acceptodds
Under review as a conference paper at ICLR 2027

PCPD: Partial Complementary Policy Distillation for Multi-Dimensional AIGC Image Quality Assessment

Abstract

Currently, AI generated content (AIGC) for image generation has achieved impressive results. Since the training process of AIGC models often requires evaluating the quality of the generated images, a high performance Image Quality Assessment (IQA) model is urgently needed. However, due to the exceptionally high quality of current AIGC images, IQA models trained based on degradation injection levels tend to yield spuriously high scores, making it difficult to accurately assess AIGC images. To address this issue, the common practice is to construct dedicated datasets of AIGC images with human-annotated quality scores and train IQA models on them. Nevertheless, the evaluation metrics for AIGC generated images are inherently multi-dimensional. Relying solely on Supervised Fine Tuning and Reinforcement Learning for model training often leads to conflicting training gradients across different dimensions, preventing the model from achieving optimal performance. To address this, we propose Partial Complementary Policy Distillation. This approach partitions the multiple dimensions into two mutually overlapping subsets. By leveraging tool calling and skill self-evolution, we guide the model to induce solutions for resolving these dimensional conflicts, ultimately internalizing these solutions into the model parameters. Extensive experiments demonstrate that our model achieves state-of-the-art (SOTA) performance on AIGC image evaluation tasks compared to other IQA models, effectively resolving the issue of dimensional conflicts.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large Language Model, Image Quality Assessment

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_JihHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Partial Complementary Policy Distillation (PCPD) for evaluating AI-generated images across aesthetics, lighting, shadow, and realism. It trains two experts on overlapping subsets of these dimensions, separating aesthetic and shadow supervision while retaining lighting and realism in both. The shadow expert uses visual tools and self-generated checking rules during training. After a tool-use internalization step, the experts exchange their complementary capabilities through dimension-masked mutual on-policy distillation. A final joint optimization stage produces a single evaluator that does not require external tools. The authors also introduce AIGC-IQA-Bench, containing 10,000 images with human ratings, and report improved correlations with human judgments on this dataset, AGIQA, and PIPAL.

Critical strengths

The paper studies a useful distinction between visual appeal and local physical plausibility. Table 4 shows that joint training can reduce aesthetic and shadow performance relative to separate training, which gives the proposed specialization strategy an empirical motivation. The four-dimensional annotations in AIGC-IQA-Bench could also provide a useful resource for studying errors that an overall quality score would obscure. The reported results show an improvement in shadow assessment without sacrificing aesthetic performance. In Table 2, PCPD raises shadow PLCC from Q-Probe’s 0.683 to 0.821, while aesthetic PLCC increases from 0.848 to 0.866. Table 6 also shows complementary strengths in the two experts and improved performance in the final unified model. These results support the practical value of the training pipeline, even though they do not fully establish its proposed mechanism.

Critical weaknesses

The evidence for resolving gradient conflict is weaker than the paper’s central claim. Table 1 diagnoses conflict at selected points during standard joint training, but the paper does not provide comparable measurements after PCPD or during the final four-dimensional GRPO stage. Restricting distillation losses to selected output tokens does not, by itself, establish that gradients in shared parameters no longer interfere. The performance gains are encouraging, but they cannot distinguish conflict reduction from other benefits of specialization and additional supervision. This matters because resolving dimensional conflict is presented as the main methodological contribution. Although the appendix reports further distillation ablations, the current evidence does not establish how much of the gain requires the proposed complementary distillation scheme rather than the broader, more expensive training pipeline. This leaves the added value of the central design uncertain.

Questions for authors

1.Could the authors provide gradient-conflict measurements comparable to Table 1 before and after PCPD, including during the final joint GRPO stage? Does interference between aesthetics and shadow remain lower when all four rewards are optimized together? This would help determine whether the proposed mechanism explains the observed gains. 2.Which baselines were trained on the same images and human annotations as PCPD, and how were their training budgets determined? Is there a comparison with joint training or a simpler distillation pipeline under a comparable compute budget? Such a comparison would materially affect my assessment of the added value of complementary distillation. 3.In Section 3.2, does masking tool calls and returned observations mean removing them from the model’s input context, or only excluding them from the SFT loss? Please clarify the conditioning information used in this stage and report Model α’s shadow performance immediately before and after conversion to tool-free inference. This would help assess whether the capability needed by the subsequent distillation stage is preserved.

Reviewer_zVrJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses multi-dimensional AIGC image quality assessment (IQA) across four dimensions: aesthetics, lighting, shadow, and realism. The authors identify a training conflict between global aesthetic priors and local physical consistency (shadow), which they characterize as “Halo Effect” and “Masking Effect” (local-to-global attention masking).

Critical strengths
  1. The problem formulation is well-motivated and non-trivial.

  2. The training pipeline is thoughtfully designed with deployment efficiency in mind.

Critical weaknesses
  1. The paper’s entire method rests on the premise that aesthetics and shadow are fundamentally irreconcilable in joint training. However, the evidence provided (Table 1) is limited to a single training run on a single dataset, with no statistical significance testing, no multiple seeds, and no cross-dataset validation. Furthermore, the paper does not justify why this specific pair (Aesthetics vs. Shadow) is uniquely conflicting, rather than other plausible pairs (e.g., Realism vs. Aesthetics, or Lighting vs. Shadow). Without broader evidence, the claimed “fundamental” conflict may be an artifact of the specific dataset or training configuration.

  2. The primary results are reported on AIGC-IQA-Bench, which is constructed by the authors. While the annotation protocol is described in detail, the benchmark itself is not publicly released, and no comparison is made against existing multi-dimensional AIGC IQA benchmarks (e.g., AGIQA-3K, AIGIQA-20K, SA-BENCH) on a per-dimension basis. The external benchmarks (AGIQA, PIPAL) are only reported for Overall scores, which does not demonstrate the method’s claimed advantage in fine-grained dimensional evaluation. The PIPAL dataset is also not designed for AIGC-specific shadow physics, raising questions about its suitability as an external validation set.

  3. The paper does not provide a rigorous formulation of what constitutes an optimal dimensional split, why 3+3 with overlapping items is preferred over other configurations (e.g., 2+2, four independent experts, or a learned split), or how the overlapping dimensions (Lighting, Realism) avoid re-introducing gradient conflicts during distillation. The method appears to be a well-engineered heuristic rather than a principled framework. The paper would benefit from either a theoretical analysis of why partial overlap is optimal, or a systematic search over splitting strategies.

  4. While the unified model achieves strong Overall scores, the per-dimension results are less convincing. The paper claims “SOTA multi-dimensional evaluation performance,” but the evidence is primarily on a private benchmark with a limited set of baselines. A more comprehensive comparison with recent RL-based IQA methods (e.g., Q-Insight, Reasoning-as-Representation) is needed.

Questions for authors

Please refer to Weaknesses.

1 comment
AuthorsOpenReview ↗

Rebuttal Revision Uploaded.

Dear Reviewer,

We have responded to all the questions raised in the rebuttal, revised the PDF, and submitted a Rebuttal Revision version of the PDF. The newly added content is marked in blue. If the paper is fortunately accepted, we will remove the blue markings. If unfortunately it is not accepted, we will incorporate the reviewers’ comments in our subsequent submission.

If you have any other questions, please feel free to raise them. However, due to time constraints, we would appreciate it if they could be raised as soon as possible. If there are no further questions, we sincerely hope that the reviewers can consider raising the score. Thank you very much.

Reviewer_iRyiOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Summary

1、The paper points out the conflicts between different tasks in multi-dimensional image quality assessment.

2、The authors build their own AIGC-IQA dataset and benchmark.

3、They propose PCPD and Self-evolving Skills to address the conflicts between different tasks.

4、Several comparisons and ablation studies are conducted to evaluate the proposed method.

Critical strengths

The motivation is reasonable. Gradient conflicts between different tasks are a real problem in multi-task learning, and it makes sense to study how to reduce such conflicts in multi-dimensional image quality assessment.

Critical weaknesses

W1. Limited novelty. I don't see much novelty in the proposed method. Most of the components are existing techniques. The overall approach feels more like a combination of existing methods (A+B) and engineering tricks, rather than a new technical contribution.

W2. poor presentation. The paper is not well written. The connections between the motivation and different parts of the method are often unclear.

W3. It is unclear how the proposed method actually addresses gradient conflicts, which is the main problem addressed in this paper but lack theoretical prove

W4. The ablation studies for PCPD are not sufficient. The PCPD pipeline involves quite a few steps, but the necessity of several designs is not well justified. For example, a much simpler alternative would be to train the two experts separately, distill only the Shadow capability from model α to model β, and then apply the same final joint GRPO training. Would this achieve similar results?

W5. The motivation and effectiveness of Self-evolving Skills are not convincing. The authors attribute the model's prediction errors to cognitive biases such as the Halo Effect, and then introduce LLM-based reflection to generate skills that supposedly correct these biases. However, there is little evidence that these errors are actually caused by such cognitive biases in the first place. The motivation feels somewhat speculative.

W6. The experimental comparisons is not be fair. The paper does not clearly state whether the competing methods were retrained or fine-tuned on the proposed AIGC-IQA-Bench. The performance gap could largely come from domain adaptation rather than the proposed method itself.

W7. The design of the new benchmark is not well justified. The authors introduce four evaluation dimensions: aesthetics, lighting, shadow, and realism. However, it is unclear why these four dimensions were chosen and whether they are sufficient to represent AIGC image quality.

Questions for authors

see weaknesses

3 comments
AuthorsOpenReview ↗

Rebuttal Revision Uploaded.

Dear Reviewer,

We have responded to all the questions raised in the rebuttal, revised the PDF, and submitted a Rebuttal Revision version of the PDF. The newly added content is marked in blue. If the paper is fortunately accepted, we will remove the blue markings. If unfortunately it is not accepted, we will incorporate the reviewers’ comments in our subsequent submission.

If you have any other questions, please feel free to raise them. However, due to time constraints, we would appreciate it if they could be raised as soon as possible. If there are no further questions, we sincerely hope that the reviewers can consider raising the score. Thank you very much.

Thanks for the response, but I still feel the paper is below iclr acceptance bar. The novelty remains limited, and the response doesn't clearly explain how the method differs from existing approaches. I'm also concerned about the fairness of the comparisons, as the baselines were originally designed for different tasks, and it's unclear how they were retrained. And I still don't see how denser supervision from OPD can resolves gradient conflicts, which is the main motivation of the paper.

AuthorsOpenReview ↗

We do not propose fundamental improvements to GRPO or OPD but we want to clarify our work's value.

1. Novelty and Difference from Existing Approaches

We understand that the reviewers still have concerns about novelty, and we understand that ICLR reviewers may prefer to see papers that improve the OPD mechanism. However, the primary area we selected is applications to computer vision, audio, language, and other modalities. In fact, our research has substantial practical value, especially in major companies such as Douyin, Meituan, JD, Tencent, and Alibaba, where AIGC image generation is widely used. Among these applications, the use of VLM-assisted AIGC image generation models is also widespread. However, the capabilities of base image generation models are constantly improving and are actually extremely powerful. Only certain specific business scenarios require certain specific dimensions, so post-training is widespread. Yet these dimensions may conflict even with more elaborate data design. Therefore, our method emerges to address this problem. Solving dimension conflicts through complementary distillation is certainly first proposed by us and has general value. When other businesses require different dimensions, one can also identify the conflicting dimensions, disentangle them, and apply complementary distillation. Of course, we do acknowledge that we do not propose fundamental improvements to GRPO or OPD.

2. Fairness of Comparisons and Baseline Retraining

The fairness of our method is indeed ensured. In fact, this is precisely where our motivation comes from. For all MLLM-based methods, we have consistently trained and tested them on business data. This was our initial work. We did find that traditional approaches struggle to achieve good results on the four dimensions we require. Although these methods may perform well when trained on any single dimension sampled arbitrarily, as long as we want the model to possess all four capabilities simultaneously, the training results are always unsatisfactory.

3. How Denser Supervision from OPD Resolves Gradient Conflicts

Let me present a scenario: the shadow break is perfectly executed and lies flat on the sofa, and its structure and break are highly consistent with physical laws. This matters. From the shadow perspective, it deserves a high score of 4.7/5. Aesthetically, however, the presence of the shadow undermines the harmony between the main subject of the indoor scene—the sofa—and the overall background, and should receive 3.5/5. During training, it often happens that when shadow quality is high, the model develops a bias that shadows do not affect aesthetic quality. Under sparse rewards, the model readily biases toward learning the shadow dimension, while for the aesthetic dimension the model enters a confused state. This is what we mean by the model continuously fitting toward accuracy in the shadow dimension; it may abandon the aesthetic direction or learn it as a secondary direction. This is especially likely in the early stage of training. However, using dense rewards such as OPD forces the model to treat the reward signals from both dimensions equally, thereby blocking this behavior.

Comments on the paper

1 comment
AuthorsOpenReview ↗

As the deadline is approaching, I would like to make a final clarification of the two most important points.

Dear Area Chair and Reviewers,

Thank you for your constructive feedback. After the rebuttal, I believe I have clarified most of the reviewers’ concerns. Of course, one reviewer has not yet responded. As the deadline is approaching, I would like to make a final clarification of the core motivation and value of this paper, as well as the key point that the dimension partitioning is not heuristic but is based on experimental analysis and can ultimately be validated experimentally.

1. Clarification of Novelty, Practical Value, and Motivation

We understand the reviewer’s concern about novelty and that ICLR reviewers may prefer to see papers that improve the OPD mechanism. However, the primary area we selected is applications to computer vision, audio, language, and other modalities, and our work has substantial practical value, especially in major companies such as Douyin, Meituan, JD, Tencent, and Alibaba, where AIGC image generation and VLM-assisted AIGC image generation models are widely used. Although base image generation models are constantly improving and are already extremely powerful, certain specific business scenarios still require particular dimensions, so post-training is widespread; yet these dimensions may conflict even with more elaborate data design. Therefore, our method emerges to address this problem. Solving dimension conflicts through complementary distillation is, to our knowledge, first proposed by us and has general value: when other businesses require different dimensions, they can also identify the conflicting dimensions, disentangle them, and apply complementary distillation.

We acknowledge that we do not propose fundamental improvements to GRPO or OPD, and we thank the reviewer for this comment; we would like to clarify the novelty of our work and restate the connection between our motivation and our method. In indoor smart-home scenarios, the four most important dimensions are realism, illumination, shadow, and aesthetics. Through various experiments, we found that training shadow and aesthetics simultaneously causes mutual interference, and our analysis shows that these two dimensions conflict with each other. We therefore adopt a complementary distillation approach: non-conflicting dimensions are treated as ordinary dimensions, while conflicting dimensions are decomposed and mutually distilled in a complementary manner. In practice, this greatly alleviates gradient conflicts and ensures that the two models do not deviate too far from each other, which would otherwise make distillation too difficult to learn. Complementary distillation is therefore not merely an A+B combination of existing techniques; it is the core technical contribution of this paper.

2. Direct Evidence and Ablation Support for the Aesthetics–Shadow Conflict

The claim that Aesthetics and Shadow conflict is not based on a single training run or a single diagnostic. We ran the experiments multiple times and averaged the test data, and we added extensive ablations on gradient conflicts. The conflict is directly measured by gradient diagnostics, not assumed:

Training ConfigurationMean Aes–Shadow Gradient CosineMinimum CosineProportion of Negative-Similarity StepsNormRatio DeviationLoss Drop on One Side Caused by Unilateral UpdateConflict Trigger Frequency D>δAesthetics PLCCShadow PLCC
Standard SFT+GRPO joint training-0.52-0.8160%3.40.42High0.770.68
Unified model after PCPD mutual distillation-0.08-0.2113%1.10.06Low0.810.84
Final unified GRPO with PCPD+0.05-0.127%1.10.06Very low0.850.88
PCPD without token mask-0.61-0.8466%4.20.45High0.710.62
Off-policy instead of on-policy distillation-0.35-0.6742%2.80.22Medium-high0.850.73

The Aesthetics–Shadow pair is not arbitrarily selected. In indoor smart-home scenarios, the four practically relevant dimensions are Aesthetics, Lighting, Shadow, and Realism. Gradient diagnostics show that Aesthetics and Shadow are the main conflicting pair, while Lighting and Realism are non-conflicting shared dimensions. Other plausible pairs do not exhibit comparable negative gradient cosine or negative-similarity steps. The conflict is therefore a property of the multi-dimensional scoring structure under joint GRPO, not an artifact of one dataset or one configuration. The single-direction distillation ablation further supports this: distilling only Shadow α→β or only Aesthetics β→α before joint GRPO degrades performance relative to PCPD, so the conflict is real and complementary mutual distillation is necessary.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.