Skip to content

Intra-finger Variability of Diffusion-based Latent Fingerprint Generation

Conference: CVPR 2026 arXiv: 2604.10040 Code: None Area: Image Generation / Biometrics Keywords: fingerprint synthesis, diffusion models, latent fingerprints, identity consistency, style diversity

TL;DR

This paper systematically evaluates the intra-finger variability of diffusion-model-based fingerprint synthesis. By constructing a latent fingerprint style library spanning 40 surface types and 15 development techniques, it enhances generation diversity and quantifies both local and global identity inconsistencies introduced during the generation process.

Background & Motivation

Background: Generative AI (GANs and DDPMs) has demonstrated the capability to produce high-quality synthetic fingerprint datasets. Fingerprint synthesis typically proceeds in two stages: generating unique identities (inter-finger variability) and generating multiple variants of the same identity (intra-finger variability). The second stage is particularly critical for latent fingerprint applications.

Limitations of Prior Work: (1) Existing models rely on holistic or random style transfer and cannot precisely specify forensic scenarios (e.g., "a latent fingerprint lifted from a glass bottle and developed with fluorescent powder"); (2) the stochasticity of the generation process may alter ridge structures and minutiae, compromising identity authenticity.

Key Challenge: An inherent tension exists between diversity and identity preservation — increasing style diversity may introduce greater identity inconsistency.

Goal: (1) Enhance style diversity in latent fingerprint generation; (2) rigorously quantify identity preservation capability.

Key Insight: A style library of 28,000 real latent fingerprints is constructed to enable precise style control, and a semi-automated framework is designed to evaluate identity consistency.

Core Idea: The style library enables controllable generation across 40+ latent fingerprint styles, while also revealing local inconsistencies in low-quality regions and global inconsistencies arising from style–reference mismatches within diffusion models.

Method

Overall Architecture

An extension of the second stage of the GenPrint framework: (1) a style library of 28,000 real latent fingerprints is curated from 7 datasets, covering 40 surface types and 15 development techniques; (2) style embeddings are used to precisely control the style of generated latent fingerprints; (3) a semi-automated framework compares minutiae and ridge patterns between reference and generated images.

Key Designs

  1. Latent Fingerprint Style Library:

    • Function: Enables precise style control for latent fingerprint generation.
    • Mechanism: 28,000 real latent fingerprints are curated from 7 diverse datasets and categorized into 40 surface types (glass, metal, paper, etc.) and 15 development techniques (powder, chemical, optical, etc.), yielding 40+ discrete styles. Style embeddings are extracted for each category and used as conditional inputs to GenPrint.
    • Design Motivation: The original GenPrint model is limited to generic or randomly styled latent fingerprints and cannot generate fingerprints corresponding to specific forensic scenarios.
  2. Semi-Automated Identity Consistency Evaluation Framework:

    • Function: Quantifies minutiae and ridge changes introduced during the generation process.
    • Mechanism: Synthetic versions are generated from manually annotated reference fingerprints and then aligned for comparison along two dimensions: (1) local inconsistencies — addition and removal of minutiae (detected automatically and verified manually); (2) global inconsistencies — hallucinated ridge patterns absent from the reference image.
    • Design Motivation: Conventional automated matching scores lack sufficient granularity; minutiae-level analysis is required to understand the specific mechanisms through which identity is corrupted.
  3. Root Cause Analysis of Inconsistencies:

    • Function: Explains the origin of identity inconsistencies.
    • Mechanism: Local inconsistencies primarily occur in regions where the reference image is of low quality — the model tends to "fabricate" details under uncertainty. Global inconsistencies arise when the reference image does not match the selected style embedding, causing the model to hallucinate non-existent ridge patterns.
    • Design Motivation: Understanding why inconsistencies occur is a prerequisite for guiding model improvement.

Loss & Training

The method builds upon GenPrint's ID-Net (a fine-tuned ControlNet), conditioned on style embeddings and text prompts. The training strategy follows the GenPrint framework without modification.

Key Experimental Results

Main Results

Evaluation Dimension Result
Style Coverage 40 surface types × 15 development techniques
Data Scale 28,000 real latent fingerprints
Identity Preservation Preserved in most cases; minor local inconsistencies observed
Global Hallucination Occurs under style–reference mismatch conditions

Ablation Study

Condition Local Inconsistency Global Inconsistency Notes
High-quality reference Low Low Best-case scenario
Low-quality reference High Low Poor-quality regions induce minutiae changes
Style mismatch Low High Mismatch between reference and style embedding

Key Findings

  • The generation process preserves identity in most cases, but low-quality regions are more prone to local inconsistencies.
  • Mismatch between style embeddings and reference images is the primary cause of global hallucinations.
  • These findings provide clear directions for improving synthetic fingerprint generators.

Highlights & Insights

  • Systematic Identity Consistency Analysis: This work is the first to quantify identity preservation in diffusion-model-based fingerprint generation at the minutiae level.
  • Forensic Scenario Controllability: The style library spanning 40 surface types × 15 development techniques endows latent fingerprint generation with practical forensic training value.
  • Actionable Root Cause Analysis: The two identified root causes — low-quality reference regions and style–reference mismatch — directly inform model improvement strategies.

Limitations & Future Work

  • The semi-automated evaluation framework still requires human involvement and cannot be fully automated.
  • Although broad, the style library's coverage may remain incomplete.
  • The paper provides analysis only and does not propose methods to resolve the identified inconsistencies.
  • vs. Wyzykowski et al.: Supports only 3 coarse-grained styles (good/bad/ugly), whereas this work achieves 40+ fine-grained styles.
  • vs. Joshi et al.: Employs neural style transfer but lacks style control; this work achieves precise control through the style library.

Rating

  • Novelty: ⭐⭐⭐ The style library construction is valuable, though methodological innovation is limited.
  • Experimental Thoroughness: ⭐⭐⭐⭐ The identity consistency analysis is thorough and rigorous.
  • Writing Quality: ⭐⭐⭐⭐ Problem formulation is clear and well-structured.
  • Value: ⭐⭐⭐ Offers direct practical value to the forensic and fingerprint recognition communities.