One Model, Two Minds: Task-Conditioned Reasoning for Unified Image Quality and Aesthetic Assessment
Authors: Wen Yin, Cencen Liu, Dingrui Liu, Bing Su, Yuan-Fang Li, Tao He
Organizations: University of Electronic Science and Technology of China, Chengdu, Sichuan, China · Jiigan Technology, Chengdu, Sichuan, China · Faculty of Information Technology, Monash University, Melbourne, Australia · The Laboratory of Intelligent Collaborative Computing of UESTC, Chengdu, Sichuan, China
Unifying Image Quality Assessment (IQA) and Image Aesthetic Assessment (IAA) in a single multimodal large language model is appealing, yet existing methods adopt a task-agnostic recipe that applies the same reasoning strategy and reward to both tasks. We show this is fundamentally misaligned: IQA relies on low-level, objective perceptual cues and benefits from concise distortion-focused reasoning, whereas IAA requires deliberative semantic judgment and is poorly served by point-wise score regression. We identify these as a reasoning mismatch and an optimization mismatch, and provide empirical evidence for both through controlled probes. Motivated by these findings, we propose TATAR (Task-Aware Thinking with Asymmetric Rewards), a unified framework that shares the visual-language backbone while conditioning post-training on each task's nature. TATAR combines three components: fast--slow task-specific reasoning construction that pairs IQA with concise perceptual rationales and IAA with deliberative aesthetic narratives; two-stage SFT+GRPO learning that establishes task-aware behavioral priors before reward-driven refinement; and asymmetric rewards that apply Gaussian score shaping for IQA and Thurstone-style completion ranking for IAA. Extensive experiments across eight benchmarks demonstrate that TATAR consistently outperforms prior unified baselines on both tasks under in-domain and cross-domain settings, remains competitive with task-specific specialized models, and yields more stable training dynamics for aesthetic assessment. Our results establish task-conditioned post-training as a principled paradigm for unified perceptual scoring. Our code is publicly available at https://github.com/yinwen2019/TATAR.
Figures & tables
Figure 1. From shared to task-conditioned post-training. Prior unified methods apply the same response behavior and reward to both IQA and IAA, conflating two tasks with fundamentally different decision regimes. TATAR decouples what is shared (the VL backbone) from what is adapted (post-training): Stage 1 instills task-specific reasoning modes via fast–slow SFT, and Stage 2 refines scoring with asymmetric rewards matched to each task’s supervision geometry.
Figure 2. Motivation for task-conditioned post-training. (a) Adding a CoT instruction yields asymmetric effects: it consistently helps IAA but hurts IQA, demonstrating that a uniform reasoning strategy is suboptimal. (b) Under shared-reward RFT, IQA converges to short, stable completions while IAA produces long, variable ones, revealing that the two tasks inhabit different response and optimization regimes.
Figure 3. Task-conditioned reasoning construction. IQA rationales are synthesized by score-conditioned reverse inference from image–score pairs, while IAA rationales are obtained by summarizing structured aesthetic annotations. A unified judge then filters low-quality generations, producing the final fast–slow reasoning corpus used in Stage 1 SFT.
Figure 4. Overview of TATAR. TATAR is a unified framework for IQA and IAA with task-conditioned reasoning and reward design. In Stage 1, format-oriented SFT aligns the model to a shared <think>/<answer> output schema while teaching distinct response modes, with concise reasoning for IQA and more deliberative reasoning for IAA. In Stage 2, task-conditioned GRPO further refines the model with a shared optimization procedure but asymmetric rewards, where IQA is optimized with a Gaussian score reward Rfmt+Rscore and IAA is optimized with a ranking reward Rfmt+Rrank .
Figure 5. Two examples from the constructed reasoning corpus ( QACoT-score dataset). IQA (top) samples use concise, distortion-focused rationales, while IAA (bottom) samples use longer and more integrative aesthetic rationales.
Models
IAA task
IQA task
ArtiMuse-10K
AVA
TAD66K
FLICKR-AES
Avg.
KonIQ-10K
SPAQ
KADID
PIPAL
Avg.
Proprietary Models
Large-scale pretraining, cross-domain testing
GPT-4o ( Achiam et al., 2023 )
33.3/27.6
50.9/48.5
27.8/28.2
60.5/59.7
43.1/41.0
69.5/74.4
87.4/88.1
67.7/64.6
32.5/34.9
64.3/65.5
Llama-4 ( AI, 2025 )
20.4/14.7
34.5/32.9
23.6/21.0
54.8/50.6
33.3/29.8
50.3/65.3
-4.1/0.7
-9.9/-0.4
-0.7/2.3
8.9/17.0
Gemini-2.5-Pro ( Team et al., 2023 )
18.7/3.5
24.8/10.0
14.3/3.7
35.7/20.6
23.4/9.5
58.2/31.6
8.7/21.2
43.6/27.4
22.5/-1.9
33.3/19.6
Claude-4.5 ( Lindsey et al., 2025 )
4.1/2.7
0.3/1.3
4.0/4.7
3.7/4.9
3.0/3.4
-3.7/-4.3
3.6/8.5
22.3/27.3
-13.1/-8.8
2.3/5.7
Table 1. Performance comparison on unified IAA and IQA scoring. Results are reported as SRCC/PLCC (%). For specialized models, red columns denote in-domain and green columns denote out-of-domain evaluation. ♣ marks results cited from UniPercept ( Cao et al., 2025a ) , ♠ denotes single-task evaluation, and * denotes models retrained under our protocol.
Settings
IAA task
IQA task
SFT
RL
Reward
ArtiMuse
AVA
TAD66K
FLICKR
Avg.
KonIQ
SPAQ
KADID
PIPAL
Avg.
✓
45.9/45.1
30.3/34.2
18.9/19.7
36.3/37.8
32.9/34.2
89.2/91.8
81.5/82.3
70.9/69.0
40.2/41.2
70.5/71.1
✓
Hard-only
40.0/43.1
56.1/55.1
30.5/31.3
53.4/53.4
45.0/45.7
90.1/92.2
82.1/81.7
70.3/68.5
41.3/41.5
70.9/71.0
✓
Gauss.-only
40.8/43.7
52.1/52.3
31.2/31.9
54.0/54.1
44.5/45.5
90.8/92.6
82.6/82.1
69.9/68.2
42.0/42.1
71.3/71.2
✓
Rank-only
41.3/44.1
55.8/51.4
30.9/31.6
54.6/54.8
45.6/45.5
91.2/92.9
82.3/81.9
70.8/69.1
42.6/42.8
71.7/71.7
✓
Task-cond.
41.7/44.6
55.6/51.3
31.3/32.0
55.0/55.2
46.0/45.9
91.7/93.2
82.8/82.3
71.3/69.6
43.0/43.2
72.2/72.1
Table 2. Ablation study on training paradigm and reward design. Results are reported as SRCC/PLCC (%). Bold and underlined values indicate first and second place. Task-cond. means Gaussian for IQA and Rank for IAA.
Figure 6. Training dynamics analysis covering completion behavior, reward trends, and optimization stability.
Figure 7. Case study comparing TATAR and UniPercept on representative IQA (top) and IAA (bottom) examples. Ground-truth scores and prediction errors are shown for each model.
Recent advances in reasoning-induced image quality assessment (IQA) have demonstrated the power of reinforcement learning to rank (RL2R) for training vision-language models (VLMs) to assess perceptual quality. However, existing approaches operate at a single granularity, predicting only an overall quality score, while overlooking the multi-dimensional nature of human quality perception, which encompasses attributes such as sharpness, color fidelity, noise level, and compositional aesthetics. In this paper, we propose MG-IQA (Multi-Granularity IQA), a multi-granularity reasoning framework that extends RL2R to jointly assess overall image quality and fine-grained quality attributes within a single inference pass. Our approach introduces three key innovations: (1) an attribute-aware prompting strategy that elicits structured multi-attribute reasoning from VLMs; (2) a multi-dimensional Thurstone reward model that computes attribute-specific fidelity rewards for group relative policy optimization; and (3) a cross-domain alignment mechanism that enables stable joint training across synthetic distortion, authentic distortion, and AI-generated image datasets without perceptual scale re-alignment. Extensive experiments on eight IQA benchmarks demonstrate that MG-IQA consistently outperforms state-of-the-art methods in both overall quality prediction (average SRCC improvement of 2.1%) and attribute-level assessment, while generating interpretable, human-aligned quality descriptions.
Multimodal large language models (MLLMs) have shown strong potential for image quality assessment (IQA) by improving consistency between quality ratings and their underlying reasoning. However, most approaches supervise reasoning through human-provided ratings and rarely examine whether it faithfully reflects image quality. Rating accuracy alone does not ensure faithful reasoning; a shared reward also obscures supervision sources and may reinforce unfaithful reasoning when a correct rating occurs by chance. To improve the faithfulness and reliability of blind IQA, we aim to (1) decouple credit assignment for reasoning and rating and (2) provide verifiable supervision for faithful reasoning. We introduce MR-IQA-2, an actor-editor-judge framework that operationalizes reasoning-editing-reflection. The actor generates quality reasoning for an input image, and the editor revises the image according to the identified quality factors. A frozen judge compares the original and edited images and provides reflective supervision for the actor's reasoning. MR-IQA-2 further uses fine-grained credit assignment to decouple reasoning and rating supervision. Judge feedback supervises reasoning, whereas human ratings supervise the predicted rating. Masked token-specific updates distinguish these signals while preserving the causal relation from reasoning to rating. Across IQA benchmarks, MR-IQA-2 achieves competitive rating alignment with humans. Visual reflection also enables richer and more faithful visual understanding beyond rating, which may inform image-quality optimization and related downstream tasks. Code is available at https://github.com/RobinY99/MR-IQA-2.
Image Quality Assessment (IQA) in open-world environments remains challenging due to limited generalization and interpretability. Recent approaches based on multimodal large language models (MLLMs) introduce textual reasoning for quality prediction, yet their judgments rely heavily on semantically biased internal representations, making them insensitive to low-level perceptual degradations. We propose IQA-T1, a tool-based visual evidence reasoning framework that augments MLLM reasoning with explicit perceptual observations. During inference, the model autonomously invokes specialized analysis tools to generate structured visual evidence, such as noise residual maps, gradient statistics, and frequency spectra, which are progressively integrated into the reasoning process. To support this paradigm, we construct Q-Tool, a dataset containing 11k multimodal reasoning chains grounded in tool-generated evidence. Extensive experiments on seven IQA benchmarks show that IQA-T1 achieves the best overall performance across datasets while producing interpretable and evidence-grounded quality assessments. Code and dataset are available at https://github.com/zibuyu-02/IQA-T1.
Jinjian Wu, Jiaqi Tang, Wei Wei +5
School of Computer Science, Northwestern Polytechnical University, Xi’an, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China