Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HyperAlign: Hyperbolic Entailment Cones for Adaptive Text-to-Image Alignment Assessment

Wenzhi Chen, Bo Hu, Leida Li, Lihuo He, Wen Lu, Xinbo Gao

TL;DR

HyperAlign introduces a hyperbolic-geometry framework for Text-to-Image Alignment Assessment that captures hierarchical entailment between text prompts and images. By projecting CLIP features into the Lorentz hyperbolic space and extracting geometric primitives (distance, exterior angle, entailment aperture), it uses dynamic supervision to encode alignment as continuous geometric constraints and employs an adaptive modulation regressor to calibrate Euclidean cosine similarity for final scoring. The approach delivers state-of-the-art performance on multiple AGIQA benchmarks and shows strong cross-database generalization, attributed to geometry-driven representations that are robust to distribution shifts. Overall, HyperAlign advances T2IAA by uniting hyperbolic hierarchical modeling with sample-specific score calibration, enabling more accurate and transferable alignment assessments in AI-generated imagery.

Abstract

With the rapid development of text-to-image generation technology, accurately assessing the alignment between generated images and text prompts has become a critical challenge. Existing methods rely on Euclidean space metrics, neglecting the structured nature of semantic alignment, while lacking adaptive capabilities for different samples. To address these limitations, we propose HyperAlign, an adaptive text-to-image alignment assessment framework based on hyperbolic entailment geometry. First, we extract Euclidean features using CLIP and map them to hyperbolic space. Second, we design a dynamic-supervision entailment modeling mechanism that transforms discrete entailment logic into continuous geometric structure supervision. Finally, we propose an adaptive modulation regressor that utilizes hyperbolic geometric features to generate sample-level modulation parameters, adaptively calibrating Euclidean cosine similarity to predict the final score. HyperAlign achieves highly competitive performance on both single database evaluation and cross-database generalization tasks, fully validating the effectiveness of hyperbolic geometric modeling for image-text alignment assessment.

HyperAlign: Hyperbolic Entailment Cones for Adaptive Text-to-Image Alignment Assessment

TL;DR

HyperAlign introduces a hyperbolic-geometry framework for Text-to-Image Alignment Assessment that captures hierarchical entailment between text prompts and images. By projecting CLIP features into the Lorentz hyperbolic space and extracting geometric primitives (distance, exterior angle, entailment aperture), it uses dynamic supervision to encode alignment as continuous geometric constraints and employs an adaptive modulation regressor to calibrate Euclidean cosine similarity for final scoring. The approach delivers state-of-the-art performance on multiple AGIQA benchmarks and shows strong cross-database generalization, attributed to geometry-driven representations that are robust to distribution shifts. Overall, HyperAlign advances T2IAA by uniting hyperbolic hierarchical modeling with sample-specific score calibration, enabling more accurate and transferable alignment assessments in AI-generated imagery.

Abstract

With the rapid development of text-to-image generation technology, accurately assessing the alignment between generated images and text prompts has become a critical challenge. Existing methods rely on Euclidean space metrics, neglecting the structured nature of semantic alignment, while lacking adaptive capabilities for different samples. To address these limitations, we propose HyperAlign, an adaptive text-to-image alignment assessment framework based on hyperbolic entailment geometry. First, we extract Euclidean features using CLIP and map them to hyperbolic space. Second, we design a dynamic-supervision entailment modeling mechanism that transforms discrete entailment logic into continuous geometric structure supervision. Finally, we propose an adaptive modulation regressor that utilizes hyperbolic geometric features to generate sample-level modulation parameters, adaptively calibrating Euclidean cosine similarity to predict the final score. HyperAlign achieves highly competitive performance on both single database evaluation and cross-database generalization tasks, fully validating the effectiveness of hyperbolic geometric modeling for image-text alignment assessment.
Paper Structure (16 sections, 9 equations, 5 figures, 2 tables, 1 algorithm)

This paper contains 16 sections, 9 equations, 5 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Text-to-Image Alignment Assessment Methods
  4. Hyperbolic Vision-Language Models
  5. Method
  6. Overall Framework
  7. Hyperbolic Space Mapping
  8. Dynamic-Supervision Entailment Modeling
  9. Adaptive Modulation Regressor
  10. Experiments
  11. Experimental Setup
  12. Single Database Performance Evaluation
  13. Ablation Studies
  14. Visualization Analysis
  15. Cross Database Performance Evaluation
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Illustration of the proposed geometric constraint mechanism.
  • Figure 2: Overall architecture of HyperAlign.
  • Figure 3: Visualization of dynamic-supervision entailment cones. High-alignment samples are constrained within narrow cones, while low-alignment samples are allowed wider cones.
  • Figure 4: Ablation study of HyperAlign on three standard datasets.
  • Figure 5: Feature space visualization on AIGCIQA2023 dataset using t-SNE for Euclidean features and CO-SNE for hyperbolic features.