acceptodds
Under review as a conference paper at ICLR 2027

Aes6D: A Dataset and Benchmark for Aesthetic-Guided 6-DoF Image Editing

Abstract

In real-world photography, the subject and scene content are often determined by the shooting environment and creative intent, creating a need to improve image composition and visual presentation through fine-grained viewpoint adjustment while preserving the original content. To this end, we introduce Aesthetic-Guided 6-DoF Image Editing. The task presents two key challenges: composition optimization typically requires achieving substantial improvement through small viewpoint changes within a limited camera-motion range, while real viewpoint adjustment involves both 3D translation and 3D rotation and thus requires full 6-DoF modeling. However, there is no dataset that couples fine-grained 6-DoF camera changes with substantial composition improvement. Aesthetic priors are particularly valuable for this task, as they indicate which compositional factors most affect visual quality and how they should be adjusted. Reliable supervision should therefore use aesthetic priors to guide the corresponding DoFs and motion directions, while covering the full 6-DoF space of translations, rotations, and their combinations. Based on this insight, we construct Aes6D, a dataset and benchmark for this task. Aes6D covers nine composition patterns, provides complete 6-DoF supervision with both Single and Multi samples, and uses aesthetic guidance to make small viewpoint changes consistently produce substantial and compositionally meaningful differences. Building on Aes6D, we further conduct an initial exploration of structured 6-DoF learning with DoF-GRPO, which introduces per-DoF geometric rewards, field-wise relative advantages, and token-level credit assignment to better align GRPO with the geometric structure of 6-DoF camera parameters. Experiments show that models trained on Aes6D achieve substantial composition improvement while preserving scene content, outperforming existing composition and image editing methods and validating Aes6D for this task.

Primary area:
datasets and benchmarks
Keywords:
Aesthetic-Guided Image Editing, Image Composition, 6-DoF Camera Control

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_zkPNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Aesthetic-Guided 6-DoF Image Editing, a novel framework that improves photographic composition through small camera translations and rotations while aiming to preserve scene content. To overcome the absence of such a dataset for direct supervision, the authors construct Aes6D, a dataset and benchmark containing 26,834 samples across nine composition patterns. Starting from high-quality photographs, the authors apply composition-aware camera perturbations and view synthesis to generate degraded views, using the inverse motions as corrective supervision. The dataset includes both single-DoF and multi-DoF adjustments, composition reasoning, and quality filtering. On top of the dataset, the authors further explore a two-stage training approach combining supervised fine-tuning with DoF-GRPO, which uses per-DoF geometric rewards, field-wise relative advantages, and token-level credit assignment to refine camera-motion predictions. The predicted adjustments are applied through a view-synthesis pipeline to produce edited images. Experiments show that the resulting method improves composition while largely preserving scene content, outperforming the evaluated composition and image editing baselines.

Critical strengths
  1. The pipeline of constructing a dataset is well designed. Instead of applying random perturbations to good photos, the framework intentionally degrades composition principles to create bad views, using the reverse motion as the ground truth. Its greatest strength lies in establishing a supervisory signal that bridges minor camera adjustments with meaningful compositional changes.
  2. The framework that converts the abstract goal of making a photo more aesthetically pleasing into explicit translation and rotation parameters are novel and convincing. By addressing both single and multi-axis adjustments while providing compositional reasoning, the system makes the direction of improvement easy to interpret and enables potential integration with real-world camera controls.
  3. Their proposed objectives are well validated by ablation studies. Applying DoF-GRPO shows clear improvements over standard SFT and conventional GRPO.
Critical weaknesses
  1. The evidence for generalization in composition improvement remains limited. On the 50 OOD images in Table 3, the human-evaluated win rate against the input view is 56% for the proposed method and 54% for PhotoFramer. A larger and more diverse OOD evaluation, together with confidence intervals, would help establish whether this small advantage is reliable. The paper should also clarify the annotators’ expertise and evaluation instructions and report inter-annotator agreement.
  2. Since the authors construct ground-truth camera movements, I wonder about the accuracy of the trained MLLM's predicted camera parameters. Providing a quantitative evaluation of these parameter predictions would make the paper significantly more convincing.
  3. The final image generation process is not explained with sufficient detail to ensure reproducibility. While the main text briefly notes that the predicted camera movement is applied to a view-synthesis pipeline, the details provided in Appendix C.4 is related to the data generation stage. The paper should explicitly specify how the reference image and model inputs are configured at inference time, and clarify how this inference workflow differs from the data generation process.
Questions for authors
  1. Could the authors report per-DoF prediction errors against the ground-truth camera motions to directly assess the accuracy of the learned camera control?
  2. Could the authors specify the reference images and model inputs used during inference and clarify any differences from the dataset-construction pipeline described in Appendix C.4?
2 comments

Thank you for answering my questions and sorry for late reply. The expanded human evaluation provides stronger evidence that the method improves composition while preserving scene content on images amenable to viewpoint correction. The per-DoF results on 1,000 examples also address my question about camera-control accuracy, and the clarification of the inference inputs resolves my concern about the editing pipeline. Although I am keeping my score at Weak Accept, I now support acceptance more strongly.

Thank you very much for taking the time to get back to us, we truly appreciate it. We are glad to hear that our response have addressed your concerns and strengthened your support for our work. Thank you again for your thoughtful feedback and constructive suggestions, which have helped us improve the paper.

Reviewer_WvMxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Aesthetic-Guided 6-DoF Image Editing, a task that improves photographic composition through camera translation and rotation while preserving scene content. To support the task, the authors construct Aes6D by applying composition-guided camera motions to high-quality Good Views to create Bad Views. The inverse motions provide 6-DoF supervision, alongside generated composition reasoning. The paper also proposes DoF-GRPO, which uses per-DoF geometric rewards and token-level credit assignment to refine motion predictions. Human and model-based evaluations indicate that the resulting edited images are preferred over several baselines on the reported benchmark subset.

Critical strengths
  1. The data construction is closely aligned with the task objective. Aes6D designs camera motions based on composition categories to create Good–Bad view pairs with known motion parameters, thereby explicitly linking composition problems to camera adjustments.
  2. The final image quality is partially validated by human evaluation. The proposed method achieves an 83% human win rate over Bad Views on approximately 100 benchmark samples, outperforming the listed baselines. Ablation studies also support the benefits of combining Single/Multi samples and using DoF-GRPO.
Critical weaknesses
  1. Limited completeness of the proposed work. Achieving the final editing result requires combining the proposed approach with existing methods, which reduces its completeness as a standalone contribution.
  2. Insufficient evidence for precise 6-DoF control and generalization. The experiments mainly assess image preference and scene consistency, without reporting per-DoF prediction errors or direction accuracy. The 50-image out-of-distribution evaluation is also limited: the human win rate is 56%, compared with 54% for the strongest baseline, with no uncertainty estimates.
Questions for authors
  1. How were the nine composition categories and spatial attribute labels assigned? Please report the sample count for each category, any human validation or inter-annotator agreement, and the retention rate after quality filtering.
  2. Were the training set and benchmark split by original Good View? Are all Single and Multi samples derived from the same photograph kept in one split, and were near-duplicate photographs across data sources checked?
  3. Could you report per-DoF parameter errors and motion-direction accuracy on the full 1,000-sample benchmark for SFT, standard GRPO, and DoF-GRPO? How do these measures relate to human preferences for the final images?
  4. How was the 10% subset used for human evaluation selected? How many annotators judged each comparison, and how were disagreements and ties handled? Please provide confidence intervals for the main benchmark and the 50-image OOD results.
Reviewer_oFrPOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Aesthetic-Guided 6-DoF Image Editing. Given a photo, the model predicts a fine-grained viewpoint adjustment in the full 6-DoF camera-motion space so that the composition improves, while the subject, scene structure, and semantic content stay largely unchanged. To support this task, the paper builds Aes6D, a dataset and benchmark that covers nine composition patterns and provides full 6-DoF supervision through Single and Multi samples. It then explores structured 6-DoF learning with DoF-GRPO and experiments shows their effectiveness

Critical strengths
  1. The task is interesting. Early image editing mostly changes the subject, the scene content, or the appearance, and it rarely corrects composition. This paper focuses on a real photographic need: improving composition and visual presentation through small viewpoint changes, without substantially changing the original content.

  2. The data construction is reasonable. Aes6D starts from high-quality Good Views, turns their composition principles into directional 6-DoF camera motions, and renders Bad Views as the model input.

  3. The method is also reasonable. The paper uses GRPO to learn structured 6-DoF camera control. DoF-GRPO introduces per-DoF geometric rewards, field-wise relative advantages, and token-level credit assignment, so the training signal follows the geometric structure of the 6-DoF camera parameters instead of one shared sequence-level score.

Critical weaknesses
  1. It remains unclear how exactly a "good composition" should be measured. This paper primarily employs an aesthetic evaluation model to determine whether a "good view" is superior to a "bad" one. However, a higher aesthetic score does not, in itself, prove that the predicted 6-DoF motion is correct. Aesthetic preference and the quality of camera adjustments are distinct concepts, yet the current evaluation approach fails to differentiate between them.

  2. Furthermore, the paper does not justify why all six DOF must be jointly optimized within a single GRPO model. It might be equally feasible to train separate models for each DOF and subsequently merge their weights.

Questions for authors

N/A

2 comments

Thank the author for the author for the reply. The author has addressed all the concerns I raised, and I have no further comments. I will maintain my original rating.

Thank you for your positive feedback and thoughtful review. We are glad that our response and additional experiments have addressed your main concern. We sincerely appreciate your time and constructive suggestions, which have helped us improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.