Research map / Image generation / Editing and style transfer

Non-Colliding Biometric Identities for Digital Entities: Geometry, Capacity, and Million-Scale Virtual Identity Provisioning

NeurIPS 2026 · Image generation › Editing and style transfer · approach: Diffusion face generation and editing · 0 citations (OpenAlex)

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.

Introduces Biometric Identity Provisioning (BIP): repulsion-based allocation of non-colliding virtual face identity embeddings against a real gallery, plus GapGen (InstantID fine-tuned with a curriculum) to realize them as 1024px faces, and the v-LFW benchmark.

Problem

Digital entities (AI agents, humanoid robots) need persistent biometric identities that do not collide with any enrolled real human face in a recognition system, while remaining inter-class separable and realizable as high-fidelity face images.

Approach

Treat virtual identity provisioning as a constrained packing problem on the real identity manifold: repulsion direction away from nearest real neighbors plus PCA-aware noise, with exact hard cosine checks against the gallery and among virtual identities; a gap-aware generator (GapGen) then realizes out-of-distribution embeddings via a real/virtual curriculum with round-trip identity loss.

Key results

• BIP allocation provisions 10M virtual embeddings against 360K real identities at tau=0.391 with 100% non-collision and 100% inter-class separability (99.98% at 10M).
• BIP+GapGen at 1M achieves 98.07% non-collision, FID 56.47, and 88.75% round-trip pass rate at 1024x1024, versus Vec2Face at 53.92% non-collision and Arc2Face at 0% RT pass rate.
• On the combined LFW+v-LFW set, BIP+IAPCT reaches 99.80% R-R accuracy, 0.01% R-V FAR, and 98.13% real-vs-virtual detection AUC.

Surprising findings

• Vec2Face-generated identities collide with real enrollees at non-trivial rates (about 46% collision at 10K), showing that existing synthetic-identity generators do not avoid real identities.
• Arc2Face achieves high non-collision (93.9%) yet a 0% round-trip pass rate, indicating embedding fidelity fails after re-encoding.

Limitations

• Non-collision is guaranteed only against the enrolled gallery at provisioning time; future enrollees require a safety buffer and periodic revocation.
• Capacity bounds rely on a spherical submanifold approximation rather than a precise face-manifold characterization.
• Image realization validated only up to 1M identities; GapGen fidelity degrades at large alpha and only ArcFace ResNet-100 is tested.

Experimental setup

Datasets: Glint360K FFHQ LFW WebFace4M IJB-B AIFaceFairnessBench

Benchmarks: v-LFW LFW IJB-B

Metrics: non-collision rate inter-class separability FID round-trip pass rate verification accuracy FAR AUC

Baselines: DCFace Arc2Face Vec2Face CAS (real) DCFace+AdaFace

Builds on: InstantID ArcFace DCFace Vec2Face HyperFace

Closest papers by meaning

• $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.84)
• Non-Adaptive Adversarial Face Generation (NeurIPS 2025 · similarity 0.84)
• Gaussian Differentially Private Human Faces Under a Face Radial Curve Representation (ICLR 2025 · similarity 0.84)
• Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.83)
• ID-to-3D: Expressive ID-guided 3D Heads via Score Distillation Sampling (NeurIPS 2024 · similarity 0.83)
• HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.83)