MiniMax H3 Face Refine Plan / 远景脸修复规划 (Advanced)

Every H3 user eventually hits it: the wide shot where your character's face is 60 pixels of mush. H3 is a joint AV transformer, so you can't just inpaint a face like it's SDXL - but you can run a second, low-denoise H3 pass on just the face region. That's the whole face-refine pipeline this pack builds, and MiniMaxH3FaceRefinePlanT8Advanced is where it starts. It's the planning node: it looks at your source video, finds and tracks faces, spots scene cuts, and produces the cropped face strips plus a plan that every downstream face-refine node consumes.

It's shot-aware, which matters more than it sounds. Real video cuts - a new camera angle, a hard cut - mean the face tracking should reset, not try to keep one wobbly track alive through a cut. The node detects cuts via a low-resolution RGB mean absolute difference (scene_cut_threshold, default 0.28) and resets its smoothing at every one. max_track_jump (0.18) and max_gap_frames (4) govern when a face is the same face across frames versus a new detection.

The inputs you'll actually touch:

• frames - the exact source IMAGE batch at the stated fps. Feed it the real frames, not a downscaled copy; the plan is source-bound.
• detector_mode - default local_opencv_yunet: OpenCV's YuNet face detector, MIT-licensed, running on CPU. There's an anime ONNX route (EXP), a manual_static_roi fallback, and a local Ultralytics option. Nothing downloads a model at execution time.
• confidence (0.35) - detection threshold; raise it if you're catching false faces, lower it if real faces are missed.
• crop_context_scale (3) - how much context around the face each crop keeps. More context = more stable, but bigger latents.
• canvas_size (default auto_512) and require_h3_grid (true) - crops get padded to H3's grid constraints, which is non-negotiable for the model to accept them.

The manual ROI fields (manual_roi_x/y/width/height) are there if detection is hopeless - a fixed rectangle beats a detector that keeps locking onto the wrong face. Outputs: face_plan (the H3_T8_FACE_REFINE_PLAN everything downstream consumes), crops, preview, report_json, plus canvas_width, canvas_height and frame_count for building your layout.

Two things to know before you build around it. First, it's deliberately conservative: it never downloads models and it never auto-accepts a candidate - the plan is a proposal you wire into the rest of the pipeline. Second, there's a real-world quirk in the description: the node destroys its detector object after planning, but OpenCV may retain process-global CPU allocator pages. That's not a leak in your workflow, it's a documented behavior - if your RAM usage looks sticky after a plan run, that's why.

Install: ComfyUI Manager → "MiniMax H3 Audio T8", or clone:

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

Restart after. No forced pip deps, no model downloads - you'll need the H3 model, CLIP, VAE and LoRAs yourself for the actual refine pass, and YuNet comes from your local OpenCV. The plan node is CPU-only and cheap; the expensive part is the second H3 pass you're planning for.