Pith. sign in

REVIEW 5 major objections 8 minor 38 references

An LLM policy can learn prompt- and timestep-varying diffusion sampling parameters that beat fixed defaults without touching the diffusion weights.

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

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

T0 review · grok-4.5

2026-07-30 21:00 UTC pith:SVKVQ3HZ

load-bearing objection Solid systems paper: freeze the diffuser, RL-train an LLM to schedule inference knobs—real gains, compositional novelty, reward-judge caveats. the 5 major comments →

arxiv 2607.23488 v1 pith:SVKVQ3HZ submitted 2026-07-26 cs.LG cs.CV

Learning Sampling Parameters for Diffusion Models

Arisrei Lim , Yossi Gandelsman This is my paper
classification cs.LG cs.CV
keywords diffusion modelssampling parametersreinforcement learningGRPOclassifier-free guidanceprompt schedulingtext-to-imageVLM-as-a-judge
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Text-to-image diffusion models leave a lot of quality on the table because people pick one set of sampling settings—prompt wording, negative prompt, guidance scale, noise schedule—and freeze them for every prompt and every denoising step. This paper argues that those choices should instead be predicted per prompt and varied over the trajectory. LeSAMP trains a language model with reinforcement learning so that, given a user prompt, it emits schedules for those parameters; the diffusion backbone stays frozen and is only run at inference. Rewards come from a vision-language judge and from human-preference scores. On Flux and Stable Diffusion 3.5 the method wins more often than default sampling and prior LLM scheduling baselines, and a human study backs the automatic numbers. The practical point is a cheaper, complementary knob: better images by learning how to sample, not by retraining the generator.

Core claim

LeSAMP shows that prompt-conditioned, timestep-varying sampling-parameter policies—especially positive-prompt schedules that grow more detailed over denoising—improve outputs of frozen text-to-image diffusion models, reaching pairwise win rates up to about 68% on human-preference scores and 73% under a VLM judge versus defaults and prior schedulers, with human preference up to about 59% over SAP and default SD 3.5.

What carries the argument

LeSAMP: cast sampling-parameter choice as an MDP whose actions are LLM tokens parsed into schedules (positive/negative prompt, CFG, noise) over denoising spans, then fine-tune the LLM with GRPO so end-of-episode image rewards (VLM pairwise win rate or preference models) improve while the diffusion model stays frozen.

Load-bearing premise

The main training signal—a VLM judge’s pairwise win rate—is assumed to track real human preference closely enough that optimizing it yields genuinely better images rather than judge-specific artifacts.

What would settle it

If independent human raters, on held-out ordinary and contradictory prompts, no longer prefer LeSAMP images over default, SAP, and R2F once the automatic VLM judge is removed from both training and scoring, the central claim fails.

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

If this is right

  • Inference-time sampling policies can raise image quality without backpropagation through the denoiser or changing base model weights.
  • Gradually enriching the positive prompt over timesteps can handle both ordinary user requests and contextually contradictory prompts without concept-substitution tricks.
  • Positive-prompt schedules are the highest-leverage single action among prompt, negative prompt, CFG, and noise on the tested backbones.
  • The same external-policy recipe can be tried on stronger frontier generators while keeping their trained distribution intact.
  • Reward choice is decisive: preference models that favor aesthetics over prompt adherence will steer the policy into reward hacking.

Where Pith is reading between the lines

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

  • If frozen-backbone sampling policies keep working on newer models, labs may ship a small LLM scheduler beside each release instead of only full RL fine-tunes of the denoiser.
  • Joint multi-parameter policies underperforming prompt-only schedules suggests the search space or credit assignment—not just model capacity—is the next bottleneck to study.
  • A two-stage reward (VLM gate on adherence, then an aesthetic scorer) is a natural fix for the HPSv3-style hacking the paper already documents.
  • Training on deliberately hard or contradictory prompts could close the remaining failure cases the authors flag when the train set is ordinary web requests.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 8 minor

Summary. The paper introduces LeSAMP, which treats the choice of diffusion sampling parameters — positive prompt, negative prompt, CFG scale, and noise schedule — as a prompt-conditioned, timestep-varying policy emitted as JSON by an LLM, trained with GRPO against image-level rewards (primarily Gemini 3 Flash pairwise win rate; HPSv3 also studied). The diffusion backbone (SD 3.5, FLUX.1 [dev], and Reve 2.0 in an appendix) remains frozen. On two evaluation suites (Whoops-Hard ∪ ContraBench and a web-collected User Requests set), LeSAMP reports tie-adjusted pairwise win rates of 60–73% against default parameters, SAP, and R2F under Gemini and HPSv3 judges, with a 511-label user study giving win rates of 53.7–59.5%. Ablations show positive-prompt schedules drive most gains; timestep variation beats a static-prompt control (60.10% vs 55.62%); CFG and noise schedules are near or below 50%; joint multi-parameter training does not beat positive-prompt-only. The authors also document HPSv3 reward hacking (App. B) and use it to justify the Gemini-based reward.

Significance. If the results hold, the paper establishes inference-time parameter scheduling as a genuinely useful, low-cost axis for improving frozen diffusion models — complementary to weight-space RL post-training and without test-time search. Notable strengths: (i) a controlled ablation isolating timestep-variation from generic prompt rewriting (static 55.62% vs scheduled 60.10%, §4.1), which is the key mechanistic evidence for the paper's thesis; (ii) per-parameter ablations across two backbones (Table 4); (iii) a held-out preference model (HPSv3) and a 511-label human study as partial checks against judge overfitting; (iv) an unusually candid analysis of reward hacking when training on HPSv3 (App. B, Table 6), including the striking Flux positive-prompt Gemini win rate of 12.50%; and (v) a baseline-strengthening choice (replacing SAP/R2F's GPT-4o with GPT-5.4 mini) that reduces the risk that gains come from a stronger base LLM. The observation that policies learn progressively-detailed prompt schedules — and that concept-substitution is not necessary for contradictory prompts — is a useful negative result.

major comments (5)
  1. [§4, Table 1; §3.4] The headline numbers (68.12% HPSv3, 73.37% Gemini in the abstract; Table 1) are dominated by Gemini-as-judge, which is also the training reward (§3.4). The paper's own controls show this is load-bearing: when training on HPSv3 instead (Table 6), the Gemini win rate for Flux positive prompt collapses to 12.50%, demonstrating that VLM-judge win rates are highly judge-specific and partially hackable. Meanwhile several HPSv3-as-held-out entries in Table 1 have confidence intervals that include or nearly include 50% (e.g., Ours vs SAP Flux, 51.54±8.33; Ours vs SD 3.5 User Requests, 52.08±6.32). The central claim would be substantially strengthened by (a) evaluating with a second VLM judge from a different model family (e.g., GPT- or Claude-class) that was not used in training, and (b) reframing the abstract/results so the lead number is a held-out metric rather than the training reward. As wr
  2. [§4 (User Study), Table 2; App. A, Table 5; App. C] The user study is the main evidence that gains transfer to genuine human preference, but it is under-powered and under-reported: 511 labels from 13 evaluators across two datasets and two comparisons means roughly 120–130 labels per cell, no confidence intervals are given in Table 2 (unlike Table 1), and there is no inter-annotator agreement or per-annotator analysis. Annotators were compensated 'as part of their regular job requirements' (App. C), which deserves more disclosure (are annotators independent of the authors' team?). In App. A, the Reve 2.0 study uses a non-standard tie rule (pairs with 4–6 wins out of 10 are reclassified as ties); this post-hoc binning should be justified, and raw win/loss counts reported alongside. Given win rates of 53.7–59.5%, plausible sampling noise could move several cells to 50%; CIs are needed for the human results to carry the weight the paper place
  3. [§4.1, Table 4; §5 (joint training)] Contribution 2 advertises a 'unified action space' over four parameters, but the ablation shows the effect is essentially carried by the positive prompt. Noise-schedule policies sit at ~50% win rates on both backbones (Table 4: 51.46/48.33 and 51.04/51.25), Flux CFG is below 50% on Gemini (48.85), Flux negative prompt scores 35.00 on HPSv3 against 62.29 on Gemini (Table 4 — an internal contradiction that is not discussed), and joint training over all four parameters does not improve over positive-prompt-only (60.42 vs 60.10 Gemini; §5). The paper should state plainly that the positive-prompt schedule is the only reliably beneficial control under this training setup, and either soften the unified-action-space framing or provide evidence (e.g., longer training, larger action-space exploration) that the other channels can contribute. The Flux negative-prompt HPSv3=35.00 cell in particular n
  4. [§1 (claims), §5; missing evaluation] The paper claims the method 'preserves the base distribution' and is 'complementary to existing post-training methods' (§1, §5), but neither claim is tested. On preservation: reward-hacking mitigation via a frozen backbone is asserted structurally, yet App. B shows the policy itself hacks HPSv3 through prompt rewriting — the frozen backbone does not prevent reward hacking, it merely changes its channel. On complementarity: no experiment combines LeSAMP with any RL post-training method (e.g., Flow-GRPO), so 'complementary' is speculation. Either run the combination experiment on one backbone, or downgrade the claim. Additionally, no diversity or distributional metric (FID, CLIP-diversity, FID against a reference set) is reported anywhere; since positive-prompt schedules progressively add detail, a natural failure mode is homogenization of style toward judge-preferred aesthetics, and the p
  5. [§4 (datasets); App. D] Training data and the User Requests evaluation set are both 'user image generation requests from the web' (App. D). The paper states the evaluation sets are held out, but gives no detail on collection, deduplication, or near-duplicate filtering between the two web-sourced corpora. Since the policy learns a somewhat generic strategy (progressively lengthening prompts; mostly-universal negative prompt template), even partial distributional overlap would inflate User Requests results relative to the more distinctive Whoops-Hard∪ContraBench set — and indeed Gemini win rates on User Requests (60.10–69.27, Table 1) are not clearly higher, but the human-study gap (55.81 vs 59.46, Table 2) goes the other way on different comparisons, making contamination hard to rule out. Please describe the collection pipeline and report a deduplication/near-dup analysis.
minor comments (8)
  1. [App. A, Table 5] Table 5's caption says 'evaluated on User Requests' but the table includes a Whoops-Hard∪ContraBench row; caption or table should be fixed.
  2. [§4.1; App. F] The normalization of policy outputs (CFG values divided by 10, sigmas by 10000) is only mentioned in App. F but is needed to interpret the learned CFG schedule 'near 4 rising to 8' discussed in §4.1; move this to the main text.
  3. [App. D; App. E] The KL schedule for prompt policies (β_KL = 2e-2 for 150 steps, then 1e-8) is abrupt and unmotivated; a sentence on why the KL is effectively switched off after warmup would help reproducibility. Similarly, App. E mentions runs 'stopped early due to crashes or policy collapse' — how frequent was collapse, and were the reported runs selected?
  4. [App. G] App. G notes typos in the output-format specification of the prompts (e.g., empty JSON skeletons in the user prompts for positive prompt, CFG, and noise schedule). Even with constrained decoding and parse penalties, please clarify whether these malformed templates were present during all training runs or only some.
  5. [Tables 2, 3, 6] Tables 2, 3, and 6 lack confidence intervals while Table 1 has them; please report CIs consistently, and state the per-cell label counts for Table 2.
  6. [Figures 2–3] Figure 3's dense prompt text is hard to parse at print size; consider separating prompts into a caption or appendix table. Figure 2 would benefit from indicating which dataset each prompt comes from inline.
  7. [§2; §5] Inference overhead is never quantified: the policy requires an LLM forward pass per prompt at deployment. A latency/cost comparison vs. SAP, R2F, and NPC-style test-time search (§2) would substantiate the 'single generation, no test-time search' advantage.
  8. [§4 (Baselines)] The reader's ledger notes the claim that replacing GPT-4o with GPT-5.4 mini makes baselines 'more competitive'; please also state which LLM SAP/R2F's gains in the original papers used and whether the re-implementation was verified against published numbers.

Circularity Check

0 steps flagged

No definitional circularity; mild train/eval metric overlap on Gemini is standard RL practice and is cross-checked by held-out HPSv3 and humans.

full rationale

LeSAMP’s chain is empirical RL, not a closed derivation: an LLM policy emits sampling-parameter schedules; a frozen diffusion model renders images; external scorers (Gemini VLM-as-a-judge, HPSv3, human raters) assign rewards or preferences; win rates vs defaults/SAP/R2F are measured on held-out prompt sets. Nothing in the MDP, GRPO update, or schedule parameterization forces those win rates by construction—the policy can and does fail on some prompts, and ablations (static vs timestep-varying prompt, per-parameter policies, joint vs single-parameter) are free to underperform. Training primarily on Gemini pairwise win rate and also reporting Gemini tables is ordinary optimize-then-evaluate overlap, not self-definitional circularity or a fitted constant renamed as a prediction; the paper explicitly holds out HPSv3 when training on Gemini, reports human study win rates up to ~59%, and documents HPSv3 reward hacking in Appendix B rather than hiding it. Self-citation to Reve 2.0 is only an extra backbone evaluation, not a load-bearing uniqueness or ansatz claim. No uniqueness theorem, no ansatz smuggled via overlapping authors, and no renaming of a known identity as a first-principles result. Score 1 only for the residual (non-load-bearing) coincidence of primary automatic train and report metric; central claims remain externally falsifiable.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The work is empirical ML systems research. Load-bearing commitments are standard RLHF-style assumptions (preference judges proxy quality), frozen diffusion samplers as environment, and hand-chosen training hyperparameters—not new physical entities. Free parameters are optimization and reward-design choices that shape the learned policy.

free parameters (4)
  • GRPO/PPO hyperparameters (lr 1e-6, G=8, clip ε=0.2, β_KL schedules, 1000 steps, batch 24, fail penalty λ_fail=2) = lr=1e-6; G=8; ε=0.2; β_KL=2e-2→1e-8 (prompts) or 1e-8 (CFG/noise); 1000 steps
    Chosen by authors; determine exploration, stability, and how harshly invalid JSON is punished. Central empirical wins depend on successful optimization under these settings.
  • Primary reward model choice (Gemini 3 Flash pairwise win rate) = Gemini 3 Flash group win rate (main)
    Defines the optimization target; alternative HPSv3 yields different, often prompt-non-adherent policies. Headline automatic metrics partly reflect this choice.
  • Denoising budget and resolution (50 steps, 1024×1024) = T=50, 1024×1024
    Fixes the schedule horizon the policy partitions; results may not transfer to other step counts/samplers without retuning.
  • CFG/noise output normalizations and allowed ranges in prompts = CFG/10 and sigma/10000 normalizations noted in App. F
    Action space bounds (e.g., CFG integers ≥10 or 0; sigma anchors from 10000 downward) constrain what can be learned and how outputs are interpreted.
axioms (4)
  • domain assumption Pairwise VLM or HPS preference scores are valid scalar rewards for improving human-desired image quality and prompt adherence.
    Section 3.4 and experiments treat judge win rate / HPSv3 as training and evaluation truth; Appendix B shows this can fail under HPSv3.
  • domain assumption Standard diffusion/flow sampling with CFG and a noise schedule correctly implements the frozen generator environment.
    Preliminaries §3.1 and problem setting §3.2 embed the diffusion model in the MDP without modifying weights.
  • domain assumption GRPO group-normalized advantages with end-of-episode image reward yield a stable policy gradient for long JSON actions.
    Method §3.1–3.3 adopts DeepSeekMath-style GRPO; every token shares terminal advantage.
  • ad hoc to paper Invalid or unparsable schedules may be assigned a group-relative penalty and dropped when no valid rollout exists.
    Equation (5) defines the failure reward; shapes learning of output format as much as image quality.
invented entities (1)
  • LeSAMP policy (LLM emitting JSON multi-span schedules over c+, c−, w, σ) no independent evidence
    purpose: Map user prompts to timestep-varying sampling parameters without updating diffusion weights.
    Central engineered object; not a new physical entity, but a new learned controller interface defined by the paper’s action schema.

pith-pipeline@v1.2.0-grok45-kimik3 · 23686 in / 3565 out tokens · 70989 ms · 2026-07-30T21:00:21.484321+00:00 · methodology

0 comments
read the original abstract

Text-to-image diffusion models expose many inference-time sampling parameters, including prompts, negative prompts, classifier-free guidance scales, and noise schedules. These parameters are typically manually chosen once and then held fixed across prompts and denoising timesteps, even though different prompts and stages of generation can benefit from different parameter values. We introduce LeSAMP, a framework for learning prompt-conditioned, timestep-varying sampling parameters. We formulate parameter selection as a reinforcement learning problem: Given a user prompt, a large language model is trained to emit schedules for the chosen sampling parameters. We optimize our model using rewards from human preference models and VLM-as-a-judge. We evaluate our model on Flux.1 [dev] and Stable Diffusion 3.5, and find that compared to baselines, LeSAMP has a win rate of up to 68.12% using human preference scores and 73.37% using VLM-as-a-judge. These gains are validated in a user study where we achieve win rates of up to 59.46% over previous baselines. Our results suggest that learned sampling-parameter policies provide a complementary approach to existing post-training methods for improving diffusion model outputs.

Figures

Figures reproduced from arXiv: 2607.23488 by Arisrei Lim, Yossi Gandelsman.

Figure 1
Figure 1. Figure 1: LeSAMP optimization overview. Using GRPO and VLM-as-a-judge preference reward, we train an LLM policy to predict timestep-varying positive prompt, negative prompt, CFG scale, or noise schedule given a user prompt. a VLM-based critic that performs pairwise comparison over generated images. Compared to RL post-training of diffusion weights, this design has two attractive properties. First, the diffusion mode… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison between LeSAMP, SAP, and R2F, applied to SD 3.5 and Flux. Each row [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results from four LeSAMP-trained policies predicting schedules for positive prompt, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison between Reve 2.0 and LeSAMP applied to Reve 2.0. Please zoom in for the [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Example of HPSv3’s bias toward aesthetic but non-prompt-adherent images. The left image [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative CFG and noise schedules emitted by trained policies. The CFG policy for [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

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

38 extracted references · 23 linked inside Pith

  1. [1]

    Denoising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. URLhttps://arxiv.org/abs/2006.11239

    Pith/arXiv arXiv 2020

  2. [2]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations, 2021. URLhttps://arxiv.org/abs/2011.13456

    Pith/arXiv arXiv 2021

  3. [3]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling, 2023. URLhttps://arxiv.org/abs/2210.02747

    Pith/arXiv arXiv 2023

  4. [4]

    High- resolution image synthesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models, 2022. URL https://arxiv.org/ abs/2112.10752

    Pith/arXiv arXiv 2022

  5. [5]

    Scalable diffusion models with transformers, 2023

    William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. URL https://arxiv.org/abs/2212.09748

    Pith/arXiv arXiv 2023

  6. [6]

    Laion-5b: An open large-scale dataset for training next generation image-text models, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models,...

    Pith/arXiv arXiv 2022

  7. [7]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps, 2022. URL https://arxiv.org/abs/2206.00927

    Pith/arXiv arXiv 2022

  8. [8]

    Classifier-free diffusion guidance, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. URL https:// arxiv.org/abs/2207.12598

    Pith/arXiv arXiv 2022

  9. [9]

    Applying guidance in a limited interval improves sample and distribution quality in diffusion models.Advances in Neural Information Processing Systems, 37:122458–122483, 2024

    Tuomas Kynkäänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models.Advances in Neural Information Processing Systems, 37:122458–122483, 2024

    2024

  10. [10]

    Dynamic classifier-free diffusion guidance via online feedback, 2025

    Pinelopi Papalampidi, Olivia Wiles, Ira Ktena, Aleksandar Shtedritski, Emanuele Bugliarello, Ivana Kajic, Isabela Albuquerque, and Aida Nematzadeh. Dynamic classifier-free diffusion guidance via online feedback, 2025. URLhttps://arxiv.org/abs/2509.16131

    arXiv 2025

  11. [11]

    Image generation from contextually-contradictory prompts, 2026

    Saar Huberman, Or Patashnik, Omer Dahary, Ron Mokady, and Daniel Cohen-Or. Image generation from contextually-contradictory prompts, 2026. URLhttps://arxiv.org/abs/ 2506.01929

    arXiv 2026

  12. [12]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv.org/abs/ 2402.03300

    Pith/arXiv arXiv 2024

  13. [13]

    Hpsv3: Towards wide-spectrum human preference score, 2025

    Yuhang Ma, Yunhao Shui, Xiaoshi Wu, Keqiang Sun, and Hongsheng Li. Hpsv3: Towards wide-spectrum human preference score, 2025. URL https://arxiv.org/abs/2508.03789

    Pith/arXiv arXiv 2025

  14. [14]

    Flux.https://github.com/black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/black-forest-labs/flux, 2024

    2024

  15. [15]

    Scaling rectified flow transformers for high-resolution image synthesis, 2024

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. URL https://arxiv.org/abs/ 2403.03206

    Pith/arXiv arXiv 2024

  16. [16]

    Rare-to-frequent: Unlocking compositional generation power of diffusion models on rare concepts with llm guidance, 2025

    Dongmin Park, Sebin Kim, Taehong Moon, Minkyu Kim, Kangwook Lee, and Jaewoong Cho. Rare-to-frequent: Unlocking compositional generation power of diffusion models on rare concepts with llm guidance, 2025. URLhttps://arxiv.org/abs/2410.22376. 11

    arXiv 2025

  17. [17]

    Denoising diffusion implicit models, 2022

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. URLhttps://arxiv.org/abs/2010.02502

    Pith/arXiv arXiv 2022

  18. [18]

    Schedule on the fly: Diffusion time prediction for faster and better image generation, 2025

    Zilyu Ye, Zhiyang Chen, Tiancheng Li, Zemin Huang, Weijian Luo, and Guo-Jun Qi. Schedule on the fly: Diffusion time prediction for faster and better image generation, 2025. URL https://arxiv.org/abs/2412.01243

    Pith/arXiv arXiv 2025

  19. [19]

    Training diffusion models with reinforcement learning, 2024

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning, 2024. URLhttps://arxiv.org/abs/2305.13301

    Pith/arXiv arXiv 2024

  20. [20]

    Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models, 2023

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models, 2023. URL https://arxiv.org/abs/2305. 16381

    2023

  21. [21]

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

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation,

  22. [22]

    Directly fine-tuning diffusion models on differentiable rewards, 2024

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards, 2024. URLhttps://arxiv.org/abs/2309.17400

    Pith/arXiv arXiv 2024

  23. [23]

    Diffusion model alignment using direct preference optimization, 2023

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization, 2023. URLhttps://arxiv.org/abs/2311.12908

    Pith/arXiv arXiv 2023

  24. [24]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model,

  25. [25]

    Flow-grpo: Training flow matching models via online rl, 2025

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-grpo: Training flow matching models via online rl, 2025. URLhttps://arxiv.org/abs/2505.05470

    Pith/arXiv arXiv 2025

  26. [26]

    Optimizing prompts for text-to-image generation,

    Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation,

  27. [27]

    Guiding what not to generate: Automated negative prompting for text-image alignment, 2025

    Sangha Park, Eunji Kim, Yeongtak Oh, Jooyoung Choi, and Sungroh Yoon. Guiding what not to generate: Automated negative prompting for text-image alignment, 2025. URL https: //arxiv.org/abs/2512.07702

    arXiv 2025

  28. [28]

    Common diffusion noise schedules and sample steps are flawed, 2024

    Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed, 2024. URLhttps://arxiv.org/abs/2305.08891

    Pith/arXiv arXiv 2024

  29. [29]

    URLhttps://arxiv.org/abs/2212.09611

    Pith/arXiv arXiv

  30. [30]

    Proximal policy optimization algorithms, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. URLhttps://arxiv.org/abs/1707.06347

    Pith/arXiv arXiv 2017

  31. [31]

    VLM-gated

    Reve Team. The layout bet, June 2026. URLhttps://reve.com. Accessed: 2026-06-03. 12 Dataset Comparison Win rate Whoops-Hard∪ContraBench Ours vs Reve 2.0 55.00 User Requests Ours vs Reve 2.0 68.00 Table 5: Pairwise preference data collected from a user study. We show tie-adjusted win rates of LeSAMP against Reve 2.0, evaluated on User Requests. Diffusion m...

    2026

  32. [32]

    Seyedmorteza Sadat, Otmar Hilliges, and Romann M. Weber. Eliminating oversaturation and artifacts of high guidance scales in diffusion models, 2025. URL https://arxiv.org/abs/ 2410.02416

    Pith/arXiv arXiv 2025

  33. [35]

    Begin with a high-level or placeholder structure that is easy for the model to stabilize

  34. [36]

    Use substitute concepts that align in shape, size, and pose with the final target, and swap them in at the step where the model is ready

  35. [37]

    Replace placeholders with the intended concept as soon as the model can express it

  36. [38]

    step": 0,

    Use at most 4 entries (4 prompts total). Output format: A list of 2-4 entries. Each entry's prompt applies from its`step` until the next entry's`step`(or until the end of the run for the last entry). Each entry is an object with two fields: -`step`: integer denoising step at which this prompt takes effect. The first entry's`step`MUST be 0; subsequent`step...

    2000

  37. [2023]

    URLhttps://arxiv.org/abs/2304.05977

    Pith/arXiv arXiv

  38. [2024]

    URLhttps://arxiv.org/abs/2305.18290

    Pith/arXiv arXiv