PCQA: A Strong Baseline for AIGC Quality Assessment Based on Prompt Condition

TL;DR

This work introduces Prompt Condition Quality Assessment (PCQA), a framework that evaluates AIGC image and video quality by conditioning MOS predictions on the generating prompt. It employs a dual-source, frozen CLIP text encoder to capture prompt semantics, a trainable visual backbone with adapters, and a feature mixer to fuse modalities, followed by an MLP regressor; predictions are stabilized via an ensemble across multiple backbones. Empirical results on AIGIQA-20K and T2VQA-DB show strong SRCC and PLCC scores, outperforming several baselines and establishing a universal baseline for AIGC QoE assessment. The authors also provide extensive ablations and discuss practical considerations such as aspect-ratio preservation, spatial information, longer video sequences, and computational costs, offering a clear pathway for future improvements. Overall, PCQA demonstrates that prompt-centric multimodal representations can robustly capture perceptual quality in AI-generated content and can meaningfully guide quality assessment in real-world AIGC systems.

Abstract

The development of Large Language Models (LLM) and Diffusion Models brings the boom of Artificial Intelligence Generated Content (AIGC). It is essential to build an effective quality assessment framework to provide a quantifiable evaluation of different images or videos based on the AIGC technologies. The content generated by AIGC methods is driven by the crafted prompts. Therefore, it is intuitive that the prompts can also serve as the foundation of the AIGC quality assessment. This study proposes an effective AIGC quality assessment (QA) framework. First, we propose a hybrid prompt encoding method based on a dual-source CLIP (Contrastive Language-Image Pre-Training) text encoder to understand and respond to the prompt conditions. Second, we propose an ensemble-based feature mixer module to effectively blend the adapted prompt and vision features. The empirical study practices in two datasets: AIGIQA-20K (AI-Generated Image Quality Assessment database) and T2VQA-DB (Text-to-Video Quality Assessment DataBase), which validates the effectiveness of our proposed method: Prompt Condition Quality Assessment (PCQA). Our proposed simple and feasible framework may promote research development in the multimodal generation field.

Figures (5)

• Figure 1: Overview of the Prompt Condition Quality Assessment (PCQA) model. The AIGC content and the corresponding prompt used to generate it are input separately. The information from the prompt will be encoded by the hybrid CLIP text encoder and used as a condition for visual quality assessment, with the trainable feature adapter to align the feature from different modals. The final MOS regression result is obtained through a feature mixer and an MLP regressor.
• Figure 2: Feature mixer and regression head. Concatenation or dot product are used as feature mixer. This enables the visual features and the textual features of the prompt to interact.
• Figure 3: Overview of the final quality score computation strategy by model ensemble. The final score is average blending 3 models with different vision backbone.
• Figure 4: Ablation study on vision encoder choice. The score represents the average of SRCC and PLCC obtained through cross-validation on the AIGIQA-20K dataset.
• Figure 5: Ablation study on input resolution and model size. Input resolution between 448 to 640 leads to better results. Models with medium or larger sizes are also more likely to achieve better results.