Understanding Pure Textual Reasoning for Blind Image Quality Assessment

TL;DR

This work investigates how textual reasoning contributes to blind image quality assessment (BIQA) by framing the problem as information flow among image, text, and score. It systematically compares three training paradigms—Chain-of-Thought, Self-Consistency, and Autoencoder-like learning—using a shared multimodal language backbone, revealing that Chain-of-Thought yields limited gains while Self-Consistency substantially closes the image–text gap, achieving a gap of about $0.02$ in PLCC and $0.03$ in SRCC on several benchmarks. The Autoencoder-like approach offers direction for more natural, quality-related textual explanations and combined image–text conditioning can boost performance. The results provide actionable insights for designing interpretable text-conditioned BIQA systems and offer a general framework applicable to other vision–language tasks without intermediate reasoning annotations.

Abstract

Textual reasoning has recently been widely adopted in Blind Image Quality Assessment (BIQA). However, it remains unclear how textual information contributes to quality prediction and to what extent text can represent the score-related image contents. This work addresses these questions from an information-flow perspective by comparing existing BIQA models with three paradigms designed to learn the image-text-score relationship: Chain-of-Thought, Self-Consistency, and Autoencoder. Our experiments show that the score prediction performance of the existing model significantly drops when only textual information is used for prediction. Whereas the Chain-of-Thought paradigm introduces little improvement in BIQA performance, the Self-Consistency paradigm significantly reduces the gap between image- and text-conditioned predictions, narrowing the PLCC/SRCC difference to 0.02/0.03. The Autoencoder-like paradigm is less effective in closing the image-text gap, yet it reveals a direction for further optimization. These findings provide insights into how to improve the textual reasoning for BIQA and high-level vision tasks.

Figures (3)

• Figure 1: Image-Text Gap. Traditional BIQA relies on visual features and lacks interpretability. Although MLLM-based BIQA generates both captions and scores, their relationship remains unclear. We examine the performance gap between image-based and text-based score prediction to reveal how textual reasoning contributes to interpretable BIQA.
• Figure 2: Training Paradigms. All models share the same MLLM backbone and perform two forward passes, where "1st" and "2nd" denote the first and second forward inferences of the same model. The 1st pass is always conditioned on the image. (a) Chain-of-Thought: The model first generates $N$ caption candidates from the image; each caption then produces $M$ score predictions through an independent second-stage inference. (b) Self-Consistency: The first pass outputs $N$ (caption, score) pairs, and each caption undergoes an additional inference step for score regression, receiving a self-consistency reward. (c) Autoencoder-like: The model takes the image and the ground-truth MOS during the first pass to generate reasoning text; the second pass regresses the score solely from this reasoning text.
• Figure 3: Comprehensive Analysis of Attention Distributions. Please zoom in to check the details. We analyze model behavior on the KonIQ koniq test set (3,015 images) by examining softmax-normalized attention values to estimate token contributions to score prediction. This shows the text-conditioned results. Different learning paradigms lead to distinct shifts in the tokens emphasized by the model. Image-conditioned results are shown in Section 3.2 in supplemental materials.