Research map / Image generation / Autoregressive and latent generation

Diff-Palm: Realistic Palmprint Generation with Polynomial Creases and Intra-Class Variation Controllable Diffusion Models

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: Classical generative image synthesis · 6 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.

Proposes Diff-Palm, which uses fourth-order polynomial palm creases sampled from a Gaussian fit and a crease-conditioned diffusion model with K-step noise-sharing sampling to generate palmprint datasets that train recognizers without real-data fine-tuning.

Problem

Palmprint recognition lacks large public datasets, and synthetic palmprints from Bezier-crease conditional GANs degrade badly without real-data fine-tuning because of inaccurate crease representation and poor balance between intra-class variation and identity consistency.

Approach

Replace Bezier crease curves with fourth-order polynomial creases whose coefficients follow a multivariate Gaussian estimated from real palms, condition a channel-concatenated diffusion model on these creases, and control intra-class variation by sharing the initial noise across the same identity during the last K sampling steps.

Key results

• Recognition model trained solely on Diff-Palm synthetic data (2k IDs, K=500) reaches average TAR@FAR=1e-6 of 0.7729, versus 0.7546 for real data and 0.5561 for PCE-Palm across five public datasets.
• Scaling Diff-Palm to 60k IDs (1.2M images) raises the average TAR to 0.9026, showing performance increases with number of generated identities.
• Polynomial creases outperform Bezier creases in the diffusion setting (0.7729 vs 0.5719 average TAR).

Measured effects reported in the paper

Surprising findings

• Applying noise-sharing to the first K sampling steps yields low intra-class diversity and worse recognition, while sharing in the last K steps works better.
• Increasing the crease-distribution variance scale gamma above or below 1 both reduce recognition performance.

Limitations

• Generator is trained on a private internet-collected dataset rather than public palmprint data.
• Performance depends on hyperparameter K and gamma that must be tuned, with trade-off between identity consistency and intra-class diversity.

Experimental setup

Datasets: CASIA PolyU Tongji MPD XJTU-UP IITD NTU-CP-v1 anonymous internet palmprint dataset (48,000 images)

Benchmarks: CASIA PolyU TongJi MPD XJTU-UP

Metrics: TAR@FAR=1e-6 identity consistency intra-class diversity class uniqueness t-SNE visualization

Baselines: PCE-Palm RPG-Palm IDiff-Face Vec2Face Real data

Builds on: DDPM PCE-Palm Bezierpalm ArcFace IDiff-Face

Closest papers by meaning

• PCE-Palm: Palm Crease Energy Based Two-Stage Realistic Pseudo-Palmprint Generation (AAAI 2024 · similarity 0.91)
• PVTree: Realistic and Controllable Palm Vein Generation for Recognition Tasks (AAAI 2025 · similarity 0.84)
• Towards Explainable and Unprecedented Accuracy in Matching Challenging Finger Crease Patterns (CVPR 2025 · similarity 0.81)
• RHanDS: Refining Malformed Hands for Generated Images with Decoupled Structure and Style Guidance (AAAI 2025 · similarity 0.81)
• HanDiffuser: Text-to-Image Generation With Realistic Hand Appearances (CVPR 2024 · similarity 0.80)
• FoundHand: Large-Scale Domain-Specific Learning for Controllable Hand Image Generation (CVPR 2025 · similarity 0.80)