pith. machine review for the scientific record. sign in

arxiv: 2603.06165 · v2 · submitted 2026-03-06 · 💻 cs.CV · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Reflective Flow Sampling Enhancement

Zikai Zhou , Muyao Wang , Shitong Shao , Lichen Bai , Haoyi Xiong , Bo Han , Zeke Xie
Authors on Pith no claims yet

Pith reviewed 2026-05-15 15:27 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords text-to-image generationflow matchinginference enhancementprompt alignmentFLUXgradient ascentsampling method
0
0 comments X

The pith

RF-Sampling lets flow models like FLUX climb text-image alignment scores at inference time by combining textual representations and flow inversion.

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

The paper presents Reflective Flow Sampling as a training-free method that improves text-to-image output in flow-matching models. It formally derives that the procedure implicitly runs gradient ascent on the alignment score between prompt and image. The approach works by linearly mixing textual features and feeding them through flow inversion to steer sampling toward prompt-consistent noise regions. Experiments show gains in image quality and prompt fidelity on standard benchmarks, plus a limited form of test-time scaling on FLUX. The method is positioned as the first inference enhancement that transfers effectively to CFG-distilled flow architectures.

Core claim

RF-Sampling implicitly performs gradient ascent on the text-image alignment score by leveraging a linear combination of textual representations and integrating them with flow inversion, allowing the model to explore noise spaces more consistent with the input prompt.

What carries the argument

Reflective Flow Sampling, which applies a linear combination of textual representations together with flow inversion to perform the implicit ascent.

If this is right

  • Generation quality rises consistently across multiple text-to-image benchmarks.
  • Prompt alignment improves without any model retraining or fine-tuning.
  • Test-time scaling becomes observable to a limited degree on FLUX.
  • The same procedure extends to other CFG-distilled flow variants.

Where Pith is reading between the lines

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

  • Similar derivations may be attempted for non-distilled flow models by relaxing the CFG assumption.
  • The linear-text-combination step could be tested as a modular add-on inside existing flow samplers.
  • If the ascent property generalizes, inference-time alignment gains might appear in related generative tasks such as video or audio synthesis.

Load-bearing premise

The formal derivation that RF-Sampling performs gradient ascent holds only for CFG-distilled flow models such as FLUX.

What would settle it

A measurement that records no increase in the text-image alignment score across sampling steps on a CFG-distilled flow model would falsify the central derivation.

Figures

Figures reproduced from arXiv: 2603.06165 by Bo Han, Haoyi Xiong, Lichen Bai, Muyao Wang, Shitong Shao, Zeke Xie, Zikai Zhou.

Figure 1
Figure 1. Figure 1: Qualitative comparisons with three representative flow models. Images for each prompt are synthesized using the same [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: RF-Sampling outperforms standard sampling with the same time consumption and significantly enhances the performance [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of RF-Sampling. Compared to previous methods, RF-Sampling employs interpolation on text embeddings [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The winning rate of RF-Sampling over other methods on [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The winning rate of RF-Sampling over other methods on [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation of the gap between shigh and slow. When the gap of shigh - slow increases within a certain range, the quality of synthesized images improves. The dotted lines represents the performance of the standard method. This indicates that within a certain range of values, RF-Sampling perform better than the standard one, demonstrating the robustness of it. No low > high low < high 21.8 21.9 22.0 22.1 22.2 … view at source ↗
Figure 7
Figure 7. Figure 7: Ablation study on the effect of βlow and βhigh. No β means that we do not implement the interpolation weight in Eqn. 2. The results reveal that following the high-weight denoising → low-weight inversion paradigm can enchance the quality of synthesized images. The dotted lines represents the performance of the standard method. This indicates that within a certain range of values, RF-Sampling perform better … view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of the sampling trajectories sampled by [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Ablation study of standard guidance scale [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Robustness to the RF-Sampling steps. The horizontal axis shows the ratio of RF-Sampling operations during the whole [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: We explore the influence of merge ratio γ on Pick-a-Pic dataset. The results across 4 metrics reveal that γ = 0.5 is a better choice, where the synthesized images are the best. The dotted lines represents the performance of the standard method. This indicates that within a certain range of values, RF-Sampling perform better than the standard one, demonstrating the robustness of it [PITH_FULL_IMAGE:figure… view at source ↗
Figure 12
Figure 12. Figure 12: We combine our proposed methods with existing LoRAs in FLUX community. Our RF-Sampling can be directly [PITH_FULL_IMAGE:figures/full_fig_p009_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Visualization of synthesized images with different [PITH_FULL_IMAGE:figures/full_fig_p011_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Visualization of synthesized images with different [PITH_FULL_IMAGE:figures/full_fig_p012_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Image editing experiments on FLUX-Kontext Bench [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Visualizations of the sampling trajectories of RF-Sampling and the standard method. we randomly select two ImageNet [PITH_FULL_IMAGE:figures/full_fig_p017_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: We directly extend our proposed method to video generation task on Wan2.1-T2V-1.3B. The visualizations show the [PITH_FULL_IMAGE:figures/full_fig_p018_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Visual results of FLUX-Lite with guidance scale w = 1. The generated images remain semantically aligned with the input text prompts, demonstrating that the model’s output is still conditionally generated even at the minimum guidance scale. This empirically verifies that CFG-distilled models like FLUX do not possess a true unconditional generation mode, and setting w = 1 does not produce unconditional outp… view at source ↗
Figure 19
Figure 19. Figure 19: Impact of Merge Ratio γ on generation quality. The inverted U-shaped curves across all metrics confirm the existence of an optimal step size, balancing gradient alignment and manifold constraints. FLUX-Dev shows significantly higher robustness to large γ values than FLUX-Lite, attributed to the smoother latent manifold of the larger model. Dotted curve lines represent quadratic fits to the data. • High We… view at source ↗
Figure 20
Figure 20. Figure 20: We combine our proposed methods with existing LoRAs in FLUX community. Our RF-Sampling can be directly applied to the corresponding downstream tasks, validating the generalizability of our method [PITH_FULL_IMAGE:figures/full_fig_p025_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: We extend our proposed methods to image editing tasks on FLUX-Kontext. Our RF-Sampling can be directly applied to the corresponding downstream tasks, validating the effectiveness of our method [PITH_FULL_IMAGE:figures/full_fig_p026_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: The winning rate of RF-Sampling over other methods on SD3.5 on Pick-a-Pic dataset. The standard sampling (baseline) [PITH_FULL_IMAGE:figures/full_fig_p027_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: The winning rate of RF-Sampling over other methods on SD3.5 on DrawBench dataset. The standard sampling (baseline) [PITH_FULL_IMAGE:figures/full_fig_p027_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: The winning rate of RF-Sampling over other methods on SD3.5 on the animation subset of HPD v2 dataset. The [PITH_FULL_IMAGE:figures/full_fig_p027_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: The winning rate of RF-Sampling over other methods on SD3.5 on the photo subset of HPD v2 dataset. The standard [PITH_FULL_IMAGE:figures/full_fig_p027_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: The winning rate of RF-Sampling over other methods on SD3.5 on the concept-art subset of HPD v2 dataset. The [PITH_FULL_IMAGE:figures/full_fig_p028_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: The winning rate of RF-Sampling over other methods on SD3.5 on the painting subset of HPD v2 dataset. The standard [PITH_FULL_IMAGE:figures/full_fig_p028_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: The winning rate of RF-Sampling over the standard one on FLUX-Lite and FLUX-Dev on Pick-a-Pic and DrawBench [PITH_FULL_IMAGE:figures/full_fig_p028_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: The winning rate of RF-Sampling over the standard one on FLUX-Lite on the 4 subsets of HPD v2 datasets. The [PITH_FULL_IMAGE:figures/full_fig_p028_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: The winning rate of RF-Sampling over the standard one on FLUX-Dev on the 4 subsets of HPD v2 datasets. The [PITH_FULL_IMAGE:figures/full_fig_p029_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Synthesized images of FLUX-Lite on anime subset of HPD v2 [PITH_FULL_IMAGE:figures/full_fig_p029_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Synthesized images of FLUX-Lite on photography subset of HPD v2 [PITH_FULL_IMAGE:figures/full_fig_p029_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: Synthesized images of FLUX-Lite on painting subset of HPD v2 [PITH_FULL_IMAGE:figures/full_fig_p030_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: Synthesized images of FLUX-Lite on concept-art subset of HPD v2 [PITH_FULL_IMAGE:figures/full_fig_p030_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: Synthesized images of FLUX-Lite on GenEval [PITH_FULL_IMAGE:figures/full_fig_p031_35.png] view at source ↗
Figure 36
Figure 36. Figure 36: Synthesized images of FLUX-Lite on Pick-a-Pic and DrawBench [PITH_FULL_IMAGE:figures/full_fig_p031_36.png] view at source ↗
Figure 37
Figure 37. Figure 37: Synthesized images of FLUX-Dev on anime subset of HPD v2 [PITH_FULL_IMAGE:figures/full_fig_p032_37.png] view at source ↗
Figure 38
Figure 38. Figure 38: Synthesized images of FLUX-Dev on photography subset of HPD v2 [PITH_FULL_IMAGE:figures/full_fig_p032_38.png] view at source ↗
Figure 39
Figure 39. Figure 39: Synthesized images of FLUX-Dev on painting subset of HPD v2 [PITH_FULL_IMAGE:figures/full_fig_p033_39.png] view at source ↗
Figure 40
Figure 40. Figure 40: Synthesized images of FLUX-Dev on concept-art subset of HPD v2 [PITH_FULL_IMAGE:figures/full_fig_p033_40.png] view at source ↗
Figure 41
Figure 41. Figure 41: Synthesized images of FLUX-Dev on GenEval [PITH_FULL_IMAGE:figures/full_fig_p034_41.png] view at source ↗
Figure 42
Figure 42. Figure 42: Synthesized images of FLUX-Dev on Pick-a-Pic and DrawBench [PITH_FULL_IMAGE:figures/full_fig_p034_42.png] view at source ↗
read the original abstract

The growing demand for text-to-image generation has led to rapid advances in generative modeling. Recently, text-to-image diffusion models trained with flow matching algorithms, such as FLUX, have achieved remarkable progress and emerged as strong alternatives to conventional diffusion models. At the same time, inference-time enhancement strategies have been shown to improve the generation quality and text-prompt alignment of text-to-image diffusion models. However, these techniques are mainly applicable to conventional diffusion models and usually fail to perform well on flow models. To bridge this gap, we propose Reflective Flow Sampling (RF-Sampling), a theoretically-grounded and training-free inference enhancement framework explicitly designed for flow models, especially for the CFG-distilled variants (i.e., models distilled from CFG guidance techniques), like FLUX. Departing from heuristic interpretations, we provide a formal derivation proving that RF-Sampling implicitly performs gradient ascent on the text-image alignment score. By leveraging a linear combination of textual representations and integrating them with flow inversion, RF-Sampling allows the model to explore noise spaces that are more consistent with the input prompt. Extensive experiments across multiple benchmarks demonstrate that RF-Sampling consistently improves both generation quality and prompt alignment. Moreover, RF-Sampling is also the first inference enhancement method that can exhibit test-time scaling ability to some extent on FLUX.

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 manuscript introduces Reflective Flow Sampling (RF-Sampling), a training-free inference enhancement framework for flow-matching text-to-image models such as FLUX. It claims a formal derivation showing that RF-Sampling implicitly performs gradient ascent on the text-image alignment score by combining textual representations linearly and integrating them with flow inversion. Experiments across benchmarks report consistent gains in generation quality, prompt alignment, and limited test-time scaling behavior.

Significance. If the central derivation is rigorous, the work would be significant for supplying the first theoretically motivated inference-time method tailored to CFG-distilled flow models, filling a gap left by techniques developed primarily for diffusion models. The reported test-time scaling observation, even if partial, is a concrete strength that could inform future scaling analyses in generative sampling.

major comments (2)
  1. [Abstract / formal derivation] Abstract and formal derivation section: the claim that RF-Sampling exactly performs gradient ascent on the alignment score rests on an unverified equivalence between the sampling update and the gradient of an alignment objective; the derivation must explicitly address discretization error from numerical ODE solvers and non-differentiability introduced by CFG distillation, or quantify the approximation gap.
  2. [Abstract] Abstract: the text-image alignment score underlying the gradient-ascent interpretation is never defined; if this score is computed from quantities internal to the same model, the argument risks circularity and the proof must state the objective function explicitly.
minor comments (2)
  1. [Experiments] Experiments section: supply the precise list of benchmarks, quantitative prompt-alignment metrics, and statistical significance tests so that the reported consistent gains can be reproduced and compared.
  2. [Method] Method section: clarify the exact coefficients and normalization used in the linear combination of textual representations.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, agreeing where the manuscript requires clarification and outlining the specific revisions we will make to strengthen the formal claims.

read point-by-point responses

  1. Referee: [Abstract / formal derivation] Abstract and formal derivation section: the claim that RF-Sampling exactly performs gradient ascent on the alignment score rests on an unverified equivalence between the sampling update and the gradient of an alignment objective; the derivation must explicitly address discretization error from numerical ODE solvers and non-differentiability introduced by CFG distillation, or quantify the approximation gap.

    Authors: We agree that the current derivation is stated in the continuous-time limit and does not explicitly bound the practical errors. In the revised manuscript we will expand the formal derivation section to (i) derive the gradient-ascent equivalence under the flow ODE, (ii) introduce an error term for discretization using standard numerical ODE solvers (e.g., Euler or Heun), and (iii) analyze the effect of CFG distillation non-differentiability. We will provide a concrete approximation bound showing that the implicit ascent property holds with an additive error controlled by step size and model Lipschitz constants, thereby quantifying the gap rather than claiming exact equivalence. revision: yes

  2. Referee: [Abstract] Abstract: the text-image alignment score underlying the gradient-ascent interpretation is never defined; if this score is computed from quantities internal to the same model, the argument risks circularity and the proof must state the objective function explicitly.

    Authors: We thank the referee for highlighting this omission. The alignment score is defined as the inner product between the text embedding and the image features obtained from the model's velocity prediction at each flow step; this objective is grounded in the pre-trained flow-matching loss and is therefore independent of the reflective sampling procedure itself. In the revised abstract and derivation we will state the objective function explicitly as the maximization of this score and clarify that the linear textual combinations and flow inversion implement an approximate gradient step on it, removing any appearance of circularity. revision: yes

Circularity Check

0 steps flagged

No circularity: formal derivation stands on flow-matching mathematics independent of fitted inputs

full rationale

The paper's central claim rests on a formal derivation showing RF-Sampling performs implicit gradient ascent on an alignment score via linear combination of text embeddings plus flow inversion. This derivation is presented as following directly from the continuous ODE structure of flow models and the properties of CFG-distilled variants (e.g., FLUX), without reducing any predicted quantity to a parameter fitted on the target data or to a self-citation chain. The alignment score is treated as an external objective (text-image consistency), and the proof is not shown to be tautological with the sampling rule itself. No self-definitional steps, fitted-input predictions, or ansatz smuggling via prior author work are exhibited in the derivation chain. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract provides no explicit free parameters, new entities, or ad-hoc axioms; the approach relies on standard flow-matching assumptions and the existence of an alignment score.

axioms (1)
  • domain assumption Flow matching models admit a well-defined inversion operation that preserves the ability to explore prompt-consistent noise spaces.
    Invoked when the method integrates flow inversion with textual combinations.

pith-pipeline@v0.9.0 · 5535 in / 1086 out tokens · 50839 ms · 2026-05-15T15:27:04.695563+00:00 · methodology

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.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes
    ?
    echoes

    ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

    we provide a formal derivation proving that RF-Sampling implicitly performs gradient ascent on the text-image alignment score... ∇xJ(xt) ∝ vθ(xt,c) − vθ(xt,∅)... Δ_RF = δt · [vθ(xt,t,chigh) − vθ(xt−δt,t−δt,clow)]... J(x''t) > J(xt) ⇔ ⟨Δ_RF, ∇xJ(xt)⟩ > 0

  • IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean J_uniquely_calibrated_via_higher_derivative echoes
    ?
    echoes

    ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

    Theorem 2 (Second-Order Optimality)... ΔJ(γ) ≈ γ⟨Δ_RF,∇xJ⟩ − ½γ²|Δ⊤_RF H(xt)Δ_RF|... γ* = ⟨Δ_RF,∇xJ⟩ / |Δ⊤_RF H Δ_RF|

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

54 extracted references · 54 canonical work pages · 9 internal anchors

  1. [1]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 10 684–10 695

    work page 2022

  2. [2]

    B. F. Labs, “Flux,” https://github.com/black-forest-labs/flux, 2024

    work page 2024

  3. [3]

    Flux.1 lite: Distilling flux1.dev for efficient text-to- image generation,

    J. M. Daniel Verdú, “Flux.1 lite: Distilling flux1.dev for efficient text-to- image generation,” 2024

    work page 2024

  4. [4]

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, Z. English, K. Lacey, A. Goodwin, Y . Marek, and R. Rombach, “Scaling rectified flow transformers for high-resolution image synthesis,” 2024. [Online]. Available: https://arxiv.org/abs/2403.03206

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    Diffusion models: A comprehensive survey of methods and applications,

    L. Yang, Z. Zhang, Y . Song, S. Hong, R. Xu, Y . Zhao, W. Zhang, B. Cui, and M.-H. Yang, “Diffusion models: A comprehensive survey of methods and applications,”ACM computing surveys, vol. 56, no. 4, pp. 1–39, 2023

    work page 2023

  6. [6]

    Flow Matching for Generative Modeling

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,”arXiv preprint arXiv:2210.02747, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  7. [8]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inNeural Information Processing Systems. Virtual Event: NeurIPS, Dec. 2020, pp. 6840–6851

    work page 2020

  8. [9]

    Golden noise for diffusion models: A learning framework,

    Z. Zhou, S. Shao, L. Bai, S. Zhang, Z. Xu, B. Han, and Z. Xie, “Golden noise for diffusion models: A learning framework,” 2025. [Online]. Available: https://arxiv.org/abs/2411.09502

    work page arXiv 2025

  9. [10]

    A general framework for inference-time scaling and steering of diffusion models,

    R. Singhal, Z. Horvitz, R. Teehan, M. Ren, Z. Yu, K. McKeown, and R. Ranganath, “A general framework for inference-time scaling and steering of diffusion models,”arXiv preprint arXiv:2501.06848, 2025

    work page arXiv 2025

  10. [12]

    Classifier-Free Diffusion Guidance

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  11. [13]

    Zigzag diffusion sampling: Diffusion models can self-improve via self-reflection,

    L. Bai, S. Shao, Z. Qi, H. Xiong, Z. Xieet al., “Zigzag diffusion sampling: Diffusion models can self-improve via self-reflection,” inThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  12. [14]

    Guidance matters: Rethinking the evaluation pitfall for text-to-image generation,

    D. Xie, S. Shao, L. Bai, Z. Zhou, B. Cheng, S. Yang, J. Wu, and Z. Xie, “Guidance matters: Rethinking the evaluation pitfall for text-to-image generation,” inThe Fourteenth International Conference on Learning Representations

    work page

  13. [15]

    Coreˆ 2: Collect, reflect and refine to generate better and faster,

    S. Shao, Z. Zhou, D. Xie, Y . Fang, T. Ye, L. Bai, and Z. Xie, “Coreˆ 2: Collect, reflect and refine to generate better and faster,”arXiv preprint arXiv:2503.09662, 2025

    work page arXiv 2025

  14. [16]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representations. kigali, rwanda: OpenReview.net, May. 2023

    work page 2023

  15. [17]

    Cfg-zero*: Improved classifier-free guidance for flow matching models,

    W. Fan, A. Y . Zheng, R. A. Yeh, and Z. Liu, “Cfg-zero*: Improved classifier-free guidance for flow matching models,”arXiv preprint arXiv:2503.18886, 2025

    work page arXiv 2025

  16. [18]

    On distillation of guided diffusion models,

    C. Meng, R. Rombach, R. Gao, D. Kingma, S. Ermon, J. Ho, and T. Salimans, “On distillation of guided diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 14 297–14 306

    work page 2023

  17. [19]

    The silent assistant: Noisequery as implicit guidance for goal-driven image generation,

    R. Wang, H. Huang, Y . Zhu, O. Russakovsky, and Y . Wu, “The silent assistant: Noisequery as implicit guidance for goal-driven image generation,”arXiv preprint arXiv:2412.05101, 2024

    work page arXiv 2024

  18. [20]

    Synthetic shifts to initial seed vector exposes the brittle nature of latent-based diffusion models,

    M. Po-Yuan, S. Kotyan, T. Y . Foong, and D. V . Vargas, “Synthetic shifts to initial seed vector exposes the brittle nature of latent-based diffusion models,”arXiv preprint arXiv:2312.11473, 2023

    work page arXiv 2023

  19. [21]

    Weak-to-strong diffusion with reflection,

    L. Bai, M. Sugiyama, and Z. Xie, “Weak-to-strong diffusion with reflection,”arXiv preprint arXiv:2502.00473, 2025

    work page arXiv 2025

  20. [22]

    Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications,

    T. Salimans, A. Karpathy, X. Chen, and D. P. Kingma, “Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications,” inInternational Conference on Learning Representations, 2017

    work page 2017

  21. [23]

    Generative pretraining from pixels,

    M. Chen, A. Radford, R. Child, J. Wu, H. Jun, D. Luan, and I. Sutskever, “Generative pretraining from pixels,” inInternational conference on machine learning. PMLR, 2020, pp. 1691–1703

    work page 2020

  22. [24]

    Generative adversarial nets

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. C. Courville, and Y . Bengio, “Generative adversarial nets.” Palais des Congrès de Montréal, Montréal CANADA: NeurIPS, Dec. 2014

    work page 2014

  23. [25]

    Conditional Generative Adversarial Nets

    M. Mirza and S. Osindero, “Conditional generative adversarial nets,” arXiv preprint arXiv:1411.1784, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  24. [26]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” inNeural Information Processing Systems Workshop. Virtual Event: NeurIPS, Dec. 2021

    work page 2021

  25. [27]

    Snapfusion: Text-to-image diffusion model on mobile devices within two seconds,

    Y . Li, H. Wang, Q. Jin, J. Hu, P. Chemerys, Y . Fu, Y . Wang, S. Tulyakov, and J. Ren, “Snapfusion: Text-to-image diffusion model on mobile devices within two seconds,” inThirty-seventh Conference on Neural Information Processing Systems

    work page

  26. [28]

    Guided flows for generative modeling and decision making,

    Q. Zheng, M. Le, N. Shaul, Y . Lipman, A. Grover, and R. T. Chen, “Guided flows for generative modeling and decision making,”arXiv preprint arXiv:2311.13443, 2023

    work page arXiv 2023

  27. [29]

    Generative modeling by estimating gradients of the data distribution,

    Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” inNeural Information Processing Systems, vol. 32. NeurIPS, 2019

    work page 2019

  28. [30]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis,

    X. Wu, Y . Hao, K. Sun, Y . Chen, F. Zhu, R. Zhao, and H. Li, “Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis,” 2023

    work page 2023

  29. [31]

    Applying guidance in a limited interval improves sample and distribution quality in diffusion models,

    T. Kynkäänniemi, M. Aittala, T. Karras, S. Laine, T. Aila, and J. Lehtinen, “Applying guidance in a limited interval improves sample and distribution quality in diffusion models,”Advances in Neural Information Processing Systems, 2024

    work page 2024

  30. [32]

    Cfg++: Manifold-constrained classifier free guidance for diffusion models,

    H. Chung, J. Kim, G. Y . Park, H. Nam, and J. C. Ye, “Cfg++: Manifold-constrained classifier free guidance for diffusion models,”

    work page

  31. [33]

    Available: https://arxiv.org/abs/2406.08070

    [Online]. Available: https://arxiv.org/abs/2406.08070

    work page arXiv

  32. [34]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation,

    Y . Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy, “Pick-a-pic: An open dataset of user preferences for text-to-image generation,” 2023

    work page 2023

  33. [35]

    Photorealistic text-to-image diffusion models with deep language understanding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans et al., “Photorealistic text-to-image diffusion models with deep language understanding,”Advances in Neural Information Processing Systems, vol. 35, pp. 36 479–36 494, 2022

    work page 2022

  34. [36]

    Geneval: An object-focused framework for evaluating text-to-image alignment,

    D. Ghosh, H. Hajishirzi, and L. Schmidt, “Geneval: An object-focused framework for evaluating text-to-image alignment,” 2023. [Online]. Available: https://arxiv.org/abs/2310.11513

    work page arXiv 2023

  35. [37]

    T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation,

    K. Huang, K. Sun, E. Xie, Z. Li, and X. Liu, “T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation,”Advances in Neural Information Processing Systems, vol. 36, pp. 78 723–78 747, 2023

    work page 2023

  36. [38]

    Chronomagic-bench: A benchmark for metamor- phic evaluation of text-to-time-lapse video generation,

    S. Yuan, J. Huang, Y . Xu, Y . Liu, S. Zhang, Y . Shi, R. Zhu, X. Cheng, J. Luo, and L. Yuan, “Chronomagic-bench: A benchmark for metamor- phic evaluation of text-to-time-lapse video generation,”arXiv preprint arXiv:2406.18522, 2024

    work page arXiv 2024

  37. [39]

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

    B. F. Labs, S. Batifol, A. Blattmann, F. Boesel, S. Consul, C. Diagne, T. Dockhorn, J. English, Z. English, P. Esser, S. Kulal, K. Lacey, Y . Levi, C. Li, D. Lorenz, J. Müller, D. Podell, R. Rombach, H. Saini, A. Sauer, and L. Smith, “Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,” 2025. [Online]. Available: htt...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  38. [40]

    Imagereward: Learning and evaluating human preferences for text-to- image generation,

    J. Xu, X. Liu, Y . Wu, Y . Tong, Q. Li, M. Ding, J. Tang, and Y . Dong, “Imagereward: Learning and evaluating human preferences for text-to- image generation,” 2023

    work page 2023

  39. [41]

    Improved aesthetic predictor

    C. Schuhmann, “Improved aesthetic predictor.” [Online]. Available: https://github.com/christophschuhmann/improved-aesthetic-predictor

    work page

  40. [42]

    Wan: Open and Advanced Large-Scale Video Generative Models

    T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, J. Zeng, J. Wang, J. Zhang, J. Zhou, J. Wang, J. Chen, K. Zhu, K. Zhao, K. Yan, L. Huang, M. Feng, N. Zhang, P. Li, P. Wu, R. Chu, R. Feng, S. Zhang, S. Sun, T. Fang, T. Wang, T. Gui, T. Weng, T. Shen, W. Lin, W. Wang, W. Wang, W. Zhou, W. Wang, W. Shen, W. Yu, X. Shi, X....

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [43]

    Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models,

    M. Li*, Y . Lin*, Z. Zhang*, T. Cai, X. Li, J. Guo, E. Xie, C. Meng, J.-Y . Zhu, and S. Han, “Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models,” inThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  42. [44]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernsteinet al., “Imagenet large scale visual recognition challenge,”International Journal of Computer Vision, vol. 115, no. 3, pp. 211–252, 2015

    work page 2015

  43. [45]

    Bag of design choices for inference of high-resolution masked generative transformer,

    S. Shao, Z. Zhou, T. Ye, L. Bai, Z. Xu, and Z. Xie, “Bag of design choices for inference of high-resolution masked generative transformer,”

    work page

  44. [46]

    Available: https://arxiv.org/abs/2411.10781 JOURNAL OF LATEX CLASS FILES, VOL

    [Online]. Available: https://arxiv.org/abs/2411.10781 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, FEBRUARY 2026 14

    work page arXiv 2026

  45. [47]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” 2018

    work page 2018

  46. [48]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, J. Hays, P. Perona, D. Ramanan, C. L. Zitnick, and P. Dollár, “Microsoft coco: Common objects in context,” 2015

    work page 2015

  47. [49]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” 2021

    work page 2021

  48. [50]

    Improved techniques for training gans,

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen, “Improved techniques for training gans,” inNeural Information Processing Systems, vol. 29. Centre Convencions Internacional Barcelona, Barcelona SPAIN: NeurIPS, Dec. 2016

    work page 2016

  49. [51]

    Scaling Instruction-Finetuned Language Models

    H. W. Chung, L. Hou, S. Longpre, B. Zoph, Y . Tay, W. Fedus, Y . Li, X. Wang, M. Dehghani, S. Brahma, A. Webson, S. S. Gu, Z. Dai, M. Suzgun, X. Chen, A. Chowdhery, A. Castro-Ros, M. Pellat, K. Robinson, D. Valter, S. Narang, G. Mishra, A. Yu, V . Zhao, Y . Huang, A. Dai, H. Yu, S. Petrov, E. H. Chi, J. Dean, J. Devlin, A. Roberts, D. Zhou, Q. V . Le, and...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  50. [52]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,”arXiv preprint arXiv:2209.03003, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  51. [53]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” 2023

    work page 2023

  52. [54]

    CLIPScore: A Reference-free Evaluation Metric for Image Captioning

    J. Hessel, A. Holtzman, M. Forbes, R. L. Bras, and Y . Choi, “Clipscore: A reference-free evaluation metric for image captioning,”arXiv preprint arXiv:2104.08718, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  53. [55]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” inInternational Conference on Learning Representations. kigali, rwanda: OpenReview.net, May. 2023

    work page 2023

  54. [56]

    Inference-time scaling for diffusion models beyond scaling denoising steps,

    N. Ma, S. Tong, H. Jia, H. Hu, Y .-C. Su, M. Zhang, X. Yang, Y . Li, T. Jaakkola, X. Jia, and S. Xie, “Inference-time scaling for diffusion models beyond scaling denoising steps,” 2025. [Online]. Available: https://arxiv.org/abs/2501.09732 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, FEBRUARY 2026 15 APPENDIX A. Benchmark Pick-a-Pic.Pick-a-Pic [33] is an...

    work page arXiv 2025