Identity-Preserving Talking Head Generation on CelebV-HQ hard (cross-video)
47.7Speaker SimilarityID-LoRA (Ours)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| ID-LoRA (Ours)Video Backbone=LTX-22026.03 | 47.7 | 87.4 | 8.49 | 3.9 | 0.363 | 11.3 | |
| CosyVoice 3.0 + WAN2.2Video Backbone=WAN2.22026.03 | 39.1 | 89 | 11.4 | 1.5 | 0.249 | 36.2 | |
| Kling 2.6 ProCommercial Model=true2026.03 | 38.5 | 85.4 | 9.49 | 3.47 | 0.316 | 12.1 | |
| ElevenLabs + WAN2.2Video Backbone=WAN2.22026.03 | 35.7 | 89.4 | 11.86 | 1.72 | 0.238 | 15.4 | |
| VoiceCraft + WAN2.2Video Backbone=WAN2.22026.03 | 34.4 | 89.2 | 10.6 | 1.33 | 0.258 | 42.7 |