Identity Preservation in AI Face Swap: ArcFace vs AdaFace Benchmarks (2026)
Identity Preservation: ArcFace vs AdaFace
Identity preservation is the single most important quality dimension in face swap. The face embedding network is what tells the generator "this is the person we're transferring." Two embeddings dominate 2026 production stacks: ArcFace and AdaFace. We benchmarked both on a production face-swap pipeline.
What Each Embedding Optimizes For
ArcFace introduced additive angular margin loss in 2018, dominating face recognition benchmarks for years. Its embeddings are tightly clustered per identity — excellent for verification, strong for swap-time identity transfer.
AdaFace (2022) added quality-adaptive margin: it weights low-quality images differently during training, leading to embeddings more robust to blur, low resolution, and difficult lighting.
Benchmark Setup
- Test set: 1,000 face pairs from a held-out evaluation set covering varied angles, lighting, and image quality.
- Pipeline: Same Wan 2.2 face-swap generator, only the embedding network swapped.
- Metrics: ArcFace cosine similarity (yes, used as evaluation regardless of generation embedding — independent eval), human preference rating, artifact density.
- Hardware: H100 80GB.
Results — High-Quality Source Images
| Metric | ArcFace | AdaFace |
|---|---|---|
| Identity similarity (mean) | 0.79 | 0.78 |
| Identity similarity (p10) | 0.71 | 0.72 |
| Human preference | 52% | 48% |
| Wall-clock per swap | 1.0× baseline | 1.05× baseline |
On clean, high-quality sources, the two embeddings are statistically indistinguishable. Pick by integration ease.
Results — Low-Quality Source Images
| Metric | ArcFace | AdaFace |
|---|---|---|
| Identity similarity (mean) | 0.62 | 0.71 |
| Identity similarity (p10) | 0.48 | 0.61 |
| Human preference | 34% | 66% |
On blurry, low-resolution, or poorly-lit sources, AdaFace's quality-adaptive margin shines. The p10 case (worst-decile inputs) gap is large enough to matter at scale.
Results — Off-Angle Source
Both embeddings struggle past ~45° head turn. Neither is meaningfully better. The fix here is generator-side, not embedding-side: better source-image guidance and pose-aware generation networks.
Results — Cross-Demographic Performance
We split the test set by source demographic group (best practice for face-recognition benchmarking). AdaFace narrowed but did not eliminate the demographic performance gap that older ArcFace deployments have shown. The fundamental fix here is training data composition; both networks improve when trained on demographically balanced data.
Production Decision Guide
- Consumer face-swap with self-uploaded photos: AdaFace is the better default. User-uploaded photos vary widely in quality.
- Studio-grade workflow with curated high-quality sources: ArcFace is fine; the quality gain from AdaFace is marginal here.
- API serving mixed customer populations: AdaFace is the safer default — the worst-case is much better, and the best-case is statistically tied.
- Latency-critical real-time: ArcFace's slightly faster inference may matter at margin. Test on your hardware.
Hybrid Approaches
Some 2026 production stacks ensemble both: ArcFace for the primary identity vector, AdaFace as a quality-aware fallback when the source image quality score is below threshold. The ensemble adds ~5% latency for 10–15% better worst-case behavior. Worth it for general-audience consumer products.
Beyond ArcFace and AdaFace
Newer embeddings (CosFace 2.0, MagFace v2, several proprietary networks) have appeared in 2025 papers. None has displaced ArcFace/AdaFace as the production default yet — typically because the marginal gain doesn't justify the integration cost in established pipelines. Watch the 2026 NeurIPS and CVPR proceedings for candidates that might.
Evaluation Honesty
One trap: don't evaluate identity preservation using the same embedding network that drove generation. The model effectively optimized for that embedding's notion of identity, so the score is inflated. Always evaluate with an independent embedding (we used a separately-trained ArcFace for both ArcFace-driven and AdaFace-driven generations).
What DeepSwapAI Uses
DeepSwapAI's production pipeline uses a hybrid: AdaFace as the primary identity embedding with ArcFace as a verification check during quality scoring. This gives consumer-segment robustness with studio-grade verification. Documented in the research methodology page.
Bottom Line
For consumer or mixed-quality input scenarios, AdaFace is the better embedding in 2026. For studio-grade curated inputs, both are tied. Hybrid ensembles capture the worst-case improvement at modest cost. The bigger lever in identity preservation is generator-side architecture and training data balance — embedding choice is meaningful but not dominant.