Pith. sign in

REVIEW 4 major objections 3 minor 7 cited by

PhotoAgent claims that autonomous photo editing can be formulated as a long-horizon planning problem, solved by closed-loop tree search with a learned aesthetic reward, so that a vague "make this image better" prompt yields edits that beat

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 20:34 UTC pith:OT2ALWMP

load-bearing objection A real system and a useful dataset, but the paper's own metric contradicts its headline claim — GPT-4o beats PhotoAgent on the UGC Score. the 4 major comments →

arxiv 2602.22809 v3 pith:OT2ALWMP submitted 2026-02-26 cs.CV

PhotoAgent: Exploratory Visual Aesthetic Planning with Large Vision Models

Mingde Yao , Zhiyuan You , King-Man Tam , Menglu Wang , Tianfan Xue This is my paper
classification cs.CV
keywords autonomous image editingaesthetic planningMonte Carlo tree searchclosed-loop agentreward modelUGC-Edit datasetinstruction-based image editingvision-language model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to show that the bottleneck in instruction-based image editing is not the editor itself but the planning: users should not have to decompose "make my photo better" into precise sequential commands. It proposes PhotoAgent, a closed-loop agent that perceives the image, proposes candidate edits, searches over multi-step edit trajectories with Monte Carlo tree search, executes the best actions, evaluates the outcome, and rolls back when quality drops. The claim, supported by quantitative metrics and a small user study, is that this explore-then-commit loop consistently outperforms single-step editors and simpler agent baselines on both instruction adherence and visual quality. A new benchmark, UGC-Edit, supplies a learned reward model trained on real user photos so the evaluator reflects human aesthetic taste rather than generic image-quality scores.

Core claim

PhotoAgent's central claim is that aesthetic photo editing can be automated by treating it as a sequential decision-making problem rather than a single instruction-following step. The system uses a vision-language perceiver to generate candidate editing actions, an MCTS-based planner to simulate and rank multi-step action sequences in a low-resolution approximation before committing, a heterogeneous executor that routes edits to either procedural operators or generative models, and an ensemble evaluator that only accepts an edit if the measured aesthetic score improves. The same loop runs iteratively with memory, so the agent can build on its own prior edits. The authors also construct UGC-E

What carries the argument

The load-bearing mechanism is the closed-loop perceive-plan-execute-evaluate cycle driven by an MCTS planner. At each round the perceiver proposes candidate atomic edits; MCTS rolls out several future trajectories in reduced resolution, scores them with an ensemble evaluator that includes the learned UGC reward model, and back-propagates the rewards; the top-K actions are executed for real, re-scored at full resolution, and only the highest-scoring improvement is kept, with rollback otherwise. This combination of simulation-based lookahead, full-resolution verification, and rollback is what distinguishes PhotoAgent from open-loop agents and single-step editors.

Load-bearing premise

The load-bearing assumption is that the learned UGC reward model faithfully captures human aesthetic preferences, so that maximizing its score inside the planner actually makes photos better; the paper's main quantitative metric, UGC Score, is computed with that same model, making part of the evidence self-referential.

What would settle it

A direct falsifier: collect a set of, say, 200 real user photos, run PhotoAgent and a single-step editor on each with the vague prompt, and have a large panel (100+ independent raters) blind-compare the outputs. If the raters prefer the single-step or open-loop outputs at or below chance, while the UGC reward model still assigns higher scores to PhotoAgent, the central claim would collapse. A cheaper check is to compute the correlation between the UGC reward model's rankings and human pairwise preferences on edits that differ only in planning depth.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If correct, users can give a vague goal like "make this image better" and the agent will decompose, sequence, execute, and verify the edits itself, removing the expertise barrier.
  • The explicit aesthetic planner should reduce short-sighted, irreversible edits because MCTS evaluates long-term consequences before execution and the evaluator rejects non-improvements.
  • The UGC-Edit reward model, trained on real user photos, provides an editing-specific evaluation signal that generic aesthetics metrics miss, and the paper reports it correlates with human judgments on an external aesthetic dataset.
  • With a fixed editing backend, PhotoAgent's latency can be traded off by reducing MCTS simulations (about a quarter of the runtime at 5 simulations while keeping comparable scores), and faster generative editors can be swapped in without changing the framework.
  • The closed-loop design prevents errors from accumulating across edits because each step is re-evaluated and unhelpful actions are rolled back.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • One implicit consequence is that the same planning loop could be transferred to other image-transformation domains—restoration, scientific imaging, non-photorealistic stylization—by swapping the executor toolkit and the evaluator reward model, while the MCTS search structure remains intact.
  • The reliance on a learned reward model suggests a possible pathological equilibrium: if the reward model overfits to a particular aesthetic distribution, the agent may confidently converge to edits that look good to the reward model but not to the user; a meaningful stress test would be to compare PhotoAgent's choices against a held-out set of human pairwise preferences on ambiguous images.
  • A testable extension is to enlarge the candidate-action set with more diverse semantic operations and measure whether MCTS's advantage grows with action diversity, since the planner's value is in sequencing many possible atomic edits.
  • The paper's sim-to-real consistency numbers (75–85% top-1 retention at quarter resolution) suggest that even cheaper simulations could be used, but they also identify the simulation-evaluator alignment as the component to monitor if the system is applied to new image types.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. PhotoAgent formulates autonomous photo editing as a closed-loop agentic decision-making problem: a VLM perceiver proposes candidate edits, an MCTS planner explores action sequences, an executor applies top-K actions with heterogeneous tools, and an ensemble evaluator scores intermediate results. The authors introduce UGC-Edit, a dataset of about 7,000 authentic user photos with aesthetic scores, and train a GRPO-based reward model on it. They also construct a 1,017-photo test set. Quantitative results in Table 1, a 20-participant user study, and ablation descriptions are used to support the claim that PhotoAgent consistently improves instruction adherence and visual quality over non-agent and agent baselines.

Significance. The contribution is potentially valuable. Formulating vague-prompt photo editing as long-horizon planning with explicit tree search over editing actions is a well-motivated step beyond single-step editors. The UGC-Edit dataset addresses a real gap: generic IQA and aesthetic datasets are not tailored to authentic user-captured photographs, and a well-validated reward model for this setting could benefit other editing systems. The paper also includes runtime profiling and a discussion of failure cases, which is commendable. However, the current quantitative evidence for the central claim is not yet convincing: the main learned metric is partly self-referential, the external validation is small, and no statistical analysis is provided. With additional validation the work could become a solid contribution.

major comments (4)
  1. [Sec. 5.1, Table 1, Appendix C] The headline metric 'UGC Score' is computed with the same reward model that is part of PhotoAgent's in-loop evaluator (weight 0.8 in Appendix C). Thus Table 1's UGC Score partially measures whether the agent optimized that model, not whether outputs are aesthetically better in an external sense. The problem is concrete: GPT-4o receives UGC Score 4.210 while PhotoAgent receives 4.176, so the paper's own preference proxy does not rank PhotoAgent first. The only non-circular evidence is the small user study (Sec. 5.2, Table 2), which lacks significance testing. Please provide per-image correlation between UGC model scores and human preferences on the edited outputs, or an external validation of the reward model on final system outputs, before claiming 'visual quality' improvements.
  2. [Sec. 5.2, Table 1] No error bars, confidence intervals, or significance tests are reported for any metric. On ImageReward, GPT-4o (0.4115) actually exceeds PhotoAgent (0.4079); on CLIP Similarity the gap is 0.0239; on BRISQUE the gap is 0.0205. Without variance estimates over the 1,017 images or paired per-image tests, the claim that PhotoAgent 'consistently improves' cannot be evaluated. The user study (20 participants, 540 votes) is also not statistically analyzed; 42.0% vs. 30.2% could be within sampling noise. Please report distributions, confidence intervals, and appropriate significance tests.
  3. [Sec. 5.1, Abstract] The central claim includes 'instruction adherence,' but no metric actually measures adherence to the intended instruction. CLIP Similarity is explicitly described in Sec. 5.1 as measuring how well the edited image preserves original content, not instruction following; the vague prompt 'make this image better' has no ground truth. This leaves the 'instruction adherence' half of the headline claim without direct quantitative support. Please either operationalize instruction adherence (e.g., human ratings of instruction compliance or an instruction-following metric) or remove the term from the headline claim.
  4. [Sec. 5.3, Appendix C] The ablation studies are described only qualitatively ('significantly decreases aesthetic metrics', 'lower performance') without a table. Table 6 in Appendix F reports only simulation-budget sweeps, not the UGC-evaluator ablation or the search-depth ablation. Since the UGC evaluator is central to the method, its removal should be quantified with the same metrics as Table 1, including the UGC Score, BRISQUE, and user-study preference. Please add full ablation tables with error bars or significance tests.
minor comments (3)
  1. [Fig. 3 caption] The caption cites 'Qwen3-VL [43]', but reference [43] is the Qwen-image technical report; Qwen3-VL is reference [2]. Please correct the citation.
  2. [Appendix A, Table 3] Table 3 reports sim-to-real consistency but no sample size or confidence interval for the top-1/top-3 retention rates. Please state how many images and rollouts were used to compute these percentages.
  3. [Sec. 5.4] First bullet: 'a broad spectrum tasks' should be 'a broad spectrum of tasks'. Please also add a period after 'ranging from basic retouching ... to high-level semantic operations'. Minor grammar throughout.

Circularity Check

1 steps flagged

UGC Score is computed by the same reward model that steers the MCTS evaluator, making the headline metric partially circular; external user study and BRISQUE provide independent but limited support.

specific steps
  1. fitted input called prediction [Sec. 5.1, Sec. 4, Appendix C, Table 1]
    "Additionally, we report UGC Score (↑), which is calculated from our reward model fine-tuned on user-generated content (Section 4) to better reflect users' aesthetic preferences. ... The UGC evaluation model, serving as a quality indicator, is assigned a weight of 0.8. The final overall score is computed as a weighted sum of these individual metrics, with a total weight normalization of 5.8."

    The UGC Score in Table 1 is defined as the output of the UGC reward model, and the same model is a weighted component (0.8) of the evaluator used inside MCTS to score actions and decide retention/rollback. Thus PhotoAgent explicitly optimizes the function later reported as evidence of aesthetic quality, biasing the comparison by construction. The circularity is partial because Table 1 shows GPT-4o still scores higher on this metric (4.210 vs 4.176), and because independent metrics (BRISQUE, user study) are also reported.

full rationale

The derivation of PhotoAgent is largely self-contained: the MCTS planner, executor routing, and closed-loop rollback are defined independently of the evaluation, and I found no load-bearing self-citations or imported uniqueness theorems. The one significant circular element is the double use of the UGC reward model as both the in-loop planning objective (Appendix C, weight 0.8) and the headline UGC Score metric (Sec. 5.1, Table 1). A system that searches over edits to maximize a score is not independently validated by reporting that same score, even when the score is trained on human preferences; this is the classic objective-as-metric circularity. The paper partially mitigates this by not claiming the best UGC Score in Table 1, by reporting independent metrics (BRISQUE, ImageReward, Laion-Reward), and by providing a user study with 20 participants (Table 2). However, the user study is small, unreplicated, and not statistically analyzed, so it cannot fully remove the concern that the quantitative 'visual quality' evidence is partially self-referential. Overall score 6 reflects partial, not total, circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central claim rests on the validity of the learned reward model and on the approximation that low-resolution simulations predict full-resolution editing outcomes. The system's performance also depends on hand-chosen evaluator weights and MCTS budget. No new physical or conceptual entities are introduced beyond the learned model and dataset.

free parameters (4)
  • Evaluator ensemble weights = CLIP 1.0, aesthetic 2.0, ImageReward 2.0, UGC 0.8
    Appendix C: The final overall score is a weighted sum with weights chosen by hand; central to what the agent optimizes.
  • MCTS search budget = depth 3, 20 simulations, 3 iterations
    Section 5 / Algorithm 1: Default configuration; ablations show performance varies with these, so results depend on chosen values.
  • VLM sampling parameters = temperature 0.7, top-p 0.8
    Appendix C: Used for perceiver candidate generation; affects diversity and quality of actions.
  • UGC evaluator generation params = temperature 0.7, top-p 0.9, max 32 tokens
    Appendix C: Used in text generation component of evaluator.
axioms (6)
  • domain assumption UGC-Edit human scores are representative of general aesthetic preference for user-generated photos
    Section 4: Reward model trained on 7k UGC images; generalization checked on PARA but not exhaustive.
  • domain assumption Low-resolution MCTS rollouts are a faithful approximation of full-resolution editing outcomes
    Appendix A: reports rank correlations, but still an approximation; suboptimal actions may be selected.
  • domain assumption Instruction-based editing tools (Flux.1 Kontext, etc.) can faithfully execute natural-language actions
    Section 3 Executor: relies on external tools; failures shown in Fig. 9.
  • domain assumption The vision-language perceiver generates diverse, atomic, and semantically meaningful candidate actions
    Section 3 Perceiver: prompts designed to elicit aesthetic-driven instructions; no guarantee of coverage or correctness.
  • domain assumption The 1,017-image test set is representative of real-world user-generated content
    Section 5/E: assembled from multiple channels; no formal sampling.
  • domain assumption Standard quality metrics (CLIP Similarity, ImageReward, BRISQUE, Laion-Reward) are meaningful proxies for editing quality
    Section 5.1: Used as evaluation metrics; they are common but not validated for aesthetic editing on UGC.

pith-pipeline@v1.3.0-alltime-deepseek · 15682 in / 10692 out tokens · 88269 ms · 2026-08-02T20:34:14.387514+00:00 · methodology

0 comments
read the original abstract

With the recent fast development of generative models, instruction-based image editing has shown great potential in generating high-quality images. However, the quality of editing highly depends on carefully designed instructions, placing the burden of task decomposition and sequencing entirely on the user. To achieve autonomous image editing, we present PhotoAgent, a system that advances image editing through explicit aesthetic planning. Specifically, PhotoAgent formulates autonomous image editing as a long-horizon decision-making problem. It reasons over user aesthetic intent, plans multi-step editing actions via tree search, and iteratively refines results through closed-loop execution with memory and visual feedback, without requiring step-by-step user prompts. To support reliable evaluation in real-world scenarios, we introduce UGC-Edit, an aesthetic evaluation benchmark consisting of 7,000 photos and a learned aesthetic reward model. We also construct a test set containing 1,017 photos to systematically assess autonomous photo editing performance. Extensive experiments demonstrate that PhotoAgent consistently improves both instruction adherence and visual quality compared with baseline methods. The project page is https://mdyao.github.io/PhotoAgent/.

Figures

Figures reproduced from arXiv: 2602.22809 by King-Man Tam, Menglu Wang, Mingde Yao, Tianfan Xue, Zhiyuan You.

Figure 1
Figure 1. Figure 1: PhotoAgent autonomously performs high-level, semantically meaningful edits aligned with human aesthetic, moving beyond [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detailed loop of PhotoAgent. First, Perceiver extracts semantic cues from the current image and proposes N candidate editing actions. Second, Planner explores the candidate actions through iterative rollouts, scoring, and pruning to progressively refine edits and select the action that achieves the optimal result. Then, the executor applies these edits while the evaluator scores intermediate results, invok… view at source ↗
Figure 3
Figure 3. Figure 3: Pipeline for constructing the UGC-Edit Dataset and training reward model. We start with a diverse pool of source images from [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results. PhotoAgent generates visually pleasing edits by autonomously improving color harmony, composition, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: PhotoAgent with user-guided prompts. marks, so as to fully exploit different tools’ strengths. Third, PhotoAgent includes a UGC-oriented evaluator trained on real user photos and aesthetic ratings, so that the resulting images better reflect real users’ preferences and values instead of optimizing only generic aesthetic scores. Finally, PhotoAgent incorporates a long-horizon plan￾ning mechanism that suppor… view at source ↗
Figure 7
Figure 7. Figure 7: More visual results of PhotoAgent. Details of MCTS Algorithm 1 shows the pseudo-code for the Monte Carlo Tree Search (MCTS) planner at the core of PhotoAgent. The search starts from the current image state st and runs for a set number of simulations. Each simulation follows four main phases: Selection, Expansion, Simulation (including Evaluation), and Backpropagation. Selection: The algorithm moves from th… view at source ↗
Figure 8
Figure 8. Figure 8: More visual results of PhotoAgent. a reward G, which reflects the predicted aesthetic and semantic quality of the edits. Backpropagation: The reward G is propagated backward along the path that was traversed. This updates the visit counts N(s, a) and average rewards Q(s, a) for all visited nodes, helping the selection phase make better decisions in future simulations. After completing all simulations, the … view at source ↗
Figure 9
Figure 9. Figure 9: Some failed results where the editor may have made excessive changes. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Aurora: Unified Video Editing with a Tool-Using Agent

    cs.CV 2026-05 unverdicted novelty 7.0

    Aurora introduces a VLM-based agent that converts raw user video edit requests into structured conditioning inputs for a unified diffusion transformer, improving performance on underspecified tasks via a new benchmark.

  2. Editor's Choice: Evaluating Abstract Intent in Image Editing through Atomic Entity Analysis

    cs.CV 2026-05 unverdicted novelty 7.0

    Presents Entity-Rubrics and AbstractEdit benchmark to measure image editing models on abstract intent, finding standard models struggle to balance edit intent with image preservation.

  3. Banana100: Breaking NR-IQA Metrics by 100 Iterative Image Replications with Nano Banana Pro

    cs.CV 2026-04 unverdicted novelty 7.0

    Banana100 dataset shows that none of 21 popular NR-IQA metrics consistently rate images degraded by 100 iterative edits lower than clean originals.

  4. Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation

    cs.CV 2026-06 unverdicted novelty 6.0

    Qwen-Image-Agent bridges the context gap in text-to-image models via Context-Aware Planning and Context Grounding that integrate plan, reason, search, memory and feedback, achieving SOTA on IA-Bench and related benchmarks.

  5. Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation

    cs.CV 2026-06 unverdicted novelty 6.0

    Qwen-Image-Agent is a unified agent framework that progressively builds sufficient generation context for T2I models via Context-Aware Planning and Context Grounding, achieving SOTA on IA-Bench, Mindbench, and WISE-Verified.

  6. LumiVideo: An Intelligent Agentic System for Video Color Grading

    cs.CV 2026-04 unverdicted novelty 6.0

    LumiVideo deploys an LLM-based agent with RAG and Tree of Thoughts to generate ASC-CDL parameters and 3D LUTs for automatic cinematic color grading from raw log video, approaching expert quality.

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

    cs.CV 2026-04 unverdicted novelty 5.0

    SmartPhotoCrafter performs automatic photographic image editing by coupling an Image Critic module that identifies deficiencies with a Photographic Artist module that generates edits, trained via multi-stage pretraini...

Reference graph

Works this paper leans on

51 extracted references · 16 linked inside Pith · cited by 6 Pith papers

  1. [1]

    Stable diffusion 3.5.https : / / huggingface

    Stability AI. Stable diffusion 3.5.https : / / huggingface . co / stabilityai / stable - diffusion-3.5-large, 2024. Accessed: 2025-09-23. 3

    2024

  2. [2]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025. 2, 4

    Pith/arXiv arXiv 2025

  3. [3]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025. 4, 5

    Pith/arXiv arXiv 2025

  4. [4]

    The opencv library.Dr

    Gary Bradski. The opencv library.Dr. Dobb’s Journal: Soft- ware Tools for the Professional Programmer, 25(11):120– 123, 2000. 2, 5

    2000

  5. [5]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 18392–18402, 2023. 1, 3, 6, 7

    2023

  6. [6]

    A survey of monte carlo tree search methods

    Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samothrakis, and Simon Colton. A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games, 4(1):1–43, 2012. 2

    2012

  7. [7]

    Hunyuanimage 3.0 technical report.arXiv preprint arXiv:2509.23951, 2025

    Siyu Cao, Hangting Chen, Peng Chen, Yiji Cheng, Yutao Cui, Xinchi Deng, Ying Dong, Kipper Gong, Tianpeng Gu, Xiusen Gu, et al. Hunyuanimage 3.0 technical report.arXiv preprint arXiv:2509.23951, 2025. 3

    Pith/arXiv arXiv 2025

  8. [8]

    Monte-carlo tree search: A new framework for game ai

    Guillaume Chaslot, Sander Bakkes, Istvan Szita, and Pieter Spronck. Monte-carlo tree search: A new framework for game ai. InProceedings of the AAAI Conference on Artifi- cial Intelligence and Interactive Digital Entertainment, vol- ume 4, pages 216–217, 2008. 2

    2008

  9. [9]

    Photoartagent: Intelligent photo retouch- ing with language model-based artist agents.arXiv preprint arXiv:2505.23130, 2025

    Haoyu Chen, Keda Tao, Yizao Wang, Xinlei Wang, Lei Zhu, and Jinjin Gu. Photoartagent: Intelligent photo retouch- ing with language model-based artist agents.arXiv preprint arXiv:2505.23130, 2025. 3

    Pith/arXiv arXiv 2025

  10. [10]

    Edival-agent: An object-centric frame- work for automated, fine-grained evaluation of multi-turn editing.arXiv preprint arXiv:2509.13399, 2025

    Tianyu Chen, Yasi Zhang, Zhi Zhang, Peiyu Yu, Shu Wang, Zhendong Wang, Kevin Lin, Xiaofei Wang, Zhengyuan Yang, Linjie Li, et al. Edival-agent: An object-centric frame- work for automated, fine-grained evaluation of multi-turn editing.arXiv preprint arXiv:2509.13399, 2025. 12

    arXiv 2025

  11. [11]

    Stargan: Unified genera- tive adversarial networks for multi-domain image-to-image translation

    Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified genera- tive adversarial networks for multi-domain image-to-image translation. InProceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition, 2018. 3

    2018

  12. [12]

    Instruc- tir: High-quality image restoration following human instruc- tions

    Marcos V Conde, Gregor Geigle, and Radu Timofte. Instruc- tir: High-quality image restoration following human instruc- tions. InEuropean Conference on Computer Vision, pages 1–21. Springer, 2024. 17

    2024

  13. [13]

    Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan. Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025. 1, 3

    Pith/arXiv arXiv 2025

  14. [14]

    Aesthetic predictor v2.5: Siglip-based aes- thetic score predictor.https : / / github

    discus0434. Aesthetic predictor v2.5: Siglip-based aes- thetic score predictor.https : / / github . com / discus0434/aesthetic-predictor-v2-5, 2024. GitHub repository. 3

    2024

  15. [15]

    Monetgpt: Solving puzzles enhances mllms’ image retouch- ing skills.ACM Transactions on Graphics (TOG), 44(4):1– 12, 2025

    Niladri Shekhar Dutt, Duygu Ceylan, and Niloy J Mitra. Monetgpt: Solving puzzles enhances mllms’ image retouch- ing skills.ACM Transactions on Graphics (TOG), 44(4):1– 12, 2025. 2, 3, 4

    2025

  16. [16]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahra- mani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Wein- berger, editors,Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 3

    2014

  17. [17]

    Nano banana: Gemini 2.5 flash image editing model.https://aistudio.google.com/models/ gemini-2-5-flash-image, 2025

    Google. Nano banana: Gemini 2.5 flash image editing model.https://aistudio.google.com/models/ gemini-2-5-flash-image, 2025. Accessed: 2025- 09-23. 3

    2025

  18. [18]

    Prompt-to-prompt im- age editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt im- age editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022. 1, 3

    Pith/arXiv arXiv 2022

  19. [19]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. InAdvances in Neural Information Processing Systems (NeurIPS), 2017. 3

    2017

  20. [20]

    Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, Sumith Ku- lal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas M¨uller, Dustin Podell, Robin Rombach, Harry Saini, Axel 9 Sauer, and Luke Smith. Flux.1 kontext: Flow matching for in-context...

  21. [21]

    Next token is enough: Realistic image quality and aesthetic scoring with multimodal large language model

    Mingxing Li, Rui Wang, Lei Sun, Yancheng Bai, and Xi- angxiang Chu. Next token is enough: Realistic image quality and aesthetic scoring with multimodal large language model. arXiv preprint arXiv:2503.06141, 2025. 5, 6

    Pith/arXiv arXiv 2025

  22. [22]

    Jarvisart: Liberating human artistic creativ- ity via an intelligent photo retouching agent.arXiv preprint arXiv:2506.17612, 2025

    Yunlong Lin, Zixu Lin, Kunjie Lin, Jinbin Bai, Panwang Pan, Chenxin Li, Haoyu Chen, Zhongdao Wang, Xinghao Ding, Wenbo Li, et al. Jarvisart: Liberating human artistic creativ- ity via an intelligent photo retouching agent.arXiv preprint arXiv:2506.17612, 2025. 2, 3, 4, 8

    Pith/arXiv arXiv 2025

  23. [23]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023. 1, 4, 5

    2023

  24. [24]

    Step1x-edit: A practical framework for gen- eral image editing.arXiv preprint arXiv:2504.17761, 2025

    Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Yingming Wang, Honghao Fu, Chun- rui Han, Guopeng Li, Yuang Peng, Quan Sun, Jingwei Wu, Yan Cai, Zheng Ge, Ranchen Ming, Lei Xia, Xianfang Zeng, Yibo Zhu, Binxing Jiao, Xiangyu Zhang, Gang Yu, and Daxin Jiang. Step1x-edit: A practical framework for gen- eral image editing.arXiv...

    Pith/arXiv arXiv 2025

  25. [25]

    No-reference image quality assessment in the spatial domain.IEEE Transactions on image processing, 21(12):4695–4708, 2012

    Anish Mittal, Anush Krishna Moorthy, and Alan Con- rad Bovik. No-reference image quality assessment in the spatial domain.IEEE Transactions on image processing, 21(12):4695–4708, 2012. 5, 6

    2012

  26. [26]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Mak- ing a “completely blind” image quality analyzer.IEEE Sig- nal processing letters, 20(3):209–212, 2012. 5

    2012

  27. [27]

    Nvidia developer blog.https:// developer.nvidia.com/blog

    NVIDIA Corporation. Nvidia developer blog.https:// developer.nvidia.com/blog. Accessed: 2025-11-

    2025

  28. [28]

    Dall·e 3, 2024

    OpenAI. Dall·e 3, 2024. Accessed: 2025-09-23. 3

    2024

  29. [29]

    Gpt-4o.https://openai.com/index/ hello-gpt-4o, 2024

    OpenAI. Gpt-4o.https://openai.com/index/ hello-gpt-4o, 2024. Accessed: 2025-09-23. 1, 3, 7

    2024

  30. [30]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis.arXiv preprint arXiv:2307.01952, 2023. 1, 6, 7

    Pith/arXiv arXiv 2023

  31. [31]

    Promptir: Prompting for all-in- one image restoration.Advances in Neural Information Pro- cessing Systems, 36:71275–71293, 2023

    Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Promptir: Prompting for all-in- one image restoration.Advances in Neural Information Pro- cessing Systems, 36:71275–71293, 2023. 17

    2023

  32. [32]

    Learn- ing transferable visual models from natural language super- vision.arXiv preprint arXiv:2103.00020, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision.arXiv preprint arXiv:2103.00020, 2021. 3, 5, 6

    Pith/arXiv arXiv 2021

  33. [33]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1

    2022

  34. [34]

    Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural in- formation processing systems, 35:25278–25294, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, et al. Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural in- formation processing systems, 35:25278–25294, 2022. 5, 6

    2022

  35. [35]

    Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 5, 6

    Pith/arXiv arXiv 2024

  36. [36]

    Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face.Advances in Neural Information Processing Systems, 36:38154–38180,

    Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face.Advances in Neural Information Processing Systems, 36:38154–38180,

  37. [37]

    Animaker: Automated multi-agent animated storytelling with mcts-driven clip generation, 2025

    Haoyuan Shi, Yunxin Li, Xinyu Chen, Longyue Wang, Bao- tian Hu, and Min Zhang. Animaker: Automated multi-agent animated storytelling with mcts-driven clip generation, 2025. 3

    2025

  38. [38]

    Mastering the game of go with deep neu- ral networks and tree search.Nature, 529(7587):484–489,

    David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrit- twieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neu- ral networks and tree search.Nature, 529(7587):484–489,

  39. [39]

    V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv: Arxiv-2305.16291,

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandku- mar. V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv: Arxiv-2305.16291,

    Pith/arXiv arXiv

  40. [40]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024. 1

    Pith/arXiv arXiv 2024

  41. [41]

    Image quality assessment: From error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: From error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 3

    2004

  42. [42]

    Genartist: Multimodal llm as an agent for unified image gen- eration and editing.Advances in Neural Information Pro- cessing Systems, 37:128374–128395, 2024

    Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image gen- eration and editing.Advances in Neural Information Pro- cessing Systems, 37:128374–128395, 2024. 1

    2024

  43. [43]

    Qwen-image technical report,

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, De- qing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingk...

  44. [44]

    Show-o: One single transformer to unify multimodal understanding and generation.arXiv preprint arXiv:2408.12528, 2024

    Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation.arXiv preprint arXiv:2408.12528, 2024. 3

    Pith/arXiv arXiv 2024

  45. [45]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- 10 cessing Systems, 36:15903–15935, 2023

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- 10 cessing Systems, 36:15903–15935, 2023. 6

    2023

  46. [46]

    Personalized image aes- thetics assessment with rich attributes

    Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aes- thetics assessment with rich attributes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19861–19869, 2022. 12

    2022

  47. [47]

    ReAct: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations (ICLR), 2023. 3, 6, 7

    2023

  48. [48]

    Agent banana: High-fidelity image editing with agentic thinking and tooling.arXiv preprint arXiv:2602.09084, 2026

    Ruijie Ye, Jiayi Zhang, Zhuoxin Liu, Zihao Zhu, Siyuan Yang, Li Li, Tianfu Fu, Franck Dernoncourt, Yue Zhao, Jiacheng Zhu, et al. Agent banana: High-fidelity image editing with agentic thinking and tooling.arXiv preprint arXiv:2602.09084, 2026. 8

    arXiv 2026

  49. [49]

    Personalized image aes- thetics assessment with attribute-guided fine-grained feature representation

    Hancheng Zhu, Zhiwen Shao, Yong Zhou, Guangcheng Wang, Pengfei Chen, and Leida Li. Personalized image aes- thetics assessment with attribute-guided fine-grained feature representation. InProceedings of the 31st ACM International Conference on Multimedia, pages 6794–6802, 2023. 12

    2023

  50. [50]

    Unpaired image-to-image translation using cycle- consistent adversarial networks

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle- consistent adversarial networks. InComputer Vision (ICCV), 2017 IEEE International Conference on, 2017. 3

    2017

  51. [51]

    4kagent: agentic any image to 4k super- resolution.arXiv preprint arXiv:2507.07105, 2025

    Yushen Zuo, Qi Zheng, Mingyang Wu, Xinrui Jiang, Renjie Li, Jian Wang, Yide Zhang, Gengchen Mai, Lihong V Wang, James Zou, et al. 4kagent: agentic any image to 4k super- resolution.arXiv preprint arXiv:2507.07105, 2025. 2, 3, 4, 8 11 A. Sim-to-Real Gap in Low-Resolution Planner Simulation PhotoAgent uses reduced-resolution rollouts to make MCTS planning c...

    Pith/arXiv arXiv 2025