EVALUATION.md

# EchoMimicV2 Accelerated evaluation notes

## Expected strengths

- More expressive half-body motion than portrait-only models such as SadTalker.
- Official English and Mandarin examples and training support.
- Accelerated weights use six denoising steps instead of the slower standard path.
- Official documentation reports roughly 50 seconds for 120 frames on A100,
  compared with roughly seven minutes for the earlier path.

## Constraints

- A precomputed pose sequence is mandatory.
- Pose templates have a finite number of frames, so long audio is truncated.
- The complete dependency set is substantially larger than SadTalker or MuseTalk.
- Quality depends on matching the reference person's framing to the pose template;
  a centered, front-facing half-body image is recommended.

## Evaluation checklist

1. Identity consistency, especially eyes and hair.
2. Mandarin and English lip synchronization.
3. Hand and arm artifacts under each pose template.
4. Background stability.
5. Runtime for 48, 72, and 120 frames.
6. Peak ZeroGPU memory and whether repeated runs reuse loaded models safely.