acceptodds
Withdrawn from ICLR 2027

PASEdit: Portrait Aesthetic Structural Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative image editing models enable high-level structural editing of portraits for aesthetic enhancement, such as adjusting pose, viewpoint, and framing. However, existing editors still face two key challenges: deciding how to improve a structurally suboptimal portrait from a coarse request, and making effective structural changes without causing identity drift or unintended changes. We propose PASEdit, a planner-executor framework for controllable single-factor portrait structural editing across pose, viewpoint, and framing. For planning, we introduce Aesthetic Planning Post-Training (AesPT), which trains factor-specific experts with task-specific visual feedback and consolidates their capabilities into a unified planner through multi-teacher on-policy distillation. For execution, the executor first learns structural transformations from paired supervision and is then optimized for identity and non-target content preservation while maintaining instruction adherence. To provide controlled supervision for structural learning, we construct 12,422 portrait editing pairs derived from real photographic degradation patterns. Extensive experiments show that PASEdit achieves more effective structural improvements while maintaining strong identity and non-target content preservation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
portrait aesthetics, structural editing, multimodal planning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AH8VOpenReview ↗
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
  1. 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.

  2. 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
  1. 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.

  2. 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.

  3. 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.

  4. 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.

  5. 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
  1. 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?

  2. 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?

  3. 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.

  4. 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.

Reviewer_C28VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PASEdit, a planner-executor framework for portrait aesthetic structural editing. Unlike conventional portrait enhancement approaches that primarily focus on appearance-level improvements, PASEdit aims to modify three high-level structural factors: human pose, camera viewpoint, and framing, while preserving subject identity and non-target image content.

Critical strengths
  1. The paper addresses a meaningful and challenging practical problem.
  2. The controlled data construction pipeline is carefully designed. The authors extract real photographic failure patterns, summarize them into reusable degradation rules, and construct synthetic editing pairs through controlled degradation.
  3. The paper is generally well organized .The method description, training pipeline, and evaluation protocol are logically arranged.
Critical weaknesses
  1. Although the paper claims to support open-ended requests such as "Improve this photo," the planner is restricted to selecting one of three structural factors: pose, viewpoint, and framing. However, real-world portraits may require simultaneous improvements in multiple structural factors. Moreover, the reported factor-selection accuracy is evaluated on synthetically degraded images, each associated with a predefined structural factor. This setting does not fully demonstrate the planner's ability to identify the most appropriate editing operation for unconstrained real-world portraits.
  2. The main quantitative evaluation is conducted on 600 test images constructed using the same data generation pipeline as the training images. Although there is no identity overlap between the training and test sets, the shared data generation process may favor PASEdit, which has been specifically trained on such data. Furthermore, Table 2 only provides a comparison with AesFormer on real-world portraits. I would like to see comparisons with the major baselines in Table 1 on real-world images.
  3. The Image-Specificity Reward is one of the key contributions of this paper. However, its effectiveness is demonstrated only through qualitative examples in Figure 6, without a corresponding quantitative ablation study.
Questions for authors
  1. In Table 3, both SI and VER decrease when moving from Expert RL to Expert RL + OPD. I wonder whether applying the same preservation-aware executor post-training directly to the Expert RL variant, without OPD, could achieve performance comparable to that of the full PASEdit model.
  2. Will the authors release the complete 22-rule library?
Reviewer_jRZaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

PASEdit addresses portrait editing through a planner–executor pipeline that modifies one structural factor: pose, viewpoint, or framing. The planner is trained through factor-specific reinforcement learning and multi-teacher on-policy distillation. The executor learns from paired images and then undergoes preservation-aware DiffusionNFT training. The authors construct 12,422 editing pairs using 22 degradation rules and report improvements on synthetic and real portraits.

Critical strengths
  • The dataset construction includes degradation validation and identity-disjoint splits. These are useful safeguards for evaluating portrait editing.
  • The expert-versus-joint RL comparison matches the total number of RL updates. Table 3 shows a substantial improvement before executor post-training.
  • Executor post-training improves CP from 3.90 to 4.59 while SI rises from 3.93 to 4.07. The separate final evaluator and human study provide additional evidence beyond training rewards.
Critical weaknesses
  • The image-selectivity reward needs stronger validation. Gt=max⁡(Sd−So,0)G_t=\max(S_d-S_o,0) measures whether advice is more useful for the degraded image than its target. This does not necessarily establish image-specific planning: generic advice such as “crop tighter” can score well when it reverses the applied degradation. Figure 6 alone is insufficient; quantitative comparisons against SdS_d-only feedback and removal of GtG_t are needed.
  • Planning rewards are disconnected from execution outcomes. AesPT scores proposed instructions without executing them, while Appendix A.5 trains the executor using only reference instructions. At inference, it receives planner-generated instructions. The paper should measure whether planner reward predicts actual editing success and compare executor training on reference versus generated plans.
  • Single-factor control is not directly measured. SI measures improvement in the requested factor, while CP combines identity and content preservation. Neither explicitly measures changes to the other structural factors. For example, changing head orientation can resemble a viewpoint edit without changing the camera. The paper needs clearer factor definitions and an audit of unintended pose, viewpoint, and framing changes.
  • The final benefit of the specialization route is not fully isolated. AesPT follows the specialization-and-distillation structure of MOPD, and the executor adopts DiffusionNFT. Table 3 supports expert specialization, but only that route receives the final executor post-training. Comparing both planner routes with the same final executor would show whether the advantage survives preservation optimization.
Questions for authors
  • How much do image selectivity and pose sufficiency contribute individually, compared with scoring improvement on IdI_d alone?
  • How strongly does planner reward correlate with executed SI/CP? Does training the executor on generated plans reduce failures?
  • Can you report results on held-out degradation rules or an independently collected real-portrait set, with image-level confidence intervals?
  • How are viewpoint changes distinguished from subject rotation, and how often do edits alter a non-target structural factor?
  • Does expert RL+OPD still outperform joint RL when both use the same preservation-trained executor?

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

Related papers

Discussion (0)

Sign in to comment.