordinal (0-9). The denoise at 0.45 is the sweet spot range for this kind of pass: enough to re-synthesise detail, not enough to give you a different person. face_crop_factor 2.5 gives the detector a generous margin around the face, which is what keeps the seam of the crop from landing on a jawline.

Motion repair. motion_enabled (off), motion_preset (balanced (default) / max quality (wide plateau) / economy (tight spans)), motion_steps (6), motion_inject (0.7). The inject value is how strongly the recovered motion gets pushed back into the sample; 0.7 is a middle setting, and pushing it to 1.0 tends to fight the base generation rather than help it.

There's also an optional model socket for scheduling order, same as the pack's other settings nodes.

The catch, and it's a big one

Neither block works out of the box, because the actual processing lives in other people's packs. Face refine needs AIMixer's ComfyUI-H3-FaceRefine (H3FaceTrackCrop, H3InjectVideoLatent, H3PerFrameDenoise, H3FaceStitch) plus a face detector model on disk. Motion repair needs AIMixer's ComfyUI-MAINodes (H3ContactSheet, H3JerkOracle, H3TimeSmear, H3InjectSchedule, H3ExactRecover, H3AudioRecover) plus the five-view LoRA.

That's a deliberate design choice by the author, and a good one: with both switches off, the core Director, native continuity, second pass and audio chain never import those packs at all. No dependency, no import error, no half-installed mess. The flip side is that turning a switch on without the pack installed gets you a clear failure rather than a graceful degradation - the long-video node checks for the nodes and the detector before it starts sampling, which is the right place to fail: before you've burned twenty minutes of GPU time.

Outputs and wiring

One output, 增强设置, carrying everything above. Wire it into the Director's enhancement input. If you don't own the helper packs, leave the whole node out of the graph - or keep it with both switches off, which is a perfectly valid no-op configuration.

Category is 沐阳 H3/导演台/内部, so this is one of the nodes the Director normally manages for you, not something you build a graph around.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang

Restart and hard-refresh. The pack itself declares no Python dependencies and needs no model weights of its own - but note the README's framing: weights, LoRAs and demo media are deliberately not distributed. If you want the face-refine path you'll be downloading the face detector and the FaceRefine pack separately, and the pack's regression tests report SKIP for those paths when the helper packs aren't present, which is worth knowing before you treat a green test run as proof your setup works.

Also true of this whole pipeline: a clean seam doesn't mean content quality stays constant down a long chain. Detail loss and identity drift accumulate. A face pass buys you margin; it doesn't make the problem go away. For the underlying mechanics, see identity preservation.