MiniMax H3 Motion Quality Audit / 高速动态质量审计 (Advanced)

Sometimes you don't need to know why H3's motion went wrong - you just need to know where, so you can rerender or repair it. MiniMaxH3MotionQualityAuditT8Advanced is a temporal proxy audit: feed it decoded frames and it scores them for the failure modes that actually plague fast H3 output - sharpness drops, temporal instability, high-motion smearing, and frozen frames - then hands you the risky ranges as a report.

Read the description carefully, because the author is meticulous about scope. It's read-only and has no external dependencies - no SAM, no face model, no ML model to download. It doesn't claim face identity detection, and it never changes, uploads, or auto-accepts media. If you want face-focused analysis, you supply a reviewed face/subject mask yourself, or a manual ROI. This is a proxy audit, not a judgment; its output is designed to feed a repair plan, not to be a final verdict.

Inputs that matter

• frames and fps - what you're auditing. fps defaults to 24.
• roi_mode - full_frame (default), manual_static_roi, or connected_mask. For faces, use a mask; the roi_x/y/width/height fields only apply to manual ROI mode.
• sharpness_ratio_floor (0.55) - below this, the region is considered soft.
• temporal_instability_multiplier (2.5) - how jumpy counts as jumpy.
• high_motion_delta_floor (0.03) and freeze_delta_ceiling (0.002) - the smearing and frozen-frame gates.
• repair_context_frames (12) - how much context around a flagged range gets included when the result becomes a repair plan.
• face_mask (optional) - reviewed face/subject mask for face-focused analysis.

Outputs

risk_detected (boolean), risk_range_count, risk_ranges_json (the list of flagged ranges), and report_json. It's an output node, so it can terminate a graph and hand its findings to a human (or to the Repair Plan node).

Install and where it fits

Pack install, once: ComfyUI Manager → search MiniMax H3 Audio T8, or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into ComfyUI/custom_nodes/, restart. No pip extras. This one is refreshingly dependency-free - no models to place, which makes it the cheapest node in the whole Motion Recovery family to try.

Use it as your "was that footage good or not?" gate before spending compute on repair. Wire risk_ranges_json into MiniMaxH3MotionRepairPlanT8Advanced and it becomes a non-destructive repair plan; or just read it and rerender manually. The honest caveat: it's a proxy - "risk detected" means the numbers look bad, and a face-specific audit needs a mask you actually reviewed. And like everything in this pack's quality corner, it's Experimental and fails closed by design, so trust its "no risk" verdict less than you trust its "risk" verdict.