How do current avatar models compare on hand fidelity?

The question

Hand quality is a notorious failure mode for diffusion video generation, called out repeatedly across the survey. mimicmotion's regional loss amplification on high-confidence hand keypoints was introduced specifically to address it. Where does hand fidelity actually stand across the models surveyed?

Why it matters

For full-body avatars, hands are visible, expressive, and high-frequency in a way that makes failure modes glaringly obvious. A model that can't render hands cleanly is unusable for many real applications (sign language, gesture-heavy presentations, demonstrations).

What we currently believe

• MimicMotion explicitly addresses hand distortion via training-loss reweighting on high-confidence hand keypoints.
• Runway Act-Two markets full hand tracking as a feature.
• Wan-Animate and AnimateAnyone V2 don't explicitly call out hand fidelity strategies, but report strong human-eval scores generally.
• No direct head-to-head hand-fidelity comparison exists in the source.

Evidence we have

• 2026-05-07-ai-avatar-motion-mimicking-models-survey notes this is an unresolved comparison.
• mimicmotion's regional loss amplification is documented.

Evidence we need

• Side-by-side hand renders of the same driving sequence across MimicMotion, Wan-Animate, StableAnimator, and Act-Two.
• Whether hand-specific evaluation metrics (e.g., MPJPE on hand keypoints, hand-region FID) are used in the field.

How to resolve

• Look for hand-specific evaluation papers in 2025 video generation survey work.
• Practical: run the same driving video through 2–3 open-source models and compare frames.

Related

• mimicmotion
• runway-act-two
• wan-animate
• independent-avatar-benchmarks