Images Speak Louder Than Scores: Failure Mode Escape for Enhancing Generative Quality

Guo-Hua Wang; Jianxin Wu; Jie Shao; Ke Zhu; Minghao Fu

REVIEW 4 major objections 6 minor 54 references

Images Speak Louder Than Scores: Failure Mode Escape for Enhancing Generative Quality

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A training-free method called FaME stores a model's worst-generated images and uses their score trajectories as negative guidance, improving perceptual quality while holding FID roughly constant.

desk verdict A useful empirical trick that improves perceptual quality, but the derivation doesn't survive contact with the equations; worth a referee but needs serious revision. read the letter →

arxiv 2508.09598 v1 pith:52BZTPFE submitted 2025-08-13 cs.CV

Jie Shao , Ke Zhu , Minghao Fu , Guo-hua Wang , Jianxin Wu This is my paper

classification cs.CV

keywords diffusionmodelsclassifier-freeguidanceimagequalityassessmentnegativefailuremodesFIDperceptualNetgeneration

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

The paper argues that distribution metrics like FID hide a persistent failure: state-of-the-art class-conditional diffusion models still produce distorted or blurry images for many classes. It proposes FaME, a training-free and inference-time add-on that first uses an image-quality-assessment model to collect a small pool of low-quality generations and their sampling trajectories, then subtracts the score function of those failure samples during future sampling. Applied to DiT-XL/2 and SiT-XL/2 on ImageNet, FaME raises Q-Align quality scores from 2.30/2.32 to 2.52/2.53 and HyperIQA scores from 0.59/0.57 to 0.63, while FID stays essentially flat (2.27 to 2.18 for DiT-XL/2 and 2.15 to 2.17 for SiT-XL/2). The authors position FaME as a correction to classifier-free guidance, which they claim over-samples low-quality regions and shifts the distribution despite improving aggregate metrics.

What carries the argument

The mechanism is a three-term denoiser combination in the spirit of classifier-free guidance but with a third 'escape' term: $D_f(x,\sigma,c) = (w+f)D_1(x,\sigma,c) + (1-w)D_0(x,\sigma) - fD_1(x,\sigma_{f,c},c)$. The negative term re-uses the model's own conditional denoiser evaluated at a noise scale $\sigma_{f,c}$ sampled uniformly from stored trajectories of low-quality images, approximating the score of the failure distribution. A delay hyperparameter $\tau$, borrowed from limited-interval guidance work, controls when the escape term begins during denoising; the default $f=0.02$, $\tau=0.3$ preserves FID while improving quality.

What would settle it

Generate ImageNet class-550 images with DiT-XL/2 plus FaME but swap the negative-trajectory pool for trajectories obtained from the highest-quality images instead of the lowest; if the SQ gain persists with high-quality negative trajectories, the mechanism is not failure-mode escape. A second check: measure the Wasserstein distance between the FaME output distribution and the training distribution of that class; FaME predicts a reduction relative to CFG, and a null result would refute the distribution-repair claim.

Watch

Extended reading notes

Core claim

The central discovery is that failure modes can be made useful without retraining. After sampling a batch of images per class and scoring them with an IQA model, FaME keeps the initial noise vectors (and in the accelerated variant, a global pool of eight trajectories across all classes) that led to the worst images. At inference, the sampling ODE is modified by adding a negative guidance term: $D_f(x,\sigma,c) = (w+f)D_1(x,\sigma,c) + (1-w)D_0(x,\sigma) - fD_1(x,\sigma_{f,c},c)$, where $\sigma_{f,c}$ is a noise scale randomly drawn from the stored failure trajectories. This steers the denoising process away from low-quality regions. The paper reports that the resulting images are visually sh

Load-bearing premise

The crucial step is the approximation in Eq 6, which treats the model's denoiser evaluated at a noise scale from a stored failure trajectory as the score of the failure distribution; if that equivalence is wrong, the negative term is not actually steering away from low-quality images.

Editorial extensions

If this is right

If FaME holds up, perceptual quality of diffusion outputs can be improved at zero extra inference cost, making IQA-guided negative samples a practical alternative to hand-written negative prompts.
The method exposes a blind spot of FID: models with excellent FID can still over-produce distorted images, so evaluation suites should report per-class quality scores alongside distribution metrics.
Because FaME works by reusing a model's own denoiser, it transfers across architectures (DiT-XL/2, SiT-XL/2) and shows initial promise for text-to-image (SDXL) without architecture-specific changes.
The $f$ and $\tau$ hyperparameters give users a simple knob to trade diversity or recall for quality, since higher $f$ and earlier $\tau$ raise SQ further at the cost of FID.

Reading between the lines

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

The paper's approximation in Eq 6 suggests a testable generalization: replacing the stored failure trajectory with a trajectory from a random or high-quality image should weaken the quality gain if the method truly targets failure modes; if it does not, the gain may come from a generic variance-reduction effect of extra guidance terms.
The global pool of eight trajectories independent of class hints that failure modes may be partly class-agnostic artifacts; a direct experiment comparing per-class pools versus global pools would clarify whether the method is learning class-specific failure signals or a universal 'bad image' direction.
For text-to-image, FaME could replace textual negative prompts with visual negative examples, which the authors note are easier to specify; an extension would build a library of undesirable aesthetics (blurry, oversaturated, low-resolution) and evaluate against prompt-based negative guidance on compositional benchmarks.

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.

Desk Editor's Note

read the letter

You should know that this paper has a genuinely useful empirical result wrapped in a derivation that doesn't hold up. The method, FaME, is a training-free add-on that improves IQA scores on ImageNet class-conditional diffusion without hurting FID. The new piece is using an IQA model to select low-quality samples and then using the denoiser evaluated at a different noise level as a repulsive term. That's a new application of negative guidance, not a new framework.

What the paper does well: the observation about FID's blindness to per-class quality issues is well-illustrated and true in my experience. The experiments are straightforward, and the SQ/SH gains are consistent across two backbones. The global trajectory pool of 8 samples is a nice trick to keep inference cost unchanged, and the t2i extension with negative images is plausible.

The soft spots are real, and one is load-bearing. Eq 6 asserts that the score of the failure distribution can be approximated by the conditional score at a different noise scale. That's not a distribution change, so it doesn't give you the failure-distribution score. And the stress-test derivation is correct: even if you take Eq 6 as given, converting Eq 8 to denoiser outputs properly gives an additional σ²/σ_f² factor and an x term; Eq 9 drops both. So the implemented update is not the one derived. That's an internal inconsistency, not a quibble. The circular use of Q-Align SQ — selecting failures, tuning f and τ, and then reporting SQ as the headline — is real, but the independent HyperIQA SH gain means it's not the whole story. No code or error bars is a minor issue for this community but still a gap.

Who is this for? Someone looking for a quick empirical way to boost perceptual quality would get value from trying this; it's simple and seems to work. A theory-minded reader will be frustrated. My own view: the method is promising as a heuristic, but the paper as written does not support the 'failure mode escape' interpretation. It deserves a serious referee, not a desk reject, because the empirical result is reproducible enough and the flaw is in the explanation, not necessarily in the thing.

Send it to peer review with the message that the derivation needs to be redone or the paper needs to be reframed as an empirical guidance trick, and ask for held-out hyperparameter tuning and an independent perceptual study, human or at least a second IQA metric.

Referee Report

4 major / 6 minor

Summary. The paper argues that FID overstates progress because it ignores per-image perceptual quality, and that CFG, while useful, shifts the sampling distribution and can oversample low-quality modes. It proposes FaME, a training-free, inference-time-equal method that uses an IQA model to identify low-quality generated samples, stores their ODE sampling trajectories, and uses the conditional denoiser evaluated at a noise scale drawn from those 'failure trajectories' as a negative-guidance term. On ImageNet, applying FaME to DiT-XL/2 and SiT-XL/2 is reported to raise Q-Align SQ from 2.30/2.32 to 2.52/2.53 and HyperIQA SH from 0.59/0.57 to 0.63, with FID roughly unchanged (2.27 to 2.18 and 2.15 to 2.17). The paper also sketches an extension to SDXL for text-to-image generation.

Significance. The paper identifies a real and increasingly recognized problem: aggregate distribution metrics such as FID miss class-dependent and sample-level perceptual artifacts. FaME is a plausible training-free and inference-cost-neutral recipe, and the reported HyperIQA improvements provide some evidence that the effect is not solely a Q-Align artifact. The qualitative protocol of pairing CFG and FaME outputs from the same initial noise is also a useful evaluation practice. However, the paper's theoretical formalization—the claim that FaME steers sampling away from failure modes—is not sound as written: the implemented update in Eq. (9) does not match the derived failure-mode-escape score, and the negative term's connection to low-quality regions is not established. The main quality metric SQ is also used both for hyperparameter selection and as the headline result, creating a circularity that needs to be addressed with independent or held-out validation.

major comments (4)

[FaME: Failure Mode Escape, Eqs. (6)-(9)] Eq. (9) does not follow from Eqs. (6) and (8). With the paper's own score estimate s_θ(x,σ) = (Dθ(x,σ)-x)/σ², substituting Eq. (6) into Eq. (8) and converting back to denoiser form gives D_f = (w+f)D1(x,σ,c) + (1−w)D0(x,σ) − f(σ²/σ_f²)(D1(x,σ_f,c)−x). Eq. (9) instead subtracts f·D1(x,σ_f,c), dropping the σ²/σ_f² factor and the +x term. Since σ_f is sampled from stored failure trajectories while σ is the current ODE noise level, σ_f=σ is not generally true. The implemented update is therefore not the score of the failure-mode-escape distribution in Eq. (5); it is not the denoiser of any score of the form (D_f−x)/σ². The central mechanism is unsupported by the derivation as written.
[Eq. (6)] The approximation ∇_x log p_{−1}(x|c;σ) ≈ ∇_x log p(x|c;σ_{f,c}) conflates a noise scale with a distribution. A stored failure trajectory determines an initial noise and a sequence of σ values, but the distribution p_{−1} of 'undesired samples' is not defined by randomly selecting one of those σ values and evaluating the conditional denoiser at the current x. No argument or experiment shows that D1(x,σ_f,c) points away from low-quality regions for a general x. Because this approximation is the only link between stored failure samples and the guidance term, the claim that FaME 'escapes failure modes' requires either a derivation of Eq. (6) or a direct validation of the negative term's direction.
[Experiments, Tables 3-4] The main quality metric SQ is used both to select the hyperparameters (f, τ) in Table 3 and to report the headline improvement in Table 4. This makes the SQ gains partly a selection artifact. The independent HyperIQA SH gain (0.59/0.57 to 0.63) mitigates the concern, but the paper should still validate f and τ on a held-out set or an independent metric, and report confidence intervals for both SQ and SH. Also, Table 4 does not fully support the abstract claim of 'without compromising FID': SiT-XL/2 FID changes from 2.15 to 2.17. The claim should be qualified as within noise or supported with repeated runs.
[Eqs. (10)-(11)] The default FaME variant replaces per-class failure trajectories with a global pool of 8 trajectories sampled from arbitrary classes and applies them to every class. This assumes that a failure trajectory for class c′ provides valid negative guidance for a different class c≠c′. No class-wise analysis is given, so for a given class the negative term may be unrelated to that class's actual low-quality modes. Given that the paper emphasizes class-dependent quality imbalance (Figure 1), this transferability assumption is load-bearing for the default method and should be tested with per-class versus global-pool ablations and pool-size sweeps.

minor comments (6)

[Hyperparameters] The text says 'By default, we set τ = 1.0' but the experimental section adopts τ = 0.3 as the default. Please reconcile.
[Background, Eq. (2)] Eq. (2) uses the notation D_w, but the surrounding text uses D1 and D0. Define the notation once and use it consistently.
[Introduction / Conclusion] Typos: 'butdoes not assess' in the Introduction, and 'somietimes' in the Conclusion.
[Figure 2] The formulas and the +/− direction labels in Figure 2(d) are very small; please enlarge and add a clear legend.
[Table 2] FaMEc is introduced in Table 2 before it is defined in the text; define it at first use.
[Conclusion] The sentence 'we will focus on two aspects' is followed by three listed directions; please adjust the count or split the list.

Circularity Check

1 steps flagged · score 5.0 of 10

SQ-based failure selection and hyperparameter tuning make the headline SQ gain partly self-referential; the independent HyperIQA SH gain provides external grounding, so circularity is moderate.

fitted input called prediction [Section 'FaME: Failure Mode Escape' (Eqs. 6-7) and Section 'Experiments' (Hyperparameters, Table 3)]
"By selecting the lowest-scoring nf,c samples, we obtain a set that we believe approximates the samples drawn from the failure distribution p−1. ... Under the setting f = 0.02, τ = 0.3, the FID score even improves, and the quality score SQ increases significantly, so we adopt this configuration as our default."

The same Q-Align SQ metric is used (i) to choose the failure trajectories that define the negative guidance (lowest-SQ samples), (ii) to select the guidance strength f and interval τ (Table 3), and (iii) as the headline improvement metric (Table 4). The SQ gain is therefore a fitted result rather than an independent prediction: hyperparameters are explicitly chosen because SQ improves, and the negative examples are defined as low-SQ samples. The improvement is partly an artifact of optimizing the evaluation metric itself. The paper's independent HyperIQA SH result (0.59/0.57 -> 0.63) is not used in selection/tuning and provides external support, so the circularity is partial rather than complete.

full rationale

The paper's formal derivation is mostly self-contained: Eq. 5 defines the target distribution, Eq. 8 is its score, and Eq. 6 is an explicit approximation (ansatz) for the unknown failure score, not a disguised input. The main circularity is evaluative: Q-Align SQ is used both to construct the method (failure-mode selection and hyperparameter tuning) and as the primary reported quality gain. This is a fitted-input-called-prediction pattern, but it is mitigated by the independent HyperIQA metric and qualitative comparisons, so the central claim retains external content. The authors themselves acknowledge in the Conclusion that FaME 'relies on existing perceptual quality metric,' which is directly relevant to this self-reference. A separate correctness issue -- Eq. 9 does not follow from Eqs. 6-8 under the stated score-to-denoiser conversion because the σ²/σ_f² factor and additive x term are dropped -- is an internal inconsistency in the derivation, not a circularity, and is noted separately. No self-citations are load-bearing; the cited Q-Align and HyperIQA are external IQA models.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central construction is a reweighted distribution p_f that depends on p_-1, which is unobserved. All the paper's evidence for p_-1's usefulness is mediated by Q-Align scores, the same metric used for selection and evaluation, so the main quantity added by the paper is not independently grounded.

free parameters (4)

f (negative guidance strength) = 0.02 (default, from Table 3 sweep)
Set by scanning SQ/FID on the same ImageNet benchmark; not held out.
tau (start step for FaME guidance) = 0.3 (default from Table 3)
Tuned jointly with f to keep FID stable; no held-out validation.
n_f (global failure trajectory pool size) = 8
Chosen by hand as a memory/quality compromise; no ablation reported.
SQ tier thresholds (2.5, 2.0) = SQ>2.5 high, 2.0-2.5 unstable, <2.0 low
Defined after inspecting the score distribution and images in Figure 1; used to support the qualitative narrative.

assumptions (4)

standard math ODE score-based diffusion formulation and denoiser relation ∇ log p(x;σ) ≈ (Dθ(x;σ)-x)/σ^2
Used in Background Eq 1 and to convert Eq 8 into model-output form Eq 9.
domain assumption Q-Align SQ and HyperIQA SH scores are reliable proxies for human perceptual quality
The paper's central evidence rests on these scores (Section 'The Image Quality Pitfall', Tables 1 and 4); no human study is provided.
ad hoc to paper Eq 6: The score of the failure distribution p_-1 at a current sample can be approximated by the conditional denoiser at a randomly chosen failure noise scale σ_f
This is the load-bearing approximation that generates the negative guidance term; no derivation is given and the equation appears to change noise level rather than distribution.
ad hoc to paper A global pool of 8 failure trajectories from unrelated classes provides valid negative guidance for any class
Eq 10-11 drop the class-conditional failure pool and sample c' uniformly from stored failures; the paper assumes cross-class transfer works, with no ablation.

invented entities (1)

Failure distribution p_-1(x|c;σ)
purpose: A target distribution of undesired low-quality samples that FaME subtracts via negative guidance.
Introduced in Eq 5, never directly measured or characterized; approximated by low-SQ samples from the same IQA used as the outcome metric, so there is no falsifiable handle outside the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Images Speak Louder Than Scores: Failure Mode Escape for Enhancing Generative Quality." pith.science (2026). https://pith.science/paper/52BZTPFE

@misc{pith2026250809598,
  author       = {Pith},
  title        = {Pith review of: Images Speak Louder Than Scores: Failure Mode Escape for Enhancing Generative Quality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/52BZTPFE}},
  note         = {Machine review of arXiv:2508.09598}
}

read the original abstract

Diffusion models have achieved remarkable progress in class-to-image generation. However, we observe that despite impressive FID scores, state-of-the-art models often generate distorted or low-quality images, especially in certain classes. This gap arises because FID evaluates global distribution alignment, while ignoring the perceptual quality of individual samples. We further examine the role of CFG, a common technique used to enhance generation quality. While effective in improving metrics and suppressing outliers, CFG can introduce distribution shift and visual artifacts due to its misalignment with both training objectives and user expectations. In this work, we propose FaME, a training-free and inference-efficient method for improving perceptual quality. FaME uses an image quality assessment model to identify low-quality generations and stores their sampling trajectories. These failure modes are then used as negative guidance to steer future sampling away from poor-quality regions. Experiments on ImageNet demonstrate that FaME brings consistent improvements in visual quality without compromising FID. FaME also shows the potential to be extended to improve text-to-image generation.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 31 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Arjovsky, M.; Chintala, S.; and Bottou, L. 2017. W asserstein Generative Adversarial Networks. In International Conference on Machine Learning, 214--223

work page 2017
[4]

M.; Sun, P.; Shen, X.; Khan, F

Bakr, E. M.; Sun, P.; Shen, X.; Khan, F. F.; Erran Li, L.; and Elhoseiny, M. 2023. HRS-Bench : Holistic, Reliable and Scalable Benchmark for Text-to-Image Models. In IEEE/CVF International Conference on Computer Vision, 19984--19996

work page 2023
[5]

Barratt, S.; and Sharma, R. 2018. A note on the Inception Score . arXiv:1801.01973

arXiv 2018
[6]

J.; Arbel, M.; and Gretton, A

Bi \'n kowski, M.; Sutherland, D. J.; Arbel, M.; and Gretton, A. 2018. Demystifying mmd gans. In International Conference on Learning Representations, 1--15

work page 2018
[7]

Brock, A.; Donahue, J.; and Simonyan, K. 2019. Large Scale GAN Training for High Fidelity Natural Image Synthesis. In International Conference on Learning Representations, 1--11

work page 2019
[8]

Chen, C.; and Mo, J. 2022. IQA-PyTorch : PyTorch Toolbox for Image Quality Assessment. [Online]. Available: https://github.com/chaofengc/IQA-PyTorch

work page 2022

Show all 54 references

[9]

Dao, Q.; Phung, H.; Nguyen, B.; and Tran, A. 2023. Flow matching in latent space. arXiv:2307.08698

2023 arXiv
[10]

Dhariwal, P.; and Nichol, A. 2021. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, 8780 -- 8794

2021
[11]

Esser, P.; Kulal, S.; Blattmann, A.; et al. 2024. Scaling rectified flow Transformers for high-resolution image synthesis. In International Conference on Machine Learning, 12606 -- 12633

2024
[12]

Esser, P.; Rombach, R.; and Ommer, B. 2021. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12873--12883

2021
[13]

Gao, Y.; Guo, H.; Hoang, T.; Huang, W.; Jiang, L.; Kong, F.; Li, H.; Li, J.; Li, L.; Li, X.; et al. 2025. Seedance 1.0: Exploring the Boundaries of Video Generation Models. arXiv preprint arXiv:2506.09113

2025 arXiv
[14]

Ghosh, D.; Hajishirzi, H.; and Schmidt, L. 2023. GENEVAL : An object-focused framework for evaluating text-to-image alignment. In Advances in Neural Information Processing Systems

2023
[15]

C.; and Woods, R

Gonzalez, R. C.; and Woods, R. E. 2006. Digital Image Processing (3rd Edition). USA: Prentice-Hall, Inc. ISBN 013168728X

2006
[16]

Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; et al. 2020. Generative adversarial networks. Communications of the ACM, 63(11): 139–144

2020
[17]

Hessel, J.; Holtzman, A.; Forbes, M.; Le Bras, R.; and Choi, Y. 2021. CLIPS core: A Reference-free Evaluation Metric for Image Captioning. In 2021 Conference on Empirical Methods in Natural Language Processing, 7514--7528

2021
[18]

Heusel, M.; Ramsauer, H.; Unterthiner, T.; Nessler, B.; and Hochreiter, S. 2017. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, 6629–6640

2017
[19]

Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 6840--6851

2020
[20]

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

2022 arXiv
[21]

Hu, X.; Wang, R.; Fang, Y.; Fu, B.; Cheng, P.; and Yu, G. 2024. EllA : Equip diffusion models with LLM for enhanced semantic alignment. arXiv:2403.05135

2024 arXiv
[22]

Huang, K.; Sun, K.; Xie, E.; Li, Z.; and Liu, X. 2023. T2I-CompBench : a comprehensive benchmark for open-world compositional text-to-image generation. In Advances in Neural Information Processing Systems, 78723 -- 78747

2023
[23]

Hyv \"a rinen, A. 2005. Estimation of Non-Normalized Statistical Models by Score Matching. Journal of Machine Learning Research, 6(24): 695--709

2005
[24]

Karras, T.; Aittala, M.; Kynk \"a \"a nniemi, T.; Lehtinen, J.; Aila, T.; and Laine, S. 2024. Guiding a diffusion model with a bad version of itself. Advances in Neural Information Processing Systems, 37: 52996--53021

2024
[25]

Karras, T.; Aittala, M.; Laine, S.; and Aila, T. 2022. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 26565 -- 26577

2022
[26]

Kynk \"a \"a nniemi, T.; Aittala, M.; Karras, T.; Laine, S.; Aila, T.; and Lehtinen, J. 2024. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. Advances in Neural Information Processing Systems, 37: 122458--122483

2024
[27]

F.; Batifol, S.; Blattmann, A.; et al

Labs, B. F.; Batifol, S.; Blattmann, A.; et al. 2025. FLUX. 1 Kontext : Flow Matching for In-Context Image Generation and Editing in Latent Space. arXiv:2506.15742

2025 arXiv
[28]

Li, P.; Yan, S.; Tsai, J.; Zhang, R.; An, R.; Guo, Z.; and Gao, X. 2025. Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Masking. arXiv:2505.20199

2025 arXiv
[29]

T.; Ben-Hamu, H.; Nickel, M.; and Le, M

Lipman, Y.; Chen, R. T.; Ben-Hamu, H.; Nickel, M.; and Le, M. 2023. Flow matching for generative modeling. In International Conference on Learning Representations, 1--13

2023
[30]

Liu, Y.; Zhang, K.; Li, Y.; Yan, Z.; Gao, C.; Chen, R.; Yuan, Z.; Huang, Y.; Sun, H.; Gao, J.; et al. 2024. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177

2024 arXiv
[31]

Lu, C.; Zhou, Y.; Bao, F.; Chen, J.; Li, C.; and Zhu, J. 2022. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in neural information processing systems, 35: 5775--5787

2022
[32]

Lu, C.; Zhou, Y.; Bao, F.; Chen, J.; Li, C.; and Zhu, J. 2025. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. Machine Intelligence Research, 1--22

2025
[33]

S.; Boffi, N

Ma, N.; Goldstein, M.; Albergo, M. S.; Boffi, N. M.; Vanden-Eijnden, E.; and Xie, S. 2024. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In European Conference on Computer Vision, 23--40. Springer

2024
[34]

Ma, N.; Tong, S.; Jia, H.; Hu, H.; Su, Y.-C.; Zhang, M.; Yang, X.; Li, Y.; Jaakkola, T.; Jia, X.; et al. 2025. Inference-time scaling for diffusion models beyond scaling denoising steps. arXiv preprint arXiv:2501.09732

2025 arXiv
[35]

Nie, S.; Zhu, F.; You, Z.; Zhang, X.; Ou, J.; Hu, J.; Zhou, J.; Lin, Y.; Wen, J.-R.; and Li, C. 2025. Large language diffusion models. arXiv preprint arXiv:2502.09992

2025 arXiv
[36]

Peebles, W.; and Xie, S. 2023. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 4195--4205

2023
[37]

Podell, D.; English, Z.; Lacey, K.; Blattmann, A.; Dockhorn, T.; M \"u ller, J.; Penna, J.; and Rombach, R. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952

2023 arXiv
[38]

Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10674--10685

2022
[39]

Sadat, S.; Hilliges, O.; and Weber, R. M. 2025. Eliminating oversaturation and artifacts of high guidance scales in diffusion models. In International Conference on Learning Representations, 1--15

2025
[40]

Sajjadi, M. S. M.; Bachem, O.; Lucic, M.; Bousquet, O.; and Gelly, S. 2018. Assessing generative models via precision and recall. In Advances in Neural Information Processing Systems, 5234–5243

2018
[41]

Shen, D.; Song, G.; Xue, Z.; Wang, F.-Y.; and Liu, Y. 2024. Rethinking the Spatial Inconsistency in Classifier-Free Diffusion Guidance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9370--9379

2024
[42]

Song, J.; Meng, C.; and Ermon, S. 2021. Denoising diffusion implicit models. In International Conference on Learning Representations, 1--12

2021
[43]

Song, Y.; and Ermon, S. 2019. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32

2019
[44]

P.; Kumar, A.; Ermon, S.; and Poole, B

Song, Y.; Sohl-Dickstein, J.; Kingma, D. P.; Kumar, A.; Ermon, S.; and Poole, B. 2021. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 1--12

2021
[45]

Su, S.; Yan, Q.; Zhu, Y.; Zhang, C.; Ge, X.; Sun, J.; and Zhang, Y. 2020. Blindly Assess Image Quality in the Wild Guided by a Self-Adaptive Hyper Network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

2020
[46]

Tian, K.; Jiang, Y.; Yuan, Z.; Peng, B.; and Wang, L. 2024. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37: 84839--84865

2024
[47]

Wan, T.; Wang, A.; Ai, B.; Wen, B.; Mao, C.; Xie, C.-W.; Chen, D.; Yu, F.; Zhao, H.; Yang, J.; et al. 2025. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314

2025 arXiv
[48]

Wang, Z.; Bovik, A.; Sheikh, H.; and Simoncelli, E. 2004. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600--612

2004
[49]

Wu, H.; Zhang, Z.; Zhang, W.; Chen, C.; Liao, L.; Li, C.; Gao, Y.; Wang, A.; Zhang, E.; Sun, W.; et al. 2023. Q-align: Teaching lmms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090

2023 arXiv
[50]

Yang, L.; Tian, Y.; Li, B.; Zhang, X.; Shen, K.; Tong, Y.; and Wang, M. 2025. Mmada: Multimodal large diffusion language models. arXiv preprint arXiv:2505.15809

2025 arXiv
[51]

Yang, L.; Zhang, Z.; Song, Y.; et al. 2023. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1--39

2023
[52]

Yao, J.; Yang, B.; and Wang, X. 2025. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, 15703--15712

2025
[53]

Yu, S.; Kwak, S.; Jang, H.; Jeong, J.; Huang, J.; Shin, J.; and Xie, S. 2024. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940

2024 arXiv
[54]

A.; Shechtman, E.; and Wang, O

Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 586--595

2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Arjovsky, M.; Chintala, S.; and Bottou, L. 2017. W asserstein Generative Adversarial Networks. In International Conference on Machine Learning, 214--223

work page 2017

[4] [4]

M.; Sun, P.; Shen, X.; Khan, F

Bakr, E. M.; Sun, P.; Shen, X.; Khan, F. F.; Erran Li, L.; and Elhoseiny, M. 2023. HRS-Bench : Holistic, Reliable and Scalable Benchmark for Text-to-Image Models. In IEEE/CVF International Conference on Computer Vision, 19984--19996

work page 2023

[5] [5]

Barratt, S.; and Sharma, R. 2018. A note on the Inception Score . arXiv:1801.01973

arXiv 2018

[6] [6]

J.; Arbel, M.; and Gretton, A

Bi \'n kowski, M.; Sutherland, D. J.; Arbel, M.; and Gretton, A. 2018. Demystifying mmd gans. In International Conference on Learning Representations, 1--15

work page 2018

[7] [7]

Brock, A.; Donahue, J.; and Simonyan, K. 2019. Large Scale GAN Training for High Fidelity Natural Image Synthesis. In International Conference on Learning Representations, 1--11

work page 2019

[8] [8]

Chen, C.; and Mo, J. 2022. IQA-PyTorch : PyTorch Toolbox for Image Quality Assessment. [Online]. Available: https://github.com/chaofengc/IQA-PyTorch

work page 2022