pith. sign in

arxiv: 2605.21605 · v2 · pith:HGPLKCGFnew · submitted 2026-05-20 · 💻 cs.CV

GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

Sixiang Chen , Zhaohu Xing , Tian Ye , Xinyu Geng , Yunlong Lin , Jianyu Lai , Xuanhua He , Fuxiang Zhai
show 2 more authors
Jialin Gao Lei Zhu
This is my paper

Pith reviewed 2026-05-25 05:37 UTC · model grok-4.3

classification 💻 cs.CV
keywords image generation agentsself-evolving frameworksvisual experience distillationtool orchestrationtrajectory comparisonprompt constructionagentic generation
0
0 comments X

The pith

GenEvolve lets image generation agents self-evolve by distilling structured visual experiences from comparing multiple tool-orchestrated trajectories.

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

The paper proposes GenEvolve as a framework for developing general image-generation agents that improve through self-evolution on diverse requests. Each generation is modeled as a tool-orchestrated trajectory involving evidence gathering, reference selection, skill invocation, and prompt composition. The method compares multiple trajectories for the same request, abstracts best-worst differences into structured visual experience provided only to a privileged teacher branch, and uses on-policy self-distillation to supply dense token-level supervision to the student branch. This approach targets limitations of scalar reward methods by helping the student internalize better strategies for search, knowledge activation, reference selection, and prompt construction. Experiments on public benchmarks and the introduced GenEvolve-Bench demonstrate substantial gains over strong baselines and state-of-the-art performance among current image-generation frameworks.

Core claim

GenEvolve models each generation attempt as a tool-orchestrated trajectory and abstracts best-worst differences into structured visual experience supplied only to a privileged teacher branch, which then provides dense token-level supervision to the student branch through visual experience distillation, enabling improved performance on image generation tasks.

What carries the argument

Tool-Orchestrated Visual Experience Distillation: the mechanism that compares multiple trajectories for the same request, extracts best-worst differences into structured visual experience for a teacher branch, and uses it for dense supervision of the student branch.

Load-bearing premise

Abstracting best-worst differences from multiple tool-orchestrated trajectories into structured visual experience and supplying it only to a privileged teacher branch will produce effective dense token-level supervision that improves the student branch's search, reference selection, and prompt construction.

What would settle it

If the student branch shows no measurable improvement in generation quality metrics after receiving the distilled supervision compared to a non-distilled baseline run on GenEvolve-Bench.

Figures

Figures reproduced from arXiv: 2605.21605 by Fuxiang Zhai, Jialin Gao, Jianyu Lai, Lei Zhu, Sixiang Chen, Tian Ye, Xinyu Geng, Xuanhua He, Yunlong Lin, Zhaohu Xing.

Figure 1
Figure 1. Figure 1: Results of GenEvolve. Top: Representative generation results by our self-evolving agent across diverse open-ended and complicated requests covering architecture, creative transfer, scientific illustration, street scenes, and more, using both Nano Banana Pro and Qwen-Image-Edit as downstream generators. Bottom: Quantitative comparison on (a) our GENEVOLVE-BENCH (KScore + four judge dimensions and Knowledge-… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of GenEvolve-Data and GenEvolve-Bench. The top row presents the construction pipeline: diverse prompts are converted into tool-orchestrated teacher trajectories, audited by VLM-based checks, used to generate and filter GT image cases, and split for supervised training, self-evolution, and held-out evaluation. The bottom row illustrates a representative case, showing how the agent retrieves visual … view at source ↗
Figure 3
Figure 3. Figure 3: Overview of GenEvolve. The student agent orchestrates external search, visual references, and internal generation knowledge to produce a prompt-reference program z = (g, R). During training, multiple trajectories are judged with image/text rewards; best-worst differences are converted into visual experience and injected only into a privileged teacher. GRPO provides trajectory-level optimization, while Visu… view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison on representative GenEvolve-Bench cases. Orange marks external or uncommon knowledge requirements, while blue marks internal generation-knowledge requirements; GenEvolve substantially improves both Qwen-based and Nano Banana Pro generation frameworks. Because tokens are sampled by the old student policy under the plain context, the SDL term uses the on-policy importance ratio ρ on i,t = m… view at source ↗
Figure 5
Figure 5. Figure 5: visualizes the two-track category hierarchy, and [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: GenEvolve-Data construction statistics. The left panel summarizes prompt-to-trajectory filtering for supervised learning, and the right panel summarizes GT image generation, image filtering, self-evolution images, and held-out benchmark cases. B Additional Method Details This section provides implementation details for the rollout protocol, prompt-reference program schema, experience memory, retrieval, GRP… view at source ↗
Figure 7
Figure 7. Figure 7: Case 1 generated images. The search query “winner nationality” (best) vs. “winner national flag” (worst) led to completely different factual grounding and flag stripe colors on the snooker table felt. Case 2 — User Request “Create a retro-futuristic 1970s-style travel poster featuring the French Aérotrain I80. The poster should show the hovertrain gliding on its inverted T-shaped concrete track. In bold vi… view at source ↗
Figure 8
Figure 8. Figure 8: Case 2 generated images. Both trajectories retrieved the same correct facts (430.4 km/h, 1974). The best trajectory called text_rendering and decomposed text into explicit lines with spatial anchors. The worst skipped all skills and crammed text into one string, resulting in unreadable typography. Case 3 — User Request “Generate a street view with two famous European housing complexes side by side. On the … view at source ↗
Figure 9
Figure 9. Figure 9: Case 3 generated images. The best trajectory called spatial_layout and used frame￾relative coordinates (“midground left/right side of the frame, spaced 10 feet apart”). The worst skipped spatial_layout and used vague “side by side at equal width,” causing the buildings to merge and text signs to fail. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Token-level evidence of experience-conditioned SDL guidance. Representative tokens from a single held-out rollout illustrate the two complementary effects of the teacher signal under the prompt-keyed experience bundle. The case asks for a stylised rendering of the Wuppertal Schwebebahn that must respect a real landmark’s identity, layout and a specified visible-carriage count; the bundle instructs the age… view at source ↗
Figure 11
Figure 11. Figure 11: Self-evolution training dynamics. (a) Mean reward across training steps. The smoothed curve (window=25) shows a steady upward trend, indicating that the agent progressively produces higher-quality tool-orchestrated trajectories and prompt-reference programs. (b) SDL loss across training steps. The decreasing trend indicates that the student policy gradually converges toward the experience-conditioned teac… view at source ↗
Figure 1
Figure 1. Figure 1: The evaluation uses the original WISE release [ [PITH_FULL_IMAGE:figures/full_fig_p032_1.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional qualitative results of GenEvolve paired with Nano Banana Pro. The agent autonomously orchestrates search, reference selection, and skill activation to produce high-fidelity images across diverse categories. Examples cover spatial layout, text rendering, quantity counting, attribute binding, anatomy/pose, creative transfer, material physics, and aesthetic drawing skills. 34 [PITH_FULL_IMAGE:fig… view at source ↗
Figure 13
Figure 13. Figure 13: Additional qualitative results of GenEvolve paired with Qwen-Image-Edit. Using the same trained agent policy as in [PITH_FULL_IMAGE:figures/full_fig_p035_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Prompt used for prompt-pool construction. The recipe fields specify the prompt track, category, grounding gap, visual anchor, target capability bundle, and difficulty. 45 [PITH_FULL_IMAGE:figures/full_fig_p045_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: User-side message template used for trajectory filtering. The evaluator receives the original request, final generation prompt, selected-reference constraints, and the structured trajectory trace. 46 [PITH_FULL_IMAGE:figures/full_fig_p046_15.png] view at source ↗
read the original abstract

Open-ended image generation is no longer a simple prompt-to-image problem. High-quality generation often requires an agent to combine a model's internal generative ability with external resources. As requests become more diverse and demanding, we aim to develop a general image-generation agent that can self-evolve through trajectories and use tools more effectively across varied generation challenges. To this end, we propose GenEvolve, a self-evolving framework based on Tool-Orchestrated Visual Experience Distillation. In GenEvolve, each generation attempt is modeled as a tool-orchestrated trajectory, where the agent gathers evidence, selects references, invokes generation skills, and composes them into a prompt-reference program. Unlike existing agentic generation methods that mainly rely on image-level scalar rewards, GenEvolve compares multiple trajectories for the same request and abstracts best-worst differences into structured visual experience, provided only to a privileged teacher branch. Inspired by on-policy self-distillation, Visual Experience Distillation provides dense token-level supervision, helping the student internalize better search, knowledge activation, reference selection, and prompt construction. We further construct GenEvolve-Data and GenEvolve-Bench. Experiments on public benchmarks and GenEvolve-Bench show substantial gains over strong baselines, achieving state-of-the-art performance among current image-generation frameworks. Our website is as follows: https://ephemeral182.github.io/GenEvolve/

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes GenEvolve, a self-evolving image-generation agent framework based on Tool-Orchestrated Visual Experience Distillation. Generation attempts are modeled as tool-orchestrated trajectories involving evidence gathering, reference selection, skill invocation, and prompt composition. Multiple trajectories for the same request are compared to abstract best-worst differences into structured visual experience, supplied only to a privileged teacher branch; this provides dense token-level supervision (inspired by on-policy self-distillation) to improve the student branch's search, knowledge activation, reference selection, and prompt construction. The authors introduce GenEvolve-Data and GenEvolve-Bench and claim substantial gains over strong baselines with SOTA performance on public benchmarks and the new benchmark.

Significance. If the experimental claims hold with proper controls and ablations, the work could advance agentic image generation by replacing scalar rewards with structured visual-experience distillation for self-evolution. The construction of new datasets and benchmarks is a concrete positive contribution. The core idea of distilling trajectory differences at token level to a student branch is a plausible extension of self-distillation techniques, though its effectiveness depends on details not visible in the supplied text.

major comments (2)
  1. Abstract: the central claim of 'substantial gains over strong baselines, achieving state-of-the-art performance' is presented without any quantitative results, tables, baseline details, error bars, or statistical tests in the visible text. This directly undermines the load-bearing performance assertion.
  2. Abstract (and methods description): the Visual Experience Distillation mechanism is described only at a high level ('abstracts best-worst differences into structured visual experience' supplied to a 'privileged teacher branch' for 'dense token-level supervision'). No equations, pseudocode, or implementation specifics are visible, preventing assessment of whether the supervision actually improves the listed student capabilities as claimed.
minor comments (2)
  1. The abstract mentions 'public benchmarks' without naming them or citing the specific tables/figures that report the results.
  2. GenEvolve-Bench is introduced but no details on its construction, size, or evaluation protocol appear in the supplied text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract and the clarity of the proposed method. We address each major comment below and will revise the manuscript accordingly where appropriate.

read point-by-point responses

  1. Referee: Abstract: the central claim of 'substantial gains over strong baselines, achieving state-of-the-art performance' is presented without any quantitative results, tables, baseline details, error bars, or statistical tests in the visible text. This directly undermines the load-bearing performance assertion.

    Authors: We agree that the abstract would be strengthened by including key quantitative results. The full manuscript (Section 4, Tables 1-3) reports specific gains (e.g., +X% on public benchmarks, +Y% on GenEvolve-Bench), baseline details, and error bars from multiple runs. To address the concern directly, we will revise the abstract to incorporate 1-2 representative quantitative highlights while respecting length limits. revision: yes

  2. Referee: Abstract (and methods description): the Visual Experience Distillation mechanism is described only at a high level ('abstracts best-worst differences into structured visual experience' supplied to a 'privileged teacher branch' for 'dense token-level supervision'). No equations, pseudocode, or implementation specifics are visible, preventing assessment of whether the supervision actually improves the listed student capabilities as claimed.

    Authors: The abstract is intentionally high-level due to space constraints. The full manuscript (Section 3.2) provides the formal definition, loss formulation for token-level distillation, pseudocode for trajectory comparison and teacher-student update, and ablation studies showing impact on search/reference selection. We will add a concise sentence to the abstract referencing the core mechanism and will ensure the methods section is explicitly cross-referenced. revision: partial

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The abstract and description present GenEvolve as a conceptual framework for self-evolving agents via trajectory comparison and visual experience distillation, without any equations, parameter-fitting steps, or self-citations. No derivation chain reduces predictions to fitted inputs or self-definitions by construction; claims rest on benchmark experiments treated as external validation. This matches the default case of a self-contained paper with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Review is based solely on the abstract; no free parameters, axioms, or invented entities are specified in the provided text.

invented entities (1)
  • Visual Experience Distillation no independent evidence
    purpose: To supply dense token-level supervision from best-worst trajectory differences to a privileged teacher branch
    Core mechanism named in the abstract

pith-pipeline@v0.9.0 · 5815 in / 1182 out tokens · 28216 ms · 2026-05-25T05:37:02.216555+00:00 · methodology

Review history (2 revisions) →

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

75 extracted references · 40 canonical work pages · 26 internal anchors

  1. [1]

    Qwen3-VL Technical Report

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    Improving image generation with better captions

    James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf, 2(3):8, 2023

    2023

  3. [3]

    FLUX.1 [schnell]

    Black Forest Labs. FLUX.1 [schnell]. Hugging Face model card, 2024. URL https:// huggingface.co/black-forest-labs/FLUX.1-schnell. Accessed: 2026-05-20

    2024

  4. [4]

    FLUX.2 [klein]

    Black Forest Labs. FLUX.2 [klein]. https://huggingface.co/black-forest-labs/ FLUX.2-klein-4B, 2026. FLUX.2 [klein] model family; compact image generation and editing models. Accessed: 2026-05-07

    2026

  5. [5]

    Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer

    Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Steven Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, Xin Jin, Liangchen Li, et al. Z-image: An efficient image generation foundation model with single-stream diffusion transformer.arXiv preprint arXiv:2511.22699, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    HunyuanImage 3.0 Technical Report

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    BLIP3-o: A Family of Fully Open Unified Multimodal Models-Architecture, Training and Dataset

    Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Silvio Savarese, et al. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset.arXiv preprint arXiv:2505.09568, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  8. [8]

    PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis.arXiv preprint arXiv:2310.00426, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

    Postercraft: Rethinking high-quality aesthetic poster generation in a unified framework.arXiv preprint arXiv:2506.10741, 2025

    SiXiang Chen, Jianyu Lai, Jialin Gao, Tian Ye, Haoyu Chen, Hengyu Shi, Shitong Shao, Yunlong Lin, Song Fei, Zhaohu Xing, et al. Postercraft: Rethinking high-quality aesthetic poster generation in a unified framework.arXiv preprint arXiv:2506.10741, 2025

    work page arXiv 2025

  10. [10]

    Posteromni: Generalized artistic poster creation via task distillation and unified reward feedback.arXiv preprint arXiv:2602.12127, 2026

    Sixiang Chen, Jianyu Lai, Jialin Gao, Hengyu Shi, Zhongying Liu, Tian Ye, Junfeng Luo, Xiaoming Wei, and Lei Zhu. Posteromni: Generalized artistic poster creation via task distillation and unified reward feedback.arXiv preprint arXiv:2602.12127, 2026

    work page arXiv 2026

  11. [11]

    Emerging Properties in Unified Multimodal Pretraining

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    Hdpo: Hybrid distillation policy optimization via privileged self-distillation.arXiv preprint arXiv:2603.23871, 2026

    Ken Ding. Hdpo: Hybrid distillation policy optimization via privileged self-distillation.arXiv preprint arXiv:2603.23871, 2026

    work page arXiv 2026

  13. [13]

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transform- ers for high-resolution image synthesis.arXiv preprint arXiv:2403.03206, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  14. [14]

    Gen-Searcher: Reinforcing Agentic Search for Image Generation

    Kaituo Feng, Manyuan Zhang, Shuang Chen, Yunlong Lin, Kaixuan Fan, Yilei Jiang, Hongyu Li, Dian Zheng, Chenyang Wang, and Xiangyu Yue. Gen-searcher: Reinforcing agentic search for image generation.arXiv preprint arXiv:2603.28767, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  15. [15]

    Gemini api release notes: Gemini 3 pro preview

    Google AI for Developers. Gemini api release notes: Gemini 3 pro preview. https:// ai.google.dev/gemini-api/docs/changelog, November 2025. Official release note for gemini-3-pro-preview. Accessed: 2026-05-07

    2025

  16. [16]

    Introducing nano banana pro

    Google DeepMind. Introducing nano banana pro. https://blog.google/technology/ai/ nano-banana-pro/, November 2025. Google DeepMind product release for the Nano Banana Pro image generation and editing model built on Gemini 3 Pro. Accessed: 2026-05-06. 12

    2025

  17. [17]

    Mind-brush: Integrating agentic cognitive search and reasoning into image generation.arXiv preprint arXiv:2602.01756, 2026

    Jun He, Junyan Ye, Zilong Huang, Dongzhi Jiang, Chenjue Zhang, Leqi Zhu, Renrui Zhang, Xiang Zhang, and Weijia Li. Mind-brush: Integrating agentic cognitive search and reasoning into image generation.arXiv preprint arXiv:2602.01756, 2026

    work page arXiv 2026

  18. [18]

    Gems: Agent-native multimodal generation with memory and skills.arXiv preprint arXiv:2603.28088, 2026

    Zefeng He, Siyuan Huang, Xiaoye Qu, Yafu Li, Tong Zhu, Yu Cheng, and Yang Yang. Gems: Agent-native multimodal generation with memory and skills.arXiv preprint arXiv:2603.28088, 2026

    work page arXiv 2026

  19. [19]

    Reinforcement Learning via Self-Distillation

    Jonas Hübotter, Frederike Lübeck, Lejs Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, et al. Reinforcement learning via self-distillation.arXiv preprint arXiv:2601.20802, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  20. [20]

    Genagent: Scaling text-to-image generation via agentic multimodal reasoning.arXiv preprint arXiv:2601.18543, 2026

    Kaixun Jiang, Yuzheng Wang, Junjie Zhou, Pandeng Li, Zhihang Liu, Chen-Wei Xie, Zhaoyu Chen, Yun Zheng, and Wenqiang Zhang. Genagent: Scaling text-to-image generation via agentic multimodal reasoning.arXiv preprint arXiv:2601.18543, 2026

    work page arXiv 2026

  21. [21]

    Craft: Continuous rea- soning and agentic feedback tuning for multimodal text-to-image generation.arXiv preprint arXiv:2512.20362, 2025

    V Kovalev, A Kuvshinov, A Buzovkin, D Pokidov, and D Timonin. Craft: Continuous rea- soning and agentic feedback tuning for multimodal text-to-image generation.arXiv preprint arXiv:2512.20362, 2025

    work page arXiv 2025

  22. [22]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, et al. Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [24]

    URLhttps://arxiv.org/abs/2510.16888

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [25]

    Rethinking kl regularization in rlhf: From value estimation to gradient optimization.arXiv preprint arXiv:2510.01555, 2025

    Kezhao Liu, Jason Klein Liu, Mingtao Chen, and Yiming Liu. Rethinking kl regularization in rlhf: From value estimation to gradient optimization.arXiv preprint arXiv:2510.01555, 2025

    work page arXiv 2025

  25. [26]

    LongCat-Image Technical Report

    Meituan LongCat Team, Hanghang Ma, Haoxian Tan, Jiale Huang, Junqiang Wu, Jun-Yan He, Lishuai Gao, Songlin Xiao, Xiaoming Wei, Xiaoqi Ma, Xunliang Cai, Yayong Guan, and Jie Hu. LongCat-Image Technical Report.arXiv preprint arXiv:2512.07584, 2025. URL https://arxiv.org/abs/2512.07584

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [27]

    WISE: A World Knowledge-Informed Semantic Evaluation for Text-to-Image Generation

    Yuwei Niu, Munan Ning, Mengren Zheng, Weiyang Jin, Bin Lin, Peng Jin, Jiaqi Liao, Chaoran Feng, Kunpeng Ning, Bin Zhu, et al. Wise: A world knowledge-informed semantic evaluation for text-to-image generation.arXiv preprint arXiv:2503.07265, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [28]

    Introducing 4o Image Generation

    OpenAI. Introducing 4o Image Generation. OpenAI blog, 2025. URL https://openai.com/ index/introducing-4o-image-generation/. Accessed: 2026-05-20

    2025

  28. [29]

    Scalable Diffusion Models with Transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers.arXiv preprint arXiv:2212.09748, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  29. [30]

    Lumina-image 2.0: A unified and efficient image generative framework.arXiv preprint arXiv:2503.21758, 2025

    Qi Qin, Le Zhuo, Yi Xin, Ruoyi Du, Zhen Li, Bin Fu, Yiting Lu, Jiakang Yuan, Xinyue Li, Dongyang Liu, Xiangyang Zhu, Manyuan Zhang, Will Beddow, Erwann Millon, Victor Perez, Wenhai Wang, Conghui He, Bo Zhang, Xiaohong Liu, Hongsheng Li, Yu Qiao, Chang Xu, and Peng Gao. Lumina-image 2.0: A unified and efficient image generative framework.arXiv preprint arX...

    work page arXiv 2025

  30. [31]

    Hierarchical Text-Conditional Image Generation with CLIP Latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  31. [32]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn 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

    2022

  32. [33]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022. 13

    2022

  33. [34]

    Approximating kl divergence

    John Schulman. Approximating kl divergence. http://joschu.net/blog/kl-approx. html, 2020. Blog post

    2020

  34. [35]

    Bytedance Seed. Seed2. 0 model card: Towards intelligence frontier for real-world complexity. Technical report, Technical report, Bytedance, 2025. URL https://lf3-static. bytednsdoc. com . . . , 2026

    2025

  35. [36]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    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 mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [37]

    Stable diffusion 3.5 large

    Stability AI. Stable diffusion 3.5 large. https://huggingface.co/stabilityai/ stable-diffusion-3.5-large , 2024. Official model card for Stable Diffusion 3.5 Large. Accessed: 2026-05-07

    2024

  37. [38]

    On a few pitfalls in kl divergence gradient estimation for rl

    Yunhao Tang and Rémi Munos. On a few pitfalls in kl divergence gradient estimation for rl. arXiv preprint arXiv:2506.09477, 2025

    work page arXiv 2025

  38. [39]

    Chameleon: Mixed-Modal Early-Fusion Foundation Models

    Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models.arXiv preprint arXiv:2405.09818, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  39. [40]

    Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding

    Tencent Hunyuan Team. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding.arXiv preprint arXiv:2405.08748, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  40. [41]

    Open multimodal retrieval-augmented factual image generation.arXiv preprint arXiv:2510.22521, 2025

    Yang Tian, Fan Liu, Jingyuan Zhang, Wei Bi, Yupeng Hu, and Liqiang Nie. Open multimodal retrieval-augmented factual image generation.arXiv preprint arXiv:2510.22521, 2025

    work page arXiv 2025

  41. [42]

    Maestro: Self-improving text-to-image generation via agent orchestration.arXiv preprint arXiv:2509.10704, 2025

    Xingchen Wan, Han Zhou, Ruoxi Sun, Hootan Nakhost, Ke Jiang, Rajarishi Sinha, and Sercan Ö Arık. Maestro: Self-improving text-to-image generation via agent orchestration.arXiv preprint arXiv:2509.10704, 2025

    work page arXiv 2025

  42. [43]

    DeepGen 1.0: A Lightweight Unified Multimodal Model for Advancing Image Generation and Editing

    Dianyi Wang, Ruihang Li, Feng Han, Chaofan Ma, Wei Song, Siyuan Wang, Yibin Wang, Yi Xin, Hongjian Liu, Zhixiong Zhang, Shengyuan Ding, Tianhang Wang, Zhenglin Cheng, Tao Lin, Cheng Jin, Kaicheng Yu, Jingjing Chen, Wenjie Wang, Zhongyu Wei, and Jiaqi Wang. DeepGen 1.0: A Lightweight Unified Multimodal Model for Advancing Image Generation and Editing. arXi...

    work page arXiv 2026

  43. [44]

    Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents

    Hao Wang, Guozhi Wang, Han Xiao, Yufeng Zhou, Yue Pan, Jichao Wang, Ke Xu, Yafei Wen, Xiaohu Ruan, Xiaoxin Chen, et al. Skill-sd: Skill-conditioned self-distillation for multi-turn llm agents.arXiv preprint arXiv:2604.10674, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [45]

    Emu3: Next-Token Prediction is All You Need

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-token prediction is all you need.CoRR,...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2024

  45. [46]

    Emu3: Next-Token Prediction is All You Need

    Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  46. [47]

    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, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  47. [48]

    OmniGen2: Towards Instruction-Aligned Multimodal Generation

    Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Exploration to advanced multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  48. [49]

    Show-o: One Single Transformer to Unify Multimodal Understanding and Generation

    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 trans- former to unify multimodal understanding and generation.arXiv preprint arXiv:2408.12528, 2024. 14

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [50]

    On-Policy Context Distillation for Language Models

    Tianzhu Ye, Li Dong, Xun Wu, Shaohan Huang, and Furu Wei. On-policy context distillation for language models.arXiv preprint arXiv:2602.12275, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  50. [51]

    Du, and Xinglong Wu

    Huichao Zhang, Liao Qu, Yiheng Liu, Hang Chen, Yangyang Song, Yongsheng Dong, Shikun Sun, Xian Li, Xu Wang, Yi Jiang, Hu Ye, Bo Chen, Yiming Gao, Peng Liu, Akide Liu, Zhipeng Yang, Qili Deng, Linjie Xing, Jiyang Liu, Zhao Wang, Yang Zhou, Mingcong Liu, Yi Zhang, Qian He, Xiwei Hu, Zhongqi Qi, Jie Shao, Zhiye Fu, Shuai Wang, Fangmin Chen, Xuezhi Chai, Zhih...

    work page arXiv 2026

  51. [52]

    Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

    Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models.arXiv preprint arXiv:2601.18734, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  52. [53]

    the first reference image

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. Llamafactory: Unified efficient fine-tuning of 100+ language models. InProceedings of the 62nd annual meeting of the association for computational linguistics (volume 3: system demonstrations), pages 400–410, 2024. 15 A GenEvolve-Data Construction A.1 Prompt Pool Recipes GenEvolve-Data ...

    2024

  53. [54]

    Aérotrain I80 | Official World Speed Record: 430.4 km/h (267 mph) | 1974

    Skipped text_rendering — crammed all text into a single long string “Aérotrain I80 | Official World Speed Record: 430.4 km/h (267 mph) | 1974” instead of decomposing into separate lines with spatial anchors → text_rendering=fail

    1974

  54. [55]

    Used a second style reference image instead of skill guidance — relied on the model to mimic typography from a reference rather than applying explicit font/layout rules→aesthetic=fail,attribute=fail

  55. [56]

    lower third

    Same correct facts as best, but without text_rendering guidance the poster is unreadable: spatial_layout=fail. Case 2 — Extracted Experience Slots (Delta S1 Search strategy:Execute parallel searches for both the historical event (speed/year) and the physical design (inverted T-track) to ground both text and visuals. S2 Knowledge activation:Call text_rende...

    1974

  56. [57]

    placed side by side at equal width

    Missing spatial_layout — used vague “placed side by side at equal width” without frame-relative coordinates →buildings overlap or merge,spatial_layout=fail

  57. [58]

    Without precise spatial anchors, text signs float or attach to the wrong building →text_rendering =fail, attribute_binding=fail

  58. [59]

    midground left/right side of the frame, spaced 10 feet apart

    Missing physical_material_consistency — sign materials (wood vs metal) not properly grounded → physical_material=partial. Best(R= 0.80): correct layout, both signs legible Worst(R= 0.40): merged buildings, text failure Figure 9:Case 3 generated images.The best trajectory called spatial_layout and used frame- relative coordinates (“midground left/right sid...

    2024

  59. [60]

    Search for missing world knowledge and visual references (grounding)

  60. [61]

    Apply prompt-writing skill guidance -- spatial layout, aesthetic drawing, text rendering, creative drawing, anatomy/body coherence, attribute binding, physical/material consistency, quantity counting -- to improve the quality and controllability of the final prompt (skill integration)

  61. [62]

    FINAL STEP

    Produce a grounded AND skill-enhanced generation-ready prompt that combines both search evidence and skill refinement Output format (ULTRA-STRICT): You MUST output exactly one of the following formats per round: (1) <think> ... </think> <tool_call> ... </tool_call> OR (2) <think> ... </think> <answer> ... </answer> - You are FORBIDDEN to output more than ...

  62. [63]

    Trigger when

    Evaluate each skill independently: does the prompt GENUINELY match the " Trigger when" condition? If yes, call it. If it matches the "Do NOT trigger " condition, skip it

  63. [64]

    When you receive skill guidance, your NEXT response MUST analyze how to apply it -- explicitly state which parts of the guidance you will use and how they improve the gen_prompt

  64. [65]

    Do not call a skill and then ignore its advice

    When you call a skill, you MUST actually USE its guidance in your final gen_prompt. Do not call a skill and then ignore its advice

  65. [66]

    search" (text): confirm identities, event names, dates, locations, specs. Typically 1-2 calls are enough. -

    Multiple skills are encouraged when the prompt has multiple distinct challenges. Do not artificially limit yourself to one skill if more are genuinely needed. - "search" (text): confirm identities, event names, dates, locations, specs. Typically 1-2 calls are enough. - "image_search": find visual references for real entities. Typically 1-2 calls are enoug...

  66. [67]

    Output exactly{n}JSON objects in one JSON array

  67. [68]

    The user-facing “prompt” must be natural and mustNOTmention skill names or tool names

  68. [69]

    Each prompt must require image_search candidate visual evidence; requires_image_search must be true

  69. [70]

    For T1, most prompts should require text search to verify a concrete factual detail that affects the image

  70. [71]

    For T3, text search is optional, butimage_searchmust still be necessary

  71. [72]

    Prompts should be visually evaluable: a reward model should be able to tell if the final generated image succeeded or failed

  72. [73]

    Prefer mid-tail real entities/objects/places/events: searchable, but not trivial

  73. [74]

    Avoid unsafe/private-person content

  74. [75]

    In metadata, describe what must be verified; doNOTfill in the factual answer unless it is already explicitly present in the user-facing prompt

  75. [76]

    prompt":

    The prompt should naturally require the target skill bundle as a whole, but must not mention skill names. Do not make every item equally complex; vary how the bundle appears. For each object, use exactly this schema: { "prompt": "...", "requires_text_search": true/false, "requires_image_search": true, "factual_gap": "short explanation", "visual_anchor_nee...