Research map / Label-efficient and robust learning / Recognition and long-tail
GIF: Generative Inspiration for Face Recognition at Scale
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Replaces scalar identity labels in face recognition training with structured integer identity codes built by hierarchically clustering uniformly spread CLIP-initialized hyperspherical code vectors, so the classifier cost scales as O(log m) instead of O(m).
Problem
Softmax/Angular Margin Softmax training in face recognition scales linearly with the number of identities, and subset-sampling efficient-training methods still scale linearly and suffer from minority collapse on unbalanced datasets.
Approach
Tokenize each identity into a sequence of integer tokens (identity code) via hierarchical k-means over hyperspherical code vectors that are initialized from CLIP class means and spread uniformly with a Gaussian potential loss; the backbone then predicts l parallel v-way token classifications plus a regression loss pulling embeddings to their code vectors.
Key results
- GIF outperforms prior efficient-training methods by 1.52% TAR@FAR=1e-4 on IJB-B with WebFace42M and ResNet-100.
- GIF outperforms ArcFace by 1.15% and 1.2% TAR@FAR=1e-4 on IJB-B and IJB-C with WebFace4M and ResNet-100.
- GIF trains more than 20% faster than FC and 15% faster than PFC on WebFace42M with ResNet-100.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace scalar-label training with structured identity codes (GIF) | Efficient face recognition methods · IJB-B | TAR@FAR=1e−4: +1.52% | Helped |
| Replace scalar-label training with structured identity codes (GIF) | Efficient face recognition methods · IJB-C | TAR@FAR=1e−4: +0.6% | Helped |
| Add explicit angular-regression embedding supervision (LAR) to code-prediction loss (LC) (ablation) | GIF trained on MS1MV2 with ResNet-100 · IJB-B | TAR@FAR=1e−4 | Helped |
| Set the loss balance to γ = 1 (ablation) | GIF trained on MS1MV2 with ResNet-100 · IJB-B and IJB-C | TAR@FAR=1e−4 | Helped |
| Use only code-prediction loss LC instead of angular-regression loss LAR (ablation) | GIF trained on MS1MV2 with ResNet-100 · IJB-B and IJB-C | TAR@FAR=1e−4 | Helped |
| Replace structured identity codes with randomly assigned atomic codes (ablation) | GIF trained with code-prediction loss only · LFW, CFP-FP, CPLFW, CALFW, and AgeDB | verification accuracy | Hurt |
| Use two identity-code tokens (ablation) | GIF trained with code-prediction loss only · CASIA-WebFace, MS1MV2, and WebFace4M | average verification performance | Hurt |
| Increase code length from l = 2 to l = 4 (ablation) | GIF trained with code-prediction loss only · CASIA-WebFace, MS1MV2, and WebFace4M | average verification performance | Helped |
| Increase code length to l = 6 (ablation) | GIF trained with code-prediction loss only · CASIA-WebFace, MS1MV2, and WebFace4M | average verification performance | Hurt |
| Use optimized hyperspherical code vectors instead of initial mean CLIP embeddings (ablation) | GIF with ResNet-100 trained on MS1MV2 · IJB-B | TAR@FAR=1e-4: 24.30 -> 95.05 | Helped |
| Use optimized hyperspherical code vectors instead of initial mean CLIP embeddings (ablation) | GIF with ResNet-100 trained on MS1MV2 · IJB-C | TAR@FAR=1e-4: 29.62 -> 96.77 | Helped |
Surprising findings
- Atomic (randomly assigned) identity codes fail to train the model, performing barely above random guessing.
- Initializing code vectors from CLIP class means alone yields near-random IJB-B performance (24.30% TAR at FAR=1e-4 vs 95.05% for optimized codes).
Limitations
- Code length and token range need per-dataset tuning (v between 5 and 25).
- Relies on a fixed CLIP-based initialization and a regression loss that is only evaluated within the paper's setups.
Experimental setup
Datasets: MS1MV2 WebFace4M WebFace12M WebFace42M
Benchmarks: LFW CPLFW CALFW CFP-FP AgeDB IJB-B IJB-C
Metrics: verification accuracy TAR@FAR=1e-4
Baselines: ArcFace CosFace PFC DCQ F2C Virtual FC AdaFace CurricularFace BroadFace MagFace
Builds on: ArcFace PFC CLIP DCQ
Closest papers by meaning
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.87)
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.86)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.86)
- CemiFace: Center-based Semi-hard Synthetic Face Generation for Face Recognition (NeurIPS 2024 · similarity 0.85)
- $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.85)
- Masked Face Recognition with Generative-to-Discriminative Representations (ICML 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.