Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fine-Grained Human Pose Editing Assessment via Layer-Selective MLLMs

Ningyu Sun, Zhaolin Cai, Zitong Xu, Peihang Chen, Huiyu Duan, Yichao Yan, Xiongkuo Min, Xiaokang Yang

TL;DR

This work tackles the evaluation gap in text-guided human pose editing by introducing HPE-Bench, a fine-grained benchmark with $1{,}700$ samples from $17$ editing models and accompanying authenticity and quality annotations. It then presents a unified framework built on layer-selective multimodal LLMs, combining contrastive LoRA tuning and Layer Sensitivity Analysis to identify the optimal feature layer for joint authenticity detection and multi-dimensional quality regression. The method achieves state-of-the-art performance in both forensic detection and perceptual quality prediction, with strong alignment to human judgments and reliable model ranking. By linking forensic traces with perceptual quality through a shared representation, the approach offers a practical benchmark and a robust evaluation tool for advancing reliable pose-editing technologies.

Abstract

Text-guided human pose editing has gained significant traction in AIGC applications. However,it remains plagued by structural anomalies and generative artifacts. Existing evaluation metrics often isolate authenticity detection from quality assessment, failing to provide fine-grained insights into pose-specific inconsistencies. To address these limitations, we introduce HPE-Bench, a specialized benchmark comprising 1,700 standardized samples from 17 state-of-the-art editing models, offering both authenticity labels and multi-dimensional quality scores. Furthermore, we propose a unified framework based on layer-selective multimodal large language models (MLLMs). By employing contrastive LoRA tuning and a novel layer sensitivity analysis (LSA) mechanism, we identify the optimal feature layer for pose evaluation. Our framework achieves superior performance in both authenticity detection and multi-dimensional quality regression, effectively bridging the gap between forensic detection and quality assessment.

Fine-Grained Human Pose Editing Assessment via Layer-Selective MLLMs

TL;DR

This work tackles the evaluation gap in text-guided human pose editing by introducing HPE-Bench, a fine-grained benchmark with samples from editing models and accompanying authenticity and quality annotations. It then presents a unified framework built on layer-selective multimodal LLMs, combining contrastive LoRA tuning and Layer Sensitivity Analysis to identify the optimal feature layer for joint authenticity detection and multi-dimensional quality regression. The method achieves state-of-the-art performance in both forensic detection and perceptual quality prediction, with strong alignment to human judgments and reliable model ranking. By linking forensic traces with perceptual quality through a shared representation, the approach offers a practical benchmark and a robust evaluation tool for advancing reliable pose-editing technologies.

Abstract

Text-guided human pose editing has gained significant traction in AIGC applications. However,it remains plagued by structural anomalies and generative artifacts. Existing evaluation metrics often isolate authenticity detection from quality assessment, failing to provide fine-grained insights into pose-specific inconsistencies. To address these limitations, we introduce HPE-Bench, a specialized benchmark comprising 1,700 standardized samples from 17 state-of-the-art editing models, offering both authenticity labels and multi-dimensional quality scores. Furthermore, we propose a unified framework based on layer-selective multimodal large language models (MLLMs). By employing contrastive LoRA tuning and a novel layer sensitivity analysis (LSA) mechanism, we identify the optimal feature layer for pose evaluation. Our framework achieves superior performance in both authenticity detection and multi-dimensional quality regression, effectively bridging the gap between forensic detection and quality assessment.
Paper Structure (30 sections, 2 equations, 2 figures, 4 tables)

This paper contains 30 sections, 2 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Construction of HPE-Bench
  3. Image Collection and Prompt Engineering
  4. Source Data Curation
  5. Prompt Formulation
  6. Standardized Generation Protocol
  7. Generative Models
  8. Data Synthesis and Analysis
  9. Metadata Acquisition and Annotation
  10. Proposed Method
  11. Overview
  12. Contrastive Visual Tuning
  13. Layer Sensitivity Analysis (LSA)
  14. Task-Specific Inference
  15. Authenticity Detection.
  16. ...and 15 more sections

Figures (2)

  • Figure 1: Overview of our constructed HPE-Bench and the proposed evaluation task. HPE-Bench contains 1,700 standardized samples generated by 17 diverse state-of-the-art editing models. Our unified framework performs concurrent authenticity detection and multi-dimensional quality regression, providing scores for visual quality, editing alignment, and attribute preservation.
  • Figure 2: Overview of our proposed framework. The model employs a contrastive LoRA tuning strategy on visual encoder and MLLM to enhance sensitivity to pose-editing artifacts. A layer sensitivity analysis (LSA) module computes statistical metrics to select the optimal intermediate feature layer from the MLLM. Finally, the authentic decoder and quality assessment decoder are utilized for simultaneous authenticity detection and multi-dimensional quality scoring.