Pith. sign in

REVIEW 4 major objections 8 minor 36 references

Fast Direct: Query-Efficient Online Black-box Guidance for Diffusion-model Target Generation

T0 review · 4 major / 8 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that nudging every diffusion noise vector toward the current output's gap to a pseudo-target cuts black-box queries by 6x-44x.

desk verdict Plausible, well-tested black-box guidance for diffusion, but the query-efficiency headline is undercut by sloppy accounting and a missing same-group baseline. read the letter →

arxiv 2502.01692 v5 pith:SQCOJLOR submitted 2025-02-02 cs.LG cs.AI

Kim Yong Tan , Yueming Lyu , Ivor Tsang , Yew-Soon Ong This is my paper · ORCID
classification cs.LGcs.AI
keywords diffusionmodelsblack-boxoptimizationonlineguidancequeryefficiencynoisesequenceGaussianprocesssurrogatetargetgenerationmolecular
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Fast Direct claims that a pre-trained diffusion model can be steered toward an arbitrary black-box target at inference time by repeatedly adding a scaled copy of the gap between the current generated sample and a pseudo-target, $x^* - x_K$, to every noise vector and renormalizing each vector to its original length. This update, called guided noise sequence optimization (GNSO), is argued to stay on the data manifold and to keep working even when the pseudo-target is noisy or off-manifold, which is exactly what an online black-box setting provides. Fast Direct constructs the pseudo-target from queried objective values, either by one gradient step of a Gaussian-process surrogate or by taking the best historical sample, so no gradient of the true objective and no offline dataset are required. On twelve 1024x1024 image prompt-alignment tasks scored by Gemini and six 3D-molecule binding-affinity tasks scored by Vina, the paper reports reaching near-full scores within a 50-batch-query budget, 6x to 10x fewer queries than fine-tuning baselines on images and 11x to 44x fewer on molecules. These results are offered as evidence that black-box guidance of diffusion models can be a cheap inference-time operation instead of a fine-tuning campaign.

What carries the argument

The load-bearing object is Guided Noise Sequence Optimization (GNSO): in each outer iteration, generate $x_K$ from the current noise sequence, then update every $\epsilon_k$ as $\epsilon_k \leftarrow (\epsilon_k + \alpha(x^* - x_K))/\|\epsilon_k + \alpha(x^* - x_K)\| \cdot \|\epsilon_k\|$. The direction $x^* - x_K$ is called the universal direction; using $x_K$ rather than intermediate latents keeps the update tied to the data manifold, and the renormalization preserves the noise norm. For the Gaussian-process variant, Proposition 1 states that with a shift-invariant kernel $k(z_1, z_2) = g(\|z_1 - z_2\|_2)$, the gradient of the GP posterior mean lies in the span of $x, x_1, \ldots, x_n$, so the pseudo-target from one gradient step stays near the data. Fast Direct plugs either that pseudo-target or the historical best sample into GNSO and repeats.

What would settle it

Run Algorithm 1 on a fixed target that is pure Gaussian noise and record $\|x_K - x^*\|$ over $T$ iterations: if the renormalized universal-direction update does not reduce this distance, or if image quality collapses as the target moves off the manifold, then the central guidance mechanism is not doing the work claimed.

Watch

Extended reading notes

Core claim

The central claim is that a universal noise update—adding $\alpha(x^* - x_K)$ to every element of the diffusion noise sequence and rescaling each vector to its original norm—turns a pre-trained diffusion sampler into a guided generator that homes in on a target $x^*$ within tens of iterations. The paper argues this works even when $x^*$ is only a noisy pseudo-target produced by a Gaussian-process surrogate or by the historical best sample, rather than the true optimum. Fast Direct wraps this update in an online loop: generate a batch, query the black-box objective, store the $(x_K, y)$ pairs, form a pseudo-target, and repeat. The reported experiments show near-full Gemini alignment ratings within 50 queries on twelve 1024x1024 prompt-alignment tasks and lower Vina scores on six 3D-molecule tasks, with query-efficiency gains of 6x-10x over DDPO, DPOK, D3PO, and DNO on images and 11x-44x on molecules.

Load-bearing premise

The load-bearing premise is empirical: adding $\alpha(x^* - x_K)$ to every noise vector and renormalizing shifts the generated sample toward $x^*$ without leaving the data manifold, and the paper offers this as an observation rather than a theorem.

Editorial extensions

If this is right

  • Fast Direct never updates the diffusion model's weights, so it avoids the GPU memory and training cost of RL fine-tuning and can be used with any stochastic scheduler.
  • Within a 50-batch-query budget, Fast Direct reaches near-full Gemini alignment ratings on all twelve tested image prompts, while DDPO, DPOK, and D3PO need roughly 200 queries to catch up on the first three tasks.
  • The same inference-time mechanism transfers from 1024x1024 images to 3D molecular structures, where it reports 11x-44x query-efficiency gains over the fine-tuning baselines.
  • The pseudo-target is a plug-in: any method that proposes an updated target from the collected dataset can be inserted into Fast Direct, so GP updates and historical-best updates are instantiations of a broader recipe.
  • With the learned GP frozen, Fast Direct generalizes to unseen prompts, improving compressibility, incompressibility, and aesthetic-quality scores on new animal prompts.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The universal-direction update is not specific to diffusion: any sampling process with a renormalizable latent or noise sequence could in principle be steered the same way, though the paper only tests diffusion samplers.
  • Because the GP pseudo-target gradient lies in the span of previously queried samples, Fast Direct's search is confined to a data-defined subspace; targets outside that span may be unreachable, a ceiling the paper does not address.
  • Replacing the one-step posterior-mean gradient with an acquisition function such as expected improvement could cut queries further, since the paper only explores gradient-descent and historical-best ways to set the pseudo-target.
  • The reported gains assume a consistent black-box scorer (Gemini at temperature zero); with noisy human raters the pseudo-target would be noisier, and while GNSO's robustness to noisy targets suggests the method would cope, the exact efficiency multipliers would likely shrink.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The paper proposes Fast Direct, an inference-time, black-box, query-efficient method for steering pre-trained diffusion models toward a target specified only by an online objective function. The method has two components: (1) GNSO (Algorithm 1), which updates the diffusion noise sequence along the direction from the current generated sample x_K to a target x* and renormalizes, and (2) Fast Direct (Algorithm 2), which repeatedly proposes a pseudo-target using either a GP surrogate or the historical best sample, runs GNSO, and queries the black-box objective. The paper reports that on twelve 1024x1024 image-prompt-alignment tasks with a Gemini 1.5 oracle and six 3D-molecule Vina-score tasks, Fast Direct is 6-10x and 11-44x more batch-query-efficient than DDPO, DPOK, D3PO, and DNO, reaching near-optimal scores within a 50 batch-query budget. The paper also reports ablations, a generalization study to unseen prompts, and a proof that the GP gradient for shift-invariant kernels lies in the span of the data.

Significance. If the central efficiency claim holds, Fast Direct is practically significant: it offers a simple, inference-time, black-box alternative to RL-based fine-tuning of diffusion models, with public code and no requirement of a differentiable objective. The GP-gradient-span proposition (Proposition 1) is a useful theoretical sanity check, and the empirical robustness to noisy pseudo-targets is an interesting phenomenon that supports the online black-box setting. The main caveat is that the query-efficiency multipliers, which are the headline contribution, rest on an ambiguous and apparently inconsistent accounting of what counts as a query and on efficiency numbers computed only for a subset of the tasks.

major comments (4)
  1. [Section 4.1, Appendix D, Algorithm 2] The unit of 'query' is never fixed. Section 4.1 says 'query refers to the input to Gemini', yet the '50 batch query budget' is implemented as N=50 outer iterations, each with B=32 instances, and each instance appears to call Gemini (or, for D3PO, perhaps two images per call, see footnote 6 and Appendix G). Appendix D then describes DNO as spending 16 x 50 = 800 'batch queries' while also stating that DNO uses 32 samples per gradient approximation, which would be 16 x 50 x 32 = 25,600 individual objective evaluations if a query means one oracle input. Because Table 1's denominators N* are computed on this ambiguous basis, the reported 6-10x and 11-44x multipliers cannot be verified from the manuscript. The authors should provide a single, unified ledger of individual oracle evaluations per method, per task, and per iteration, and recompute all efficiency ratios with that ledger.
  2. [Algorithm 2, lines 6-16] Algorithm 2's inner loop 'for t <- 1 to i' means that by outer iteration i each instance performs i GNSO diffusion passes and i noise updates before issuing a single objective query at line 14. After N=50 outer iterations, Fast Direct makes sum_{i=1}^{50} i = 1275 diffusion passes per instance, while each baseline is charged one update per batch query. If Fig. 4 and Table 1 count 'algorithm updates', the comparison is not apples-to-apples because Fast Direct receives quadratic unqueried compute; if they count individual objective evaluations, the paper must state the per-method per-iteration ledger explicitly. The authors should either justify that unqueried diffusion compute is negligible relative to oracle cost for both domains, or report the total compute as a separate axis.
  3. [Table 1 and Abstract] The abstract claims 6x-10x query-efficiency improvement on 'twelve' image tasks and 11x-44x on 'six' molecule tasks, but Table 1 reports N* only for Image Tasks 1-3 and Molecule Tasks 1-2. The text says that the extra 200-query budget was granted only for 'tasks 1 to 3' (images) and 'tasks 1 and 2' (molecules), so the quantitative efficiency multipliers in Table 1 are not computed on all twelve/six tasks. The authors should either provide per-task N* and efficiency ratios for every task, or adjust the Abstract and conclusion claims to the subset actually measured.
  4. [Fig. 4, Fig. 10, Table 1] No error bars, confidence intervals, or significance tests are reported for the central quantitative comparison. The image curves and molecule curves appear to be single-run averages (DNO is averaged over 16 independent trials), and the 'Fast Direct outperforms baseline with 200 queries' claims in Table 1 depend on these trajectories. The authors should report means and variances over at least a few random seeds, or otherwise justify that the observed differences are not within run-to-run variability.
minor comments (8)
  1. [Section 4.1] The phrase '50 batch query budget' is never defined; the paper defines 'query' but not 'batch query'. Please define it explicitly and state how many individual Gemini calls one batch query corresponds to.
  2. [Algorithm 2 and Appendix A] In Appendix A, the generalization procedure says 'Lines 17 and 18 are removed from Algorithm 2', but Algorithm 2 has only 16 lines. Please fix the line numbering or the reference.
  3. [Section 4.1] Typo: 'EularDescreteScheduler' should be 'EulerDiscreteScheduler'.
  4. [Appendix G] The D3PO query prompt says 'Given this two images...' while the main-text definition of a query as a single Gemini input is unclear about whether one query can contain two images; please reconcile footnote 6, Appendix G, and the main-text query definition.
  5. [Figure 5 caption] The caption says 'the 45 common animals that were used in DDPO (Tang et al., 2024)', but the DDPO citation should be Black et al. (2023), not Tang et al. (2024).
  6. [Appendix A] The LAION aesthetics predictor reference is missing ('?'). Please add the proper citation.
  7. [Section 4.1] The GP is said to operate on the SDXL latent space of dimension d=4x128x128, but the main text never states that x_K is the latent representation rather than the decoded image. Please clarify what input is stored in dataset D and used in Eq. (2)-(3).
  8. [Footnote 4 and Table 3] The prompt is called 'deer-eleplant' in the text and figures but 'deer-elephant' in Table 3; please make the label consistent.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported 6-44x query-efficiency gains are an artifact of Algorithm 2's query-counting convention; the central claim reduces by construction to counting outer iterations rather than actual objective queries.

  1. self definitional [Algorithm 2, lines 6-16; Section 4.1 'Experiment Procedure'; Table 1]
    "for t ← 1 to i do ... Query black-box objective score y ← f (xK). ... We use N = 50 iterations to utilize the 50 batch query budget, and set the batch size as B = 32. ... Table 1: batch-query-efficiency gain (200/N∗)"

    Algorithm 2 places the objective query (Line 14) inside the inner loop 'for t ← 1 to i'. For N=50 outer iterations, a single instance therefore executes Σ_{i=1}^{50} i = 1275 objective evaluations, not 50. Yet Section 4.1 reports the experiment as a '50 batch query budget', and Table 1 computes Fast Direct's efficiency gain as 200/N* using N* in this same self-defined outer-iteration unit. The 6-44x query-efficiency multipliers are thus built into the query-counting convention rather than measured against a common query unit; the comparison to baselines is not a measured prediction but a construction of the accounting.

full rationale

The only load-bearing reduction I can exhibit is the query-efficiency comparison. Algorithm 2's query operation sits inside the inner loop 'for t ← 1 to i', so after 50 outer iterations a single Fast Direct instance has made 1275 objective evaluations, while Section 4.1 and Table 1 treat the 50 outer iterations as the entire batch query budget and compute gains 200/N*. The 6-44x multipliers therefore reduce by construction to the choice of query-counting convention. Apart from this accounting step, the paper does not disguise fitting as prediction: GNSO is explicitly presented as an empirical heuristic, with Section 5 deferring theory to future work; Proposition 1 is a proved linear-algebra statement; and the GP pseudo-target is a standard online surrogate rather than a fitted parameter renamed as a prediction. The only self-citation (CASBO, Lyu et al. 2024) appears in a related-work table and is not load-bearing. The remaining weakness, the lack of proof that the universal direction preserves the data manifold, is a correctness and empirical-support concern, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. The pseudo-target is an algorithmic construct. The main free parameters are alpha and the GP lengthscale, which are chosen by hand or from prior work. The load-bearing axiom is the empirical universal-direction update in Algorithm 1, which the authors themselves flag as lacking theoretical principles.

free parameters (4)
  • Step size alpha for image tasks = 80
    Hand-chosen in Section 4.1; controls the magnitude of the noise update. Ablation in Appendix B shows insensitivity on Task 1, but the central query-efficiency results depend on a usable alpha.
  • Step size alpha for molecule tasks = 1e-2
    Hand-chosen in Section 4.2 for Vina minimization; no sensitivity analysis is reported for this domain.
  • GP lengthscale lambda = sqrt(d), d = 4*128*128 latent dim
    Set following Hvarfner et al. (2024), per Section 4.1; affects GP surrogate smoothness and pseudo-target location.
  • Batch size B and number of iterations N = B=32, N=50 (most tasks); N=100 for aesthetic quality
    Experiment budget choices; the inner loop in Algorithm 2 runs t=1..i, so total diffusion passes grow quadratically in N.
assumptions (4)
  • ad hoc to paper Adding alpha*(x* - xK) to each noise vector, then renormalizing, moves the generated sample toward the target on the data manifold without quality loss.
    Core premise of Algorithm 1 and Fast Direct; explicitly empirical, no proof, and the paper states theory is future work (Section 5).
  • standard math GP posterior mean formula with shift-invariant kernels and its differentiability.
    Used to define the pseudo-target (Eq. 2, 3) and Proposition 1.
  • domain assumption The score landscape is smooth enough that a one-step GP gradient descent produces a pseudo-target with lower score.
    Assumed in the Remark after Eq. 2; empirically needed for the algorithm to make progress.
  • domain assumption Gemini 1.5 ratings correlate with text-image alignment, and Vina scores correlate with binding affinity.
    Used as ground-truth objectives in Sections 4.1 and 4.2; the paper does not validate these surrogates against human or experimental data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fast Direct: Query-Efficient Online Black-box Guidance for Diffusion-model Target Generation." pith.science (2026). https://pith.science/paper/SQCOJLOR

@misc{pith2026250201692,
  author       = {Pith},
  title        = {Pith review of: Fast Direct: Query-Efficient Online Black-box Guidance for Diffusion-model Target Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SQCOJLOR}},
  note         = {Machine review of arXiv:2502.01692}
}
abstract

Guided diffusion-model generation is a promising direction for customizing the generation process of a pre-trained diffusion model to address specific downstream tasks. Existing guided diffusion models either rely on training the guidance model with pre-collected datasets or require the objective functions to be differentiable. However, for most real-world tasks, offline datasets are often unavailable, and their objective functions are often not differentiable, such as image generation with human preferences, molecular generation for drug discovery, and material design. Thus, we need an $\textbf{online}$ algorithm capable of collecting data during runtime and supporting a $\textbf{black-box}$ objective function. Moreover, the $\textbf{query efficiency}$ of the algorithm is also critical because the objective evaluation of the query is often expensive in real-world scenarios. In this work, we propose a novel and simple algorithm, $\textbf{Fast Direct}$, for query-efficient online black-box target generation. Our Fast Direct builds a pseudo-target on the data manifold to update the noise sequence of the diffusion model with a universal direction, which is promising to perform query-efficient guided generation. Extensive experiments on twelve high-resolution ($\small {1024 \times 1024}$) image target generation tasks and six 3D-molecule target generation tasks show $\textbf{6}\times$ up to $\textbf{10}\times$ query efficiency improvement and $\textbf{11}\times$ up to $\textbf{44}\times$ query efficiency improvement, respectively. Our implementation is publicly available at: https://github.com/kimyong95/guide-stable-diffusion/tree/fast-direct

Figures

Figures reproduced from arXiv: 2502.01692 by the authors.

Figure 1
Figure 1. Demonstration of guided generation for a given target by Algorithm 1. Column 2 (Update [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The generated images over each number of batch queries on the prompt ”deer-eleplant” [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. The 32 randomly generated images for the prompt ”deer-eleplant” [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (31 more)
Figure 4
Figure 4. Figure 4: The average Gemini rating (from 1 to 5, higher is better) of the generated images over [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Left column: The average objective score of the generated images over each number of [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: The generated images using the unseen prompt ”hippo” in the aesthetic quality task, extra [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: The generated images using the 16 unseen prompts. Top left is the image generated by the [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Left: Gemini rating for different step size settings. Middle: Gemini rating for different [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: The update direction ⃗d = x ∗ −xK′ for K′ ∈ {K, K/2, K/4, K/8}, the generated images quality decrease as the xK′ being more more noisy. D BASELINES DETAILS For DDPO, DPOK, and D3PO, we fine-tune the model to maximize the Gemini rating. We fine-tune the model with 50 ep…
Figure 10
Figure 10. Figure 10: The Vina score (lower is better) of the generated molecules for each number of batch [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: The accumulated Gemini rating (from 1 to 5, higher is better) over the number of batch [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: The accumulated Vina score (lower is better) over the number of batch queries. The [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: The generated images over each number of batch queries on the prompt ”traffic-light”, [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: The generated images over each number of batch queries on the prompt ”apple”, extra [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]
Figure 15
Figure 15. Figure 15: The generated images over each number of batch queries on the prompt ”cyber-dog” for [PITH_FULL_IMAGE:figures/full_fig_p023_15.png]
Figure 16
Figure 16. Figure 16: The generated images over each number of batch queries on the prompt ”puppy-nose” [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]
Figure 17
Figure 17. Figure 17: The generated images over each number of batch queries on the prompt ”robot-plant” for [PITH_FULL_IMAGE:figures/full_fig_p024_17.png]
Figure 18
Figure 18. Figure 18: The generated images over each number of batch queries on the prompt ”ocean” for each [PITH_FULL_IMAGE:figures/full_fig_p024_18.png]
Figure 19
Figure 19. Figure 19: The generated images over each number of batch queries on the prompt ”sand-glass” for [PITH_FULL_IMAGE:figures/full_fig_p025_19.png]
Figure 20
Figure 20. Figure 20: The generated images over each number of batch queries on the prompt ”penguin” for [PITH_FULL_IMAGE:figures/full_fig_p025_20.png]
Figure 21
Figure 21. Figure 21: The generated images over each number of batch queries on the prompt ”basket” for each [PITH_FULL_IMAGE:figures/full_fig_p026_21.png]
Figure 22
Figure 22. Figure 22: The generated images over each number of batch queries on the prompt ”ice-cube” for [PITH_FULL_IMAGE:figures/full_fig_p026_22.png]
Figure 23
Figure 23. Figure 23: The generated images over each number of batch queries on the prompt ”cat-butterfly” [PITH_FULL_IMAGE:figures/full_fig_p027_23.png]
Figure 24
Figure 24. Figure 24: The 32 randomly generated images for the prompt ”traffic-light” guided by Fast Direct [PITH_FULL_IMAGE:figures/full_fig_p028_24.png]
Figure 25
Figure 25. Figure 25: The 32 randomly generated images for the prompt ”apple” guided by Fast Direct by [PITH_FULL_IMAGE:figures/full_fig_p029_25.png]
Figure 26
Figure 26. Figure 26: The 32 randomly generated images for the prompt ”cyber-dog” guided by Fast Direct by [PITH_FULL_IMAGE:figures/full_fig_p030_26.png]
Figure 27
Figure 27. Figure 27: The 32 randomly generated images for the prompt ”puppy-nose” guided by Fast Direct [PITH_FULL_IMAGE:figures/full_fig_p031_27.png]
Figure 28
Figure 28. Figure 28: The 32 randomly generated images for the prompt ”robot-plant” guided by Fast Direct [PITH_FULL_IMAGE:figures/full_fig_p032_28.png]
Figure 29
Figure 29. Figure 29: The 32 randomly generated images for the prompt ”ocean” guided by Fast Direct by [PITH_FULL_IMAGE:figures/full_fig_p033_29.png]
Figure 30
Figure 30. Figure 30: The 32 randomly generated images for the prompt ”sand-glass” guided by Fast Direct by [PITH_FULL_IMAGE:figures/full_fig_p034_30.png]
Figure 31
Figure 31. Figure 31: The 32 randomly generated images for the prompt ”penguin” guided by Fast Direct by [PITH_FULL_IMAGE:figures/full_fig_p035_31.png]
Figure 32
Figure 32. Figure 32: The 32 randomly generated images for the prompt ”basket” guided by Fast Direct by [PITH_FULL_IMAGE:figures/full_fig_p036_32.png]
Figure 33
Figure 33. Figure 33: The 32 randomly generated images for the prompt ”ice-cube” guided by Fast Direct by [PITH_FULL_IMAGE:figures/full_fig_p037_33.png]
Figure 34
Figure 34. Figure 34: The 32 randomly generated images for the prompt ”cat-butterfly” guided by Fast Direct [PITH_FULL_IMAGE:figures/full_fig_p038_34.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 14 canonical work pages

  1. [2]

    An overview of diffusion models: Ap- plications, guided generation, statistical rates and optimization.arXiv preprint arXiv:2404.07771,

    Minshuo Chen, Song Mei, Jianqing Fan, and Mengdi Wang. An overview of diffusion models: Ap- plications, guided generation, statistical rates and optimization.arXiv preprint arXiv:2404.07771,

    arXiv

  2. [5]

    Diffdock: Dif- fusion steps, twists, and turns for molecular docking

    Gabriele Corso, Hannes St ¨ark, Bowen Jing, Regina Barzilay, and Tommi Jaakkola. Diffdock: Dif- fusion steps, twists, and turns for molecular docking. arXiv preprint arXiv:2210.01776,

    arXiv

  3. [6]

    Laurianne David, Amol Thakkar, Roc´ıo Mercado, and Ola Engkvist

    doi: 10.1109/TPAMI.2023.3261988. Laurianne David, Amol Thakkar, Roc´ıo Mercado, and Ola Engkvist. Molecular representations in ai-driven drug discovery: a review and practical guide. Journal of Cheminformatics , 12(1):56,

    arXiv 2023

  4. [9]

    3d equiv- ariant diffusion for target-aware molecule generation and affinity prediction

    Jiaqi Guan, Wesley Wei Qian, Xingang Peng, Yufeng Su, Jian Peng, and Jianzhu Ma. 3d equiv- ariant diffusion for target-aware molecule generation and affinity prediction. arXiv preprint arXiv:2303.03543,

    arXiv

  5. [10]

    Manifold preserving guided diffusion

    Yutong He, Naoki Murata, Chieh-Hsin Lai, Yuhta Takida, Toshimitsu Uesaka, Dongjun Kim, Wei- Hsiang Liao, Yuki Mitsufuji, J Zico Kolter, Ruslan Salakhutdinov, et al. Manifold preserving guided diffusion. arXiv preprint arXiv:2311.16424,

    arXiv

  6. [11]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685,

    arXiv

  7. [12]

    Vanilla bayesian optimization performs great in high dimension

    Carl Hvarfner, Erik Orm Hellsten, and Luigi Nardi. Vanilla bayesian optimization performs great in high dimension. arXiv preprint arXiv:2402.02229,

    arXiv

  8. [13]

    Diffusion models for black- box optimization

    Siddarth Krishnamoorthy, Satvik Mehul Mashkaria, and Aditya Grover. Diffusion models for black- box optimization. arXiv preprint arXiv:2306.07180,

    arXiv

Show all 36 references
  1. [14]

    Derivative-free guidance in continuous and discrete diffusion models with soft value-based decoding

    Xiner Li, Yulai Zhao, Chenyu Wang, Gabriele Scalia, Gokcen Eraslan, Surag Nair, Tommaso Bian- calani, Aviv Regev, Sergey Levine, and Masatoshi Uehara. Derivative-free guidance in continuous and discrete diffusion models with soft value-based decoding. arXiv preprint arXiv:2408.08252,

    arXiv

  2. [15]

    Sdxl-lightning: Progressive adversarial diffusion distillation

    Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxl-lightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929,

    arXiv

  3. [16]

    Covariance-adaptive sequential black-box optimization for diffusion targeted generation.arXiv preprint arXiv:2406.00812,

    Yueming Lyu, Kim Yong Tan, Yew Soon Ong, and Ivor W Tsang. Covariance-adaptive sequential black-box optimization for diffusion targeted generation.arXiv preprint arXiv:2406.00812,

    arXiv

  4. [17]

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

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

    arXiv

  5. [18]

    Aligning text-to- image diffusion models with reward backpropagation

    Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to- image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739,

    arXiv

  6. [19]

    Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36,

    12 Published as a conference paper at ICLR 2025 Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36,

    2025

  7. [20]

    Gem- ini 1.5: Unlocking multimodal understanding across millions of tokens of context

    Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean- baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. Gem- ini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv ...

    arXiv

  8. [21]

    Equivalence between policy gradients and soft q- learning

    John Schulman, Xi Chen, and Pieter Abbeel. Equivalence between policy gradients and soft q- learning. arXiv preprint arXiv:1704.06440,

    arXiv

  9. [23]

    Tuning-free alignment of diffusion models with direct noise optimization

    Zhiwei Tang, Jiangweizhi Peng, Jiasheng Tang, Mingyi Hong, Fan Wang, and Tsung-Hui Chang. Tuning-free alignment of diffusion models with direct noise optimization. arXiv preprint arXiv:2405.18881,

    arXiv

  10. [24]

    Understanding rein- forcement learning-based fine-tuning of diffusion models: A tutorial and review

    Masatoshi Uehara, Yulai Zhao, Tommaso Biancalani, and Sergey Levine. Understanding rein- forcement learning-based fine-tuning of diffusion models: A tutorial and review. arXiv preprint arXiv:2407.13734,

    arXiv

  11. [25]

    We follow DDPO (Black et al.,

    13 Published as a conference paper at ICLR 2025 A M ORE EXPERIMENTS : C OMPRESSIBILITY , INCOMPRESSIBILITY , AND AESTHETIC QUALITY Problem. We follow DDPO (Black et al.,

    2025

  12. [26]

    Objective Function

    to evaluate our algorithm on the three black-box optimization tasks for images: compressibility, incompressibility, and aesthetic quality. Objective Function. For compressibility, we aimed to minimize the compressed JPEG size (MB) of the generated images. For incompressibility...

    2023

  13. [27]

    We set N = 50 batch query budget for compressibility and incompressibility, and N = 100 for aesthetic quality

    The prompt for each instance within a batch is sampled randomly and independently. We set N = 50 batch query budget for compressibility and incompressibility, and N = 100 for aesthetic quality. We use the same pre-trained model, hyper-parameters, and GP settings as in Section ...

    2025

  14. [28]

    Right column: The objective score of the images generated byunseen prompts, which demonstrates the generalization capability

    as the input prompts. Right column: The objective score of the images generated byunseen prompts, which demonstrates the generalization capability. Note that DNO is not applicable to this task. 15 Published as a conference paper at ICLR 2025 (a) Compressibility (b) Incompressi...

    2025

  15. [30]

    7For D3PO, to evaluate 64 images only requires 32 queries calls, see Appendix G

    sampler. 7For D3PO, to evaluate 64 images only requires 32 queries calls, see Appendix G. 18 Published as a conference paper at ICLR 2025 For DNO, we optimize the noise sequence w.r.t the Gemini rating. Recall that the Gemini is a black- box, so we use the non-differentiable m...

    2025

  16. [31]

    ✓ ✗ DDOM (Krishnamoorthy et al., 2023), RCGDM (Yuan et al., 2024), DPO (Wallace et al.,

    2023

  17. [32]

    ✗ ✓ DDPO (Black et al., 2023), DPOK (Fan et al., 2024), D3PO (Yang et al.,

    2023

  18. [33]

    ✓ ✓ Inference-time Guidance LGD (Song et al., 2023), MPGD (He et al., 2023), DNO (2024a) (Karunratanakul et al., 2024), ReNO (Eyring et al.,

    2024

  19. [34]

    Answer in the format: Score=(score), Reason=(reason)

    ✗ ✓ DNO (Tang et al., 2024), CASBO (Lyu et al., 2024), Fast Direct(ours) ✓ ✓ G E XPERIMENT DETAILS We use the following query question for the Gemini: Does the prompt $prompt accurately describe the image? Rate from 1 (inaccurate) to 5 (accurate). Answer in the format: Score=(...

    2024

  20. [35]

    The accumulated plot displays the maximum Gemini rating achieved for each number of batch queries

    20 Published as a conference paper at ICLR 2025 Figure 11: The accumulated Gemini rating (from 1 to 5, higher is better) over the number of batch queries. The accumulated plot displays the maximum Gemini rating achieved for each number of batch queries. Figure 12: The accumula...

    2025

  21. [36]

    22 Published as a conference paper at ICLR 2025 Figure 15: The generated images over each number of batch queries on the prompt ”cyber-dog” for each algorithm

    are given to the baseline methods for demonstration. 22 Published as a conference paper at ICLR 2025 Figure 15: The generated images over each number of batch queries on the prompt ”cyber-dog” for each algorithm. Figure 16: The generated images over each number of batch querie...

    2025

  22. [500]

    are given to the baseline methods for demonstration. 16 Published as a conference paper at ICLR 2025 (a) Pre-trained (b) Compressibility (c) Incompressibility (d) Aesthetic Quality Figure 7: The generated images using the 16 unseen prompts. Top left is the image generated by t...

    2025

  23. [2004]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502,

    2010 arXiv

  24. [2020]

    Aligning optimization trajec- tories with diffusion models for constrained design generation

    11 Published as a conference paper at ICLR 2025 Giorgio Giannone, Akash Srivastava, Ole Winther, and Faez Ahmed. Aligning optimization trajec- tories with diffusion models for constrained design generation. Advances in Neural Information Processing Systems, 36:51830–51861,

    2025

  25. [2021]

    Reno: Enhancing one-step text-to-image models through reward-based noise optimization

    Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. arXiv preprint arXiv:2406.04312,

    arXiv

  26. [2022]

    Directly fine-tuning diffusion models on differentiable rewards

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400,

    arXiv

  27. [2023]

    Training diffusion models with reinforcement learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301,

    arXiv

  28. [2024]

    Diffusion posterior sampling for general noisy inverse problems

    Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687,

    arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.