- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces PASEdit, a framework consisting of a planner and an executor for portrait aesthetic structural editing. Given an input portrait and a coarse request, the planner generates a concrete instruction targeting a single factor, namely pose, viewpoint, or framing. The executor then edits the image while preserving subject identity and non-target content. The authors propose Aesthetic Planning Post-Training (AesPT), which first trains three specialized planners with task-specific rewards and then consolidates them into a unified planner through multi-teacher on-policy distillation. The executor is trained with paired supervision and further optimized using DiffusionNFT, with rewards for identity preservation, non-target content preservation, and instruction following. The authors also construct a dataset of 12,422 controlled editing pairs based on failure patterns extracted from real photographic examples. Experiments include comparisons on the controlled benchmark, training-stage ablations, comparisons on real portraits, and a user study. The results show improvements in structural editing, although generalization to real-world images, the core reward design, and strict single-factor control require further validation.
Critical strengths
-
Addressing structural problems in portraits is a useful and practically relevant research direction. Compared with exposure adjustment, color correction, or skin retouching, editing pose, viewpoint, and framing requires more substantial changes. The paper considers both planning from coarse requests and preserving content during editing.
-
The image-specificity reward is an interesting design. Comparing the value of the same instruction on a degraded image and its high-quality target provides a way to distinguish targeted corrections from generic photographic advice.
Critical weaknesses
-
The methodological contribution over existing work needs to be clarified. The decomposition into planning and execution has already been explored in AesFormer, the expert training and consolidation procedure follows the general MOPD paradigm, and executor optimization uses DiffusionNFT. The main contributions therefore appear to lie in task adaptation, data construction, and reward design. It is not yet clear what generalizable methodological insights these choices provide.
-
Most of the quantitative evidence comes from a controlled test set that shares its construction procedure with the training data. Splitting by identity reduces subject overlap, but does not rule out adaptation to the 22 degradation rules or patterns introduced by the image generator. The real-image experiments are a useful addition, but the quantitative comparison on 200 real portraits only includes AesFormer, and the user study covers 20 examples. It remains unclear whether the large gains in Table 1 generalize to a broader range of real portraits.
-
The core reward design lacks sufficient quantitative ablation. The image-specificity reward Gt is an important component, but its effect is mainly illustrated through the qualitative examples in Figure 6. The paper does not systematically evaluate the removal of Gt, the reference-guidance term At, or the pose-sufficiency term BP. The effects of reward weights and probability-weighted scoring are also not analyzed, making it difficult to assess the necessity of these choices.
-
The reliability of the automatic evaluation needs further validation. Using different models for training rewards and final evaluation is reasonable, but agreement with human judgments should also be established. CP combines identity preservation and non-target content preservation, which may hide differences between them.
-
Some important details for reproducibility and understanding the method’s limitations are missing. The proprietary model used to generate degraded images is not identified, and the complete degradation rules, judge prompts, scoring anchors, and filtering acceptance rates are not fully provided. The paper also emphasizes single-factor control without independently quantifying changes to non-target structural factors. More failure analysis on multiple subjects, occlusions, complex interactions, and images that do not require editing would help clarify its applicability.
Questions for authors
-
Could the authors provide quantitative ablations of Gt, At, BP, and probability-weighted scoring, with results reported separately for each task? Which component contributes most to the final performance?
-
Which model is used to generate the degraded images? Could the authors release the complete set of 22 rules, generation and evaluation prompts, and statistics on automatic filtering?
-
Are the 200 real test portraits and the 20 user-study examples disjoint from the entire training image pool, including subject identities? The paper explicitly states that they are disjoint from the 1,000 examples used for rule extraction, but the broader exclusion criteria are less clear.
-
How are intended changes distinguished from unnecessary changes to other factors? For example, a viewpoint edit may also change subject scale, visible background, and head orientation.