Face Performance Evaluate

Every generated facial performance eventually needs a QC pass, and doing that by eye across hundreds of frames is a slow way to go blind. MKRFacePerformanceEvaluate is the MKRShift node that scores a clip numerically instead: lip/audio sync lag, blink rate, landmark smoothness, and pose jitter - then flags anything that breaks a threshold. It's the regression-test and QA node at the end of the face-performance pipeline.

If you've ever changed one retarget setting, re-rendered, and squinted at the mouth to decide if it's "better," this node exists to replace that guess with a number. Same idea as the pack's other analysis nodes (the freeze-frame detector, the scene-cut detector): measure it, don't feel it.

How it works

You feed it the same JSON frame arrays the other performance nodes produce, and it computes metrics:

• audio_frames_json - source audio features.
• refined_frames_json - from MKRFacePerformanceLipRefine (or compatible data).
• eye_frames_json - from MKRFacePerformanceEyeMotion.
• pose_frames_json - from MKRFacePerformancePoseMerge.

clip_id is just a label stored in the metrics so you can tell runs apart (default clip_alpha - rename it per shot). fps (default 60) drives all the timing math, so it must match the pipeline you're evaluating.

The evaluation utility (lib/face_performance/evaluation_utility.py) computes lip/audio sync lag in frames, blink behavior, motion outliers, and pose jitter. The outputs that actually matter:

• lag_frames (INT) - best-estimate lip-to-audio lag in frames. The headline number.
• blink_rate_per_minute (FLOAT) - estimated blink rate; humans land roughly 10–20/min, so wildly different numbers suggest the eye synthesis is doing something odd.
• pose_jitter_score (FLOAT) - mean per-frame pose delta across head and neck channels. High = wobbly.
• metrics_json - the full metric object (also where clip_id shows up).
• failures_json - array of threshold violations, from thresholds_json. Leave thresholds_json as {} to use the built-in defaults; pass overrides for any RegressionThresholds field when you have project-specific tolerances.

What it's for

The author's own framing says it best: it's for regression testing or comparing retarget settings inside a workflow. That's the killer use case - set up a baseline render, tweak a knob, re-evaluate, and read the numbers instead of re-watching both renders frame by frame. If you're iterating on max_delta_per_frame or blink intervals and want to know if you made things better or just different, this is the honest arbiter.

Installing and caveats

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/criskb/MKRShift_Nodes

Restart ComfyUI, or ComfyUI Manager search "MKRShift_Nodes". Pure Python, no new deps.

Caveats are the lane's usual ones: it's for structured JSON performance data, not raw video (all inputs must be JSON arrays, thresholds_json a JSON object - the validator checks). And there's zero community chatter about it yet, so treat the built-in thresholds as starting points, not gospel. But for its stated job - "did this change make the performance better or worse" - a numeric score beats vibes, and this node gives you one.