acceptodds
Withdrawn from ICLR 2027

SmartPhotoCrafter: Unified Reasoning, Generation and Optimization for Automatic Photographic Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Automatic photographic enhancement requires more than executing predefined editing instructions. A model must first infer whether an image should be edited, determine what needs to be changed, and estimate the appropriate strength of each adjustment. This problem becomes particularly challenging in real photographs, where technical degradations and aesthetic deficiencies often co-occur, while successful edits must preserve the original content and photographic intent. We introduce SmartPhotoCrafter, a unified framework for automatic photographic optimization across restoration, retouching, and their compositions. SmartPhotoCrafter consists of an Image Critic that diagnoses image quality and infers editing decisions, and a Photographic Artist that realizes these decisions through conditional generation. Instead of relying on a text-only cascade, the Artist is conditioned directly on the Critic's contextual and reasoning representations, providing a richer interface between visual assessment and image generation. We further align reasoning and generation through executable editing targets and outcome-aligned reinforcement learning, with task-specific signals that evaluate editing direction, photometric agreement, perceptual quality, and content preservation. Experiments on automatic optimization, mixed restoration–retouching, and conventional restoration tasks demonstrate the effectiveness of the proposed method.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Automatic Image Enhancement, Image Editing, Multimodal Learning

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

This paper proposes SmartPhotoCrafter, a framework for automatic photographic enhancement that unifies image restoration and retouching. It combines an MLLM-based Image Critic for quality assessment and editing decisions with a generative Photographic Artist for image enhancement. The authors introduce a latent decision interface to transfer the Critic's representations to the Artist and employ outcome-aligned reinforcement learning to jointly improve editing decisions and generated results. Experiments demonstrate competitive performance on automatic enhancement and restoration benchmarks. However, concerns remain regarding computational efficiency, baseline coverage, the generalizability of the proposed interface, and the evaluation of automatic editing decisions.

Critical strengths

1.The paper addresses a meaningful problem by considering image restoration and aesthetic retouching within a unified automatic enhancement framework. The motivation is clear and practically relevant. 2.The latent decision interface and executable-edit grounding provide a reasonable approach to connecting image assessment with generation. The outcome-aligned rewards are also designed to account for editing direction, photometric improvement, and perceptual consistency. 3.The paper includes quantitative comparisons, ablation studies, and human evaluations. The results demonstrate promising reference consistency and competitive enhancement quality.

Critical weaknesses
  1. Qwen-Image-Edit already employs Qwen2.5-VL for visual semantic conditioning, while the proposed Critic is also initialized from Qwen2.5-VL. This architectural relationship may contribute to the effectiveness of transferring hidden representations. Therefore, the observed improvement does not necessarily establish that the proposed interface generalizes across independently developed models. The authors should clarify how this interface differs from the existing conditioning mechanism and provide evidence of its transferability, ideally through a targeted experiment involving a different VLM or generator backbone.
  2. The framework introduces an additional 7B-parameter VLM and a multi-module inference pipeline, substantially increasing system complexity. However, the improvements are not consistently significant. The paper does not report inference latency, peak GPU memory, or computational cost. These comparisons are necessary to assess whether the additional complexity is justified by the performance gains.
  3. The paper lacks comparisons with recent agentic image restoration methods, such as AgenticIR and 4KAgent, as well as VLM-guided approaches such as FAPE-IR. These methods are closely related to the proposed framework, and comparisons on applicable tasks would better establish its effectiveness. Furthermore, although the Image Critic is trained using MOS annotations, its scoring accuracy is not independently evaluated against established IQA methods. More importantly, the ranking reward in Eq. (10) relies on the Critic's own predicted scores rather than independent quality judgments. Without evaluating scoring accuracy before and after reinforcement learning, it remains unclear whether this feedback improves quality assessment or reinforces existing prediction biases.
  4. A central claim of the paper is that the model can autonomously determine whether an image requires editing and select appropriate adjustments. However, the automatic enhancement benchmark primarily contains images with predefined restoration or retouching targets. The experiments do not sufficiently evaluate whether the model can recognize images that require little or no modification. A small-scale evaluation on already high-quality photographs would help verify whether the Critic can avoid unnecessary edits and preserve intentional photographic styles.
Questions for authors
  1. Given that Qwen-Image-Edit already incorporates Qwen2.5-VL for visual semantic conditioning, can the authors provide evidence that the proposed latent decision interface generalizes to different VLM or generator backbones?
  2. Could the authors report the inference latency and peak GPU memory usage of SmartPhotoCrafter compared with representative baselines?
  3. How does SmartPhotoCrafter compare with recent agentic and VLM-guided image restoration methods, such as AgenticIR, 4KAgent, and FAPE-IR? Additionally, could the authors evaluate the Critic's scoring accuracy against established IQA methods and clarify whether the self-referential ranking reward in Eq. (10) improves scoring reliability?
  4. How does SmartPhotoCrafter handle images that are already of high photographic quality? Can the authors provide quantitative or qualitative evidence that the Critic can recognize when no editing is necessary and avoid unnecessary modifications to intentional photographic styles?
Reviewer_5ahbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SmartPhotoCrafter, a unified framework that seamlessly integrates image restoration and retouching. The architecture is driven by two synergistic modules: an Image Critic that evaluates quality and formulates edit strategies, and a Photographic Artist that synthesizes enhanced images based on the Critic's latent representations. Crucially, the method transcends traditional text-only pipelines by allowing the Artist to directly process the Critic's hidden states. To ensure high-quality outputs, the framework incorporates executable-edit grounding and outcome-aligned reinforcement learning, effectively aligning the system with photometric improvement, perceptual consistency, and content preservation. The proposed approach is rigorously validated across diverse scenarios, including automatic enhancement, mixed editing tasks, and standard restoration benchmarks.

Critical strengths

1.The Critic-Artist decomposition is intuitive, and the latent conditioning approach is more innovative than a plain text-based pipeline. Moreover, the comparison between textual cascading and latent conditioning provides strong empirical support for the claimed design choices.

2.The method shows competitive or better performance on the included metrics, especially on the multi-edit setting and some restoration tasks.

Critical weaknesses

1.It is difficult to determine how much of the performance improvement stems from the latent interface itself, as opposed to the use of significantly stronger base models, extensive supervision, or reward engineering. For instance, Table 2 lacks a more detailed ablation study on different reward designs.

2.A substantial portion of the training signal originates from generated annotations and rule-executed edits, which may compromise realism and introduce bias. For example, the Critic model employed in the paper is merely based on Qwen-2.5-7B, which struggles to make precise evaluations of fine-grained image information. Therefore, it is necessary to conduct comparative experiments with other superior feedback models to validate the robustness of the proposed method.

3.The paper briefly presents a few case studies, but it lacks a systematic analysis of when the Critic makes erroneous judgments or when the Artist follows suboptimal suggestions.

4.The paper claims that the Artist benefits from the Critic's latent states, but it remains unclear what specific information is actually transferred and the underlying mechanism of why it works. Further analysis and explanation are required, including the theoretical rationale for the superiority of the latent representations and visualizations of these latent states across different samples.

5.The feedback branch in the model inevitably increases the training time for reinforcement learning. However, the paper fails to provide a comparative analysis of the additional time latency or computational overhead introduced by this design.

Questions for authors

Please Refer to Weakness.

Reviewer_4TBHOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents SmartPhotoCrafter to achieve automatic photographic enhancement that combines image restoration and retouching. It consists of an Image Critic, which assesses image quality and outputs suggested edits, and an Artist that generates the enhanced image conditioned on the outputs of the Critic, which are represented in hidden states. The training process consists of supervised adaptation, interface alignment, and reinforcement learning using photographic and perceptual rewards derived from reference images.

Critical strengths

It addresses a meaningful problem of automatic photo enhancement, combining the critic model and generative model, making them work synergistically and improve each other.

Critical weaknesses
  1. Incremental novelty and insufficient evidence to prove its effectiveness. The framework consists of the critic model and the generative model; each is a pre-trained backbone: QWen2.5-VL-7B and QWen-Image-Edit. The method mainly combines them to achieve judgment and enhancement simultaneously. However, current experimental results are insufficient to support the central claim that combining the critic and artist models could achieve better performance than the artist alone. An important baseline comparison should be added, which is to train a model with the same size and settings, including training strategies, losses, RL strategies, and training datasets as the current combined one. This could help to clarify whether the current performance gain mainly comes from the critic-artist combination or the end-to-end pairwise training pipeline plus the RL training objectives with references. Since this central contribution and design motivation remain unclear, the current framework seems incremental and less convincing.

  2. Unfair comparison with baselines and lack of comparisons. The proposed framework uses the pairwise training data during both the adaptation training and the RL training, but it is unclear whether the baseline methods are trained on the same pairwise datasets. Some baseline methods, for example, InstructPix2Pix, are not designed for enhancement purposes; directly comparing with it without further fine-tuning is unfair. Furthermore, comparison results with other supervised methods should be reported, such as photo retouching methods, including 3D-LUT-based methods, CSRNet, RSFNet, and other super-resolution or dehaze baselines.

Given the insufficient experiments and unfair comparison, I lean toward rejection.

Questions for authors

NA

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

Related papers

Discussion (0)

Sign in to comment.