pith. machine review for the scientific record. sign in

arxiv: 2604.08536 · v1 · submitted 2026-04-09 · 💻 cs.CV · cs.AI

Recognition: unknown

RewardFlow: Generate Images by Optimizing What You Reward

Onkar Susladkar , Dong-Hwan Jang , Tushar Prakash , Adheesh Juvekar , Vedant Shah , Ayush Barik , Nabeel Bashir , Muntasir Wahed
show 2 more authors
Ritish Shrirao Ismini Lourentzou
Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:21 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords diffusion modelsimage editingLangevin dynamicsreward optimizationcompositional generationVQA supervisionadaptive policy
0
0 comments X

The pith

RewardFlow steers pretrained diffusion models at inference time by optimizing multiple differentiable rewards through Langevin dynamics guided by a prompt-aware adaptive policy.

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

The paper establishes that an inversion-free method can improve image editing and compositional generation by combining rewards for semantic alignment, perceptual quality, object consistency, and a new differentiable VQA reward, then coordinating them dynamically. It uses multi-reward Langevin dynamics to steer sampling in existing diffusion and flow-matching models. The adaptive policy reads the prompt to infer intent and adjust each reward's weight and the sampling step size on the fly. A sympathetic reader would care because this shifts optimization from model training to inference, letting users get higher-fidelity outputs from off-the-shelf models by directly rewarding the desired properties.

Core claim

RewardFlow unifies complementary differentiable rewards for semantic alignment, perceptual fidelity, localized grounding, object consistency, and human preference, and further introduces a differentiable VQA-based reward that provides fine-grained semantic supervision through language-vision reasoning. To coordinate these heterogeneous objectives, it designs a prompt-aware adaptive policy that extracts semantic primitives from the instruction, infers edit intent, and dynamically modulates reward weights and step sizes throughout sampling. This is realized via multi-reward Langevin dynamics applied at inference time to steer pretrained diffusion and flow-matching models.

What carries the argument

Multi-reward Langevin dynamics coordinated by a prompt-aware adaptive policy that infers edit intent from the input prompt and dynamically modulates individual reward weights plus step sizes during sampling.

If this is right

  • Delivers higher edit fidelity by jointly optimizing semantic, perceptual, and consistency objectives during sampling.
  • Achieves stronger compositional alignment through the addition of fine-grained VQA-based supervision.
  • Applies directly to existing pretrained models without requiring inversion or retraining.
  • Enables automatic, prompt-dependent adjustment of reward influence instead of manual hyperparameter search.

Where Pith is reading between the lines

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

  • The approach could be tested on video or 3D generation by defining analogous differentiable rewards for temporal consistency or depth.
  • It opens a route for incorporating external reasoning models directly into the generative sampling loop rather than only at evaluation time.
  • Users might define task-specific reward combinations at runtime to steer outputs toward particular stylistic or factual constraints.
  • The framework suggests a path to reduce reliance on classifier-free guidance by letting explicit rewards take over more of the control burden.

Load-bearing premise

That the different rewards can be balanced by the adaptive policy without producing conflicts or instabilities in the sampling trajectory.

What would settle it

A controlled experiment on a standard image-editing benchmark that replaces the prompt-aware adaptive policy with fixed equal weights and measures whether edit fidelity and compositional alignment fall below the levels reported for the full method.

Figures

Figures reproduced from arXiv: 2604.08536 by Adheesh Juvekar, Ayush Barik, Dong-Hwan Jang, Ismini Lourentzou, Muntasir Wahed, Nabeel Bashir, Onkar Susladkar, Ritish Shrirao, Tushar Prakash, Vedant Shah.

Figure 1
Figure 1. Figure 1: RewardFlow enables accurate, localized, inversion-free image editing and generation using multi-reward Langevin guidance. Abstract We introduce RewardFlow, an inversion-free framework that steers pretrained diffusion and flow-matching models at in￾ference time through multi-reward Langevin dynamics. Re￾wardFlow unifies complementary differentiable rewards for semantic alignment, perceptual fidelity, locali… view at source ↗
Figure 2
Figure 2. Figure 2: Gradient localization of our differentiable rewards. We visualize the image-space gradient ∇IRtot(·) for various edit prompts. Our proposed rewards prevent semantic leakage by con￾centrating the gradient precisely on target semantic regions, demon￾strating the fine-grained spatial control enabled by RewardFlow. each step, the policy therefore computes the reward weights wi using three sources of informatio… view at source ↗
Figure 3
Figure 3. Figure 3: Image editing qualitative comparison across diverse instruction types. RewardFlow produces edits that are both semantically accurate and spatially localized, while better preserving background structure, lighting, and identity compared to prior methods. (43 NFEs and 20 sampling steps), corresponding to roughly 60–80% fewer sampling steps than gradient-based editors that typically require 50–100 steps. In t… view at source ↗
Figure 4
Figure 4. Figure 4: Text-to-image qualitative results. Across all prompts, RewardFlow produces images that exhibit higher alignment with the textual descriptions while also generating outputs with more visually appealing composition and aesthetics. Llama standing in the grass field A notebook pen and open Laptop on wooden Table beside it Coffee cup and Mobile [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Gradient localization across reward combinations. Including all rewards concentrates gradients to accurate object contours and eliminates leakage. and the ruined castle that other methods fail to realize. In the striped marbles example, RewardFlow more faithfully satisfies the specified object count and patterns. 5. Ablation Studies Reward Components. We analyze both quantitative and qualitative effects of… view at source ↗
Figure 7
Figure 7. Figure 7: Ablations illustrating the effect of removing key com￾ponents. RewardFlow (all components) achieves the best visual consistency and instruction alignment [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: High-resolution images generated by RewardFlow. 1 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Overview of the RewardFlow framework. GENEVAL. GENEVAL is a structured evaluation suite tar￾geting fine-grained text-to-image alignment at the object level. It introduces prompts designed to probe a model’s ability to generate images with correct object presence, co￾occurrence, counting, spatial arrangement, and color at￾tribution. Each generated image is evaluated using auto￾mated pipelines based on pre-t… view at source ↗
Figure 10
Figure 10. Figure 10: Prompt template used for semantic primitives and Rvqa. uses RegionCLIP image-region ψ reg img(I, rm) and text embed￾dings ψ reg text(p). Given region proposals {rm} we compute sm(p) and soft attention weights αm(p), and define Rrg(I (k) , p) = X m αm(p) sm(p). This reward encourages gradients to concentrate on spa￾tial regions that are both semantically and visually aligned with each SP, matching the beha… view at source ↗
Figure 11
Figure 11. Figure 11: Reward progression over time. Cosine similarity backbone. All rewards except Rhps and Rvqa are implemented as cosine similarities between the embeddings of semantic primitives and the current image at step k. Gradients are obtained via automatic differentiation through the corresponding vision–language encoders. KL Tether for Image Editing. For all image editing ex￾periments, we enable the clean-latent KL… view at source ↗
Figure 12
Figure 12. Figure 12: Image Editing Qualitative Results with Flux + RewardFlow. For each input image on the left, RewardFlow is instructed to apply a targeted edit (text below), and the right image shows the generated result. Tasks span from global scene modifications and object-level edits to very fine-grained, localized edits. performance remains relatively stable when scaling from 3B to 8B parameters, indicating that modera… view at source ↗
Figure 13
Figure 13. Figure 13: Image Editing Qualitative Results with Qwen Image + RewardFlow. For each input image on the left, RewardFlow is instructed to apply a targeted edit (text below), and the right image shows the generated result. Tasks span from global scene modifications and object-level edits to very fine-grained, localized edits. by a textured input image. Across all examples, edits are restricted to instruction-relevant … view at source ↗
Figure 14
Figure 14. Figure 14: Image Editing Qualitative Comparisons. Comparison across a range of challenging edits, such as diverse attribute-, style-, and object-level transformations. Each row shows the source image followed by results from strong baselines and RewardFlow. fEdit, FlowChef, and KV-Edit), while others apply the rusty texture too aggressively or inconsistently, affecting broader regions and introducing unnatural appea… view at source ↗
Figure 15
Figure 15. Figure 15: RewardFlow counting failure case. cat’s silhouette, eye color, and overall scene context. Failure Modes. While robust, RewardFlow is bounded by its components. A primary failure mode arises from VQA limitations in fine-grained reasoning like counting. As shown in [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Text-to-image qualitative results with Flux as backbone. Qualitative comparison of Flux, Flux + Global Reward, and Flux + RewardFlow across diverse prompts. 10 [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗
read the original abstract

We introduce RewardFlow, an inversion-free framework that steers pretrained diffusion and flow-matching models at inference time through multi-reward Langevin dynamics. RewardFlow unifies complementary differentiable rewards for semantic alignment, perceptual fidelity, localized grounding, object consistency, and human preference, and further introduces a differentiable VQA-based reward that provides fine-grained semantic supervision through language-vision reasoning. To coordinate these heterogeneous objectives, we design a prompt-aware adaptive policy that extracts semantic primitives from the instruction, infers edit intent, and dynamically modulates reward weights and step sizes throughout sampling. Across several image editing and compositional generation benchmarks, RewardFlow delivers state-of-the-art edit fidelity and compositional alignment.

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 introduces RewardFlow, an inversion-free framework for steering pretrained diffusion and flow-matching models at inference time via multi-reward Langevin dynamics. It unifies differentiable rewards for semantic alignment, perceptual fidelity, localized grounding, object consistency, and human preference, while introducing a new differentiable VQA-based reward for fine-grained language-vision supervision. A prompt-aware adaptive policy is proposed to extract semantic primitives from instructions, infer edit intent, and dynamically modulate reward weights and step sizes. The central claim is state-of-the-art edit fidelity and compositional alignment across image editing and compositional generation benchmarks.

Significance. If the empirical claims hold with proper validation, the work could meaningfully advance inference-time control in generative models by enabling flexible, training-free coordination of heterogeneous objectives through Langevin dynamics. The introduction of a differentiable VQA reward and the adaptive policy represent potentially useful technical ideas for multi-objective optimization in diffusion-based generation.

major comments (2)
  1. [Method (prompt-aware adaptive policy)] The description of the prompt-aware adaptive policy (in the method section) asserts that it resolves objective conflicts and prevents sampling instabilities across heterogeneous rewards, but provides no quantitative analysis of reward trade-offs, trajectory variance, or failure modes on conflicting objectives; this mechanism is load-bearing for the SOTA claim on edit fidelity and compositional alignment.
  2. [Abstract and Experiments] The abstract and results overview assert SOTA performance on several benchmarks but supply no quantitative metrics, baselines, ablation studies, or error analysis to support the central claim; without these, the effectiveness of the unified rewards and adaptive coordination cannot be verified.
minor comments (2)
  1. [Abstract] The abstract would be strengthened by including at least one key quantitative result (e.g., a metric improvement) to substantiate the SOTA assertion.
  2. [Method] Notation for the Langevin dynamics update and reward weighting should be made fully explicit with equations to aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We appreciate the emphasis on strengthening the empirical support for the prompt-aware adaptive policy and the SOTA claims. We address each major comment below and commit to revisions that provide the requested quantitative analysis and clearer experimental presentation.

read point-by-point responses

  1. Referee: [Method (prompt-aware adaptive policy)] The description of the prompt-aware adaptive policy (in the method section) asserts that it resolves objective conflicts and prevents sampling instabilities across heterogeneous rewards, but provides no quantitative analysis of reward trade-offs, trajectory variance, or failure modes on conflicting objectives; this mechanism is load-bearing for the SOTA claim on edit fidelity and compositional alignment.

    Authors: We agree that a dedicated quantitative analysis of the adaptive policy is necessary to substantiate its role in managing heterogeneous rewards. The current manuscript demonstrates the policy's benefits through end-to-end performance on benchmarks but does not isolate metrics such as reward trade-off curves, trajectory variance statistics, or explicit failure cases under conflicting objectives. In the revised version, we will add a new subsection with ablations that report these quantities (e.g., variance in reward gradients and sampling trajectories with/without adaptation, Pareto analysis on conflicting reward pairs, and qualitative/quantitative failure modes). This will directly bolster the SOTA claims on edit fidelity and compositional alignment. revision: yes

  2. Referee: [Abstract and Experiments] The abstract and results overview assert SOTA performance on several benchmarks but supply no quantitative metrics, baselines, ablation studies, or error analysis to support the central claim; without these, the effectiveness of the unified rewards and adaptive coordination cannot be verified.

    Authors: We acknowledge that the abstract and high-level results overview currently state SOTA outcomes without accompanying numbers or error analysis, which limits immediate verifiability. The detailed experimental section does contain quantitative comparisons, baselines, and ablations, but these are not sufficiently highlighted in the overview. In the revision, we will (1) update the abstract to include key quantitative highlights (e.g., specific metric improvements over baselines), (2) expand the results overview with a concise table of main metrics, baseline comparisons, and error bars, and (3) add explicit error analysis and additional ablations on reward unification and adaptive coordination. These changes will make the central claims directly verifiable from the overview. revision: yes

Circularity Check

0 steps flagged

No circularity: RewardFlow derivation is self-contained using external pretrained models and rewards

full rationale

The paper presents RewardFlow as an inference-time framework that applies multi-reward Langevin dynamics to steer existing pretrained diffusion and flow-matching models. It unifies complementary differentiable rewards drawn from prior literature (semantic, perceptual, grounding, consistency, preference, and a new VQA-based reward) and introduces a prompt-aware adaptive policy to modulate them. No equations or method steps reduce by construction to self-defined quantities, fitted inputs renamed as predictions, or load-bearing self-citations whose validity depends on the current work. The central claims of SOTA performance are evaluated on external benchmarks rather than being forced by internal definitions or prior author results invoked as uniqueness theorems. The coordination mechanism, while an assumption, is not circular; it is an empirical claim tested via benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; all elements are described at a high level without implementation details.

pith-pipeline@v0.9.0 · 5445 in / 1050 out tokens · 65317 ms · 2026-05-10T17:21:37.112811+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

57 extracted references · 10 canonical work pages · 6 internal anchors

  1. [1]

    Edicho: Consistent image editing in the wild

    Qingyan Bai, Hao Ouyang, Yinghao Xu, Qiuyu Wang, Ceyuan Yang, Ka Leong Cheng, Yujun Shen, and Qifeng Chen. Edicho: Consistent image editing in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 2

    2025

  2. [2]

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

    Black Forest Labs. FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space. arXiv preprint arXiv:2506.15742, 2025. 1, 5, 7

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023. 2

    work page internal anchor Pith review arXiv 2023

  4. [4]

    Ledits++: Limitless image editing us- ing text-to-image models

    Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing us- ing text-to-image models. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2

    2024

  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. InIEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), 2023. 2

    2023

  6. [6]

    Training-Free Reward-Guided Image Editing via Trajectory Optimal Control

    Jinho Chang, Jaemin Kim, and Jong Chul Ye. Training-free reward-guided image editing via trajectory optimal control. arXiv preprint arXiv:2509.25845, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. InIn- ternational Conference on Learning Representations (ICLR),

  8. [8]

    Scaling rectified flow trans- formers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim En- tezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow trans- formers for high-resolution image synthesis. InInternational Conference on Machine Learning (ICML), 2024. 1

    2024

  9. [9]

    Reno: Enhancing one-step text-to-image models through reward-based noise optimiza- tion.Advances in Neural Information Processing Systems (NeurIPS), 2024

    Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimiza- tion.Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 7

    2024

  10. [10]

    An image is worth one word: Personalizing text-to-image generation using textual inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. InInternational Confer- ence on Learning Representations (ICLR), 2022. 2

    2022

  11. [11]

    Controllable first-frame-guided video editing via mask-aware lora fine-tuning

    Chenjian Gao, Lihe Ding, Xin Cai, Zhanpeng Huang, Zibin Wang, and Tianfan Xue. Controllable first-frame-guided video editing via mask-aware lora fine-tuning. InInterna- tional Conference on Learning Representations (ICLR), 2025. 2

    2025

  12. [12]

    Instantedit: Text-guided few-step image editing with piecewise rectified flow

    Yiming Gong, Zhen Zhu, and Minjia Zhang. Instantedit: Text-guided few-step image editing with piecewise rectified flow. InInternational Conference on Computer Vision (ICCV),

  13. [13]

    Generative adversarial networks.Communi- cations of the ACM, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Communi- cations of the ACM, 2020. 2

    2020

  14. [14]

    Proxedit: Improving tuning-free real image editing with proximal guidance

    Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopou- los, Xiaoxiao He, Yuxiao Chen, et al. Proxedit: Improving tuning-free real image editing with proximal guidance. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 5

    2024

  15. [15]

    Prompt-to-Prompt Image Editing with Cross Attention Control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt Image Editing with Cross Attention Control. InACM SIGGRAPH Asia, 2022. 2, 5

    2022

  16. [16]

    Denoising diffu- sion probabilistic models.Advances in Neural Information Processing Systems (NeurIPS), 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models.Advances in Neural Information Processing Systems (NeurIPS), 2020. 1

    2020

  17. [17]

    T2i-compbench: A comprehensive benchmark for open- world compositional text-to-image generation.Advances in Neural Information Processing Systems (NeurIPS), 2023

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open- world compositional text-to-image generation.Advances in Neural Information Processing Systems (NeurIPS), 2023. 5

    2023

  18. [18]

    An edit friendly ddpm noise space: Inversion and manipulations

    Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5

    2024

  19. [19]

    Direct inversion: Boosting diffusion- based editing with 3 lines of code.arXiv preprint arXiv:2310.01506, 2023

    Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct Inversion: Boosting Diffusion-Based Editing with 3 Lines of Code.arXiv preprint arXiv:2310.01506, 2023. 2, 5

    work page arXiv 2023

  20. [20]

    Pnp inversion: Boosting diffusion-based editing with 3 lines of code

    Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. InInternational Conference on Learning Representations (ICLR), 2024. 5

    2024

  21. [21]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2019. 2

    2019

  22. [22]

    Imagic: Text-based real image editing with diffusion models

    Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

    2023

  23. [23]

    FlowEdit: Inversion-Free 9 Text-Based Editing Using Pre-Trained Flow Models

    Vladimir Kulikov, Matan Kleiner, Inbar Huberman- Spiegelglas, and Tomer Michaeli. FlowEdit: Inversion-Free 9 Text-Based Editing Using Pre-Trained Flow Models. InIn- ternational Conference on Computer Vision (ICCV), 2025. 2, 5

    2025

  24. [24]

    Flow matching for generative mod- eling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative mod- eling. InThe Eleventh International Conference on Learning Representations, 2022. 1

    2022

  25. [25]

    Regiondrag: Fast region- based image editing with diffusion models

    Jingyi Lu, Xinghui Li, and Kai Han. Regiondrag: Fast region- based image editing with diffusion models. InEuropean Conference on Computer Vision (ECCV), 2024. 2

    2024

  26. [26]

    Sdedit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. InInternational Conference on Learning Representations (ICLR), 2021. 2

    2021

  27. [27]

    Origen: Zero-shot 3d orienta- tion grounding in text-to-image generation.arXiv preprint arXiv:2503.22194, 2025

    Yunhong Min, Daehyeon Choi, Kyeongmin Yeo, Jihyun Lee, and Minhyuk Sung. Origen: Zero-shot 3d orienta- tion grounding in text-to-image generation.arXiv preprint arXiv:2503.22194, 2025. 2, 3

    work page arXiv 2025

  28. [28]

    Null-text inversion for editing real im- ages using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real im- ages using guided diffusion models. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

    2023

  29. [29]

    Introducing dall-e 2

    OpenAI. Introducing dall-e 2. OpenAI Blog (2022). https: //openai.com/index/dall-e-2/, 2022. 7

    2022

  30. [30]

    Introducing dall-e 3

    OpenAI. Introducing dall-e 3. OpenAI Blog (2023). https: //openai.com/dall-e-3, 2023. 7

    2023

  31. [31]

    Styleclip: Text-driven manipulation of stylegan imagery

    Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. InInternational Conference on Computer Vision (ICCV), 2021. 2

    2021

  32. [32]

    Metaxas, and Yezhou Yang

    Maitreya Patel, Song Wen, Dimitris N. Metaxas, and Yezhou Yang. FlowChef: Steering of Rectified Flow Models for Con- trolled Generations. InInternational Conference on Computer Vision (ICCV), 2025. 5

    2025

  33. [33]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. InInternational Conference on Learning Representations (ICLR), 2023. 7

    2023

  34. [34]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational Conference on Machine Learning (ICML),

  35. [35]

    Sam 2: Segment any- thing in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment any- thing in images and videos. InThe Thirteenth International Conference on Learning Representations, 2024. 2, 4

    2024

  36. [36]

    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. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7

    2022

  37. [37]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven gen- eration

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven gen- eration. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

    2023

  38. [38]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations,

  39. [39]

    Postedit: Posterior sampling for effi- cient zero-shot image editing

    Feng Tian, Yixuan Li, Yichao Yan, Shanyan Guan, Yanhao Ge, and Xiaokang Yang. Postedit: Posterior sampling for effi- cient zero-shot image editing. InThe Thirteenth International Conference on Learning Representations, 2024. 2

    2024

  40. [40]

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muham- mad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language en- coders with improved semantic understanding, localization, and dense features.arXiv preprint arXiv:2502.14786, 2025. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    Training-free text-guided image editing with visual autoregressive model

    Yufei Wang, Lanqing Guo, Zhihao Li, Jiaxing Huang, Pichao Wang, Bihan Wen, and Jian Wang. Training-free text-guided image editing with visual autoregressive model. InInterna- tional Conference on Computer Vision (ICCV), 2025. 5

    2025

  42. [42]

    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. 5, 7

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  43. [43]

    Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis.arXiv preprint arXiv:2306.09341,

    work page internal anchor Pith review arXiv

  44. [44]

    TurboEdit: Instant Text-Based Im- age Editing Using Few-Step Diffusion Models

    Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. TurboEdit: Instant Text-Based Im- age Editing Using Few-Step Diffusion Models. InEuropean Conference on Computer Vision (ECCV), 2024. 5

    2024

  45. [45]

    Dymo: Training-free diffusion model alignment with dynamic multi-objective scheduling

    Xin Xie and Dong Gong. Dymo: Training-free diffusion model alignment with dynamic multi-objective scheduling. InIEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), 2025. 2

    2025

  46. [46]

    arXiv preprint arXiv:2312.04965 (2023) 18 R.Morita et al

    Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-Free Image Editing with Natural Language. arXiv preprint arXiv:2312.04965, 2023. 5

    work page arXiv 2023

  47. [47]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2024. 7

    2024

  48. [48]

    RegionCLIP: Region- Based Language-Image Pretraining

    Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region- Based Language-Image Pretraining. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4

    2022

  49. [49]

    Training-free geometric image editing on diffusion models

    Hanshen Zhu, Zhen Zhu, Kaile Zhang, Yiming Gong, Yu- liang Liu, and Xiang Bai. Training-free geometric image editing on diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 2 10

    2025

  50. [50]

    KV-Edit: Training-Free Image Editing for Precise Back- ground Preservation

    Tianrui Zhu, Shiyi Zhang, Jiawei Shao, and Yansong Tang. KV-Edit: Training-Free Image Editing for Precise Back- ground Preservation. InInternational Conference on Com- puter Vision (ICCV), 2025. 5 11 : Generate Images by Optimizing What You Reward Supplementary Material Figure 8.High-resolution images generated by RewardFlow. 1

    2025

  51. [51]

    (2), specify the diffusion-strength schedule γk, and provide a derivation showing how Eq

    SDE Formulation In this section, we detail the stochastic differential equa- tion (SDE) that grounds the Langevin-style reverse update in Eq. (2), specify the diffusion-strength schedule γk, and provide a derivation showing how Eq. (2) emerges from sampling a prompt-tilted latent density. Prompt-tilted target density.Let qt(z|p) denote the un- conditional...

  52. [52]

    What color is the bench?

    Datasets and Evaluation To ensure a fair comparison, we adopt the same evaluation protocols and metrics as defined in the original papers of each respective dataset. T2I-COMPBENCH.T2I-COMPBENCHis a large-scale benchmark designed to evaluate compositional text-to-image generation in open-world settings, and consists of approxi- mately 6,000 prompts categor...

  53. [53]

    remove cap

    Implementation Details In this section, we provide additional implementation details for RewardFlow. An overview of the method is illustrated in Figure 9. Unless otherwise stated, we use the same hy- perparameters across all backbones, datasets, and tasks. All experiments are run on a single node with2× NVIDIA A100 GPUs (80 GB each). We implement RewardFl...

  54. [54]

    Include: • Visible subject descriptors (pose, angle, clothing items) actually present

    Extract short edit prompts:output a compact list of 5–12 atomic, actionable tags/phrases that guide the image edit. Include: • Visible subject descriptors (pose, angle, clothing items) actually present. • The edit action(s) and key visual attributes (style, color, size, placement). • Constraints to preserve identity, lighting, composition, realism, and co...

  55. [55]

    Unspecified from image

    Create exactly one Q&A pair focused on the final edited image’s appearance. • Ask **one** question that would most affect the final look (e.g., style, colorway, size/scale, placement, material/finish, mood/lighting continuity). • Give **one** concise answer based on the image/instruction; if not determinable, answer "Unspecified from image." ## Rules - **...

  56. [56]

    As shown in Table 5, RewardFlow consistently improves compositional faithfulness over both backbone models and Table 5.T2I generation on GENEVAL

    Additional Results Text-to-Image Generation.We perform additional text-to-image generation evaluation on GENEVAL. As shown in Table 5, RewardFlow consistently improves compositional faithfulness over both backbone models and Table 5.T2I generation on GENEVAL. Model Overall↑Single↑Two↑Counting↑Colors↑Position↑Color Attribution↑ SD v2.1 0.50 0.98 0.51 0.44 ...

    work page arXiv

  57. [57]

    as if on psychedelics,

    Additional Qualitative Results Image Editing Qualitative Results.Using Flux as the base model, as shown in Figure 12, RewardFlow follows a wide variety of fine-grained instructions while preserving background layout and image identity. RewardFlow can perform strong stylistic changes, such as recoloring the carved wooden ornament “as if on psychedelics,” t...