Pith. sign in

REVIEW 5 major objections 5 minor 23 references

DiffIER reduces diffusion error step by step at inference time.

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 ↗

DiffIER claims that iteratively minimizing the distance between a diffusion model's predicted noise and a random Gaussian sample at each inference step improves generation quality.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection DiffIER is a genuinely new inference-time optimization idea, but the paper's own equations don't support it: a sign error, a circular derivation, and a random target that cannot measure the error it claims to reduce. the 5 major comments →

arxiv 2508.13628 v2 pith:E7PYUFBN submitted 2025-08-19 cs.CV

DiffIER: Optimizing Diffusion Models with Iterative Error Reduction

Ao Chen , Lihe Ding , Tianfan Xue This is my paper
classification cs.CV
keywords diffusion modelsclassifier-free guidancetraining-inference gapinference-time optimizationiterative error reductiontext-to-image generationimage super-resolutiontext-to-speech
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.

The reading

The paper argues that poor conditional generation quality and sensitivity to the classifier-free guidance weight are caused by a "training-inference gap": the model's predicted score function is not the true gradient of the log-likelihood, and these per-step errors accumulate during sampling. To close that gap, it proposes DiffIER, a training-free, plug-and-play optimizer that, at each reverse step, runs gradient descent on the squared difference between the CFG prediction and a freshly drawn Gaussian noise vector. The paper's claim is that this iterative error minimization reduces the accumulated error, improves sample quality, and makes the final output far less dependent on the choice of guidance weight, with experiments across text-to-image generation, image super-resolution, and text-to-speech.

Core claim

The central claim is that the reverse process can be improved without retraining by directly minimizing the per-step approximation error rather than by tuning the guidance weight. The paper shows that when the conditional score error $e_{t,c} = \nabla_{x_t}\log q(x_t|c) - s_\theta(x_t,c)$ is zero, the optimal guidance weight is $\omega^*=1$, and it interprets CFG as a compensation for this nonzero error. Since the true gradient is intractable, DiffIER substitutes a stochastic target: it draws $\epsilon\sim\mathcal{N}(0,I)$ and minimizes $\mathcal{L}=\|\epsilon^{\mathrm{cfg}}_{\theta,\omega}(x_t,c)-\epsilon\|^2$ by gradient descent, claiming this makes the predicted distribution converge towa

What carries the argument

The central object is the training-inference gap, measured as the accumulated squared deviation between the CFG score $\epsilon^{\mathrm{cfg}}_{\theta,\omega}(x_t,c)$ and the true score $\nabla_{x_t}\log q(x_t|c)$. The carrying mechanism is DiffIER's inner-loop update: at each reverse step, sample $\epsilon\sim\mathcal{N}(0,I)$, compute $\mathcal{L}=\|\epsilon^{\mathrm{cfg}}_{\theta,\omega}(x_t,c)-\epsilon\|^2$, and take a step $\epsilon^{\mathrm{cfg}} \leftarrow \epsilon^{\mathrm{cfg}} + \eta\,\nabla_{\epsilon^{\mathrm{cfg}}}\mathcal{L}$ until a convergence threshold; the optimized noise estimate is consumed by the DDIM sampler. Iterating this loop is what the paper claims turns the trainin

Load-bearing premise

The load-bearing premise is that a freshly sampled Gaussian vector $\epsilon\sim\mathcal{N}(0,I)$ is a valid proxy for the actual noise that produced the current sample $x_t$, so that pulling the CFG prediction toward this vector reduces real error rather than injecting fresh randomness.

What would settle it

On a low-dimensional Gaussian diffusion model where the true score $\nabla\log q(x_t)$ is known analytically, run DiffIER and measure the KL divergence between the predicted reverse transition and the true reverse transition before and after the inner-loop update; if the divergence does not decrease, the target-noise proxy is wrong. A simpler image-level check: for one fixed prompt and one fixed $x_t$, run the inner loop with many different random draws of $\epsilon$; if the final generated images differ substantially across draws, the update is injecting noise rather than removing it.

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

If this is right

  • If the central claim is right, the same pretrained diffusion model yields better samples without any retraining, fine-tuning, or additional classifiers.
  • The quality of generated images, text, and audio becomes less sensitive to the CFG weight, reducing the need for per-prompt trial-and-error weight selection.
  • Because the per-step error is optimized to a small value, the accumulated training-inference gap shrinks, which the paper ties to higher fidelity and stability in conditional and unconditional sampling.
  • The method applies uniformly to different diffusion-based pipelines (latent diffusion, super-resolution, and score-based TTS), so it can be dropped into existing inference code as a plug-in.

Where Pith is reading between the lines

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

  • The update rule pushes the CFG prediction toward a fresh random noise vector, so in practice the method behaves like an inference-time denoiser; if the true posterior noise at step $t$ is not Gaussian, the random target may add noise instead of removing it.
  • The derived formula for the optimal weight implies that $\omega^*$ varies with time step and condition; DiffIER could plausibly be combined with dynamic weight schedulers or per-step adaptive weights, a combination the paper does not test.
  • A natural stress test: apply DiffIER to few-step or distilled samplers; if the gap is already small in those models, the optimization should cost almost nothing, and if the random-target assumption fails, the gains should disappear.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper proposes DiffIER, a training-free, inference-time optimization method for diffusion models. At each reverse step, the CFG-augmented noise prediction ε_cfg is iteratively adjusted to minimize L = ||ε_cfg − ε||², where ε is freshly sampled from N(0,I), and the adjusted prediction is then used by the DDIM sampler. The authors argue that this reduces a 'training-inference gap' caused by accumulated score-prediction error, and they derive an optimal CFG weight formula in Sec. 4.1 and a KL-based justification in Sec. 4.2. Experiments cover class-conditional ImageNet generation, text-to-image generation, image super-resolution with StableSR, and text-to-speech with Grad-TTS.

Significance. If the core mechanism were valid, DiffIER would be a simple and attractive plug-and-play enhancement applicable across diffusion-based tasks, with the potential to make CFG results less sensitive to the guidance weight. The paper identifies a real phenomenon—CFG output quality is sensitive to the guidance scale—and the proposed method is training-free and agnostic to the downstream model. However, the central theoretical link is not established: the optimization target is a random Gaussian independent of the actual inference trajectory, the derivation from KL divergence to the final objective is non-rigorous, and Algorithm 1 contains a sign error. As presented, the empirical improvements are not explained by the stated mechanism, and some reported numbers appear internally inconsistent. The contribution therefore does not currently meet the standard for a sound methodological claim.

major comments (5)
  1. [§4.2, Eq. (10) and Supplementary A.2, Eq. (22)] Equation (22) states pθ*(xt−1|xt) = E_{pθ*} q(xt−1|x_{0:T−1/t−1}, xT). This is self-referential: the optimal distribution appears on both sides, and the conditioning variables x_{0:T−1/t−1} are not available at step t during inference. The 'mean-field' decomposition leading to this equation is asserted rather than derived, and the jump from Eq. (22) to the concrete objective L = ||ε_cfg − ε||² in Eq. (12)/(23) is a non-sequitur. This is the load-bearing step for the paper's central claim that accumulated error is minimized, and it is not supported.
  2. [Algorithm 1, line 10 (and Algorithm 2, line 10)] The algorithm claims to minimize L = ||ε_cfg − ε||², but the update is ε_cfg ← ε_cfg + η ∇_{ε_cfg} L. Since ∇_{ε_cfg} L = 2(ε_cfg − ε), this is gradient ascent and increases the distance to ε rather than decreasing it. The objective as written would require ε_cfg ← ε_cfg − η ∇_{ε_cfg} L. This is an internal inconsistency that directly affects the proposed method.
  3. [Algorithm 1, line 8; Eq. (12)] The method samples ε ∼ N(0,I) independently at each inner iteration. During inference, there is no ground-truth noise realization relating x_t to x_0 or x_{t−1}; the reverse trajectory is generated by the model itself. Minimizing ||ε_cfg − ε||² against an independent Gaussian target does not drive ε_cfg toward ∇_{x_t} log q(x_t|c) and does not reduce the error between the model's score estimate and the true score. At best it injects stochasticity. Without an argument identifying the sampled ε with the actual noise behind x_t, the central error-reduction claim fails.
  4. [§4.1, Eq. (9) and Supplementary A.1, Eq. (15)] The optimal-weight formula is a tautology. Because e_{t,c} is defined as ∇_{x_t} log q(x_t|c) − s_θ(x_t,c), equation (15) merely expands a definition and rearranges terms. It does not establish that CFG weight compensates for accumulated inference error, and the formula contains the unknown quantity ∇ log q, so it cannot be used as a prediction. This section may be motivation, but it is presented as a derivation and should be substantially qualified.
  5. [Table 2] The reported numbers are internally inconsistent. In the LPIPS row for LSUN-b, DiffIER is listed as 20.69, which is impossible for a perceptual distance normally in [0,1]; the MUSIQ row for LSUN-b lists DiffIER as 20.96, while StableSR is 58.22 in the same metric. These values suggest column misalignment or typographical errors and undermine the quantitative evidence for the super-resolution claims. The table must be corrected before the empirical results can be assessed.
minor comments (5)
  1. [§4.2] The sentence 'mentioned in Sec.' contains an empty cross-reference.
  2. [Equations (4)–(5)] The symbol ω is used for different roles in classifier guidance and classifier-free guidance. Clarify the relationship, since Eq. (4) has (ω+1) while Eq. (5) uses ω as the CFG scale.
  3. [Supplementary A.2] Algorithm 2 in the supplement is identical to Algorithm 1 in the main text. This duplication should be removed or the supplement should be cross-referenced rather than repeated.
  4. [Table 3] The PESQ differences are very small (about 0.01–0.02). No confidence intervals or significance tests are reported, so the claim of consistent improvement is not strongly supported.
  5. [Figure 2] The caption refers to 'star' and 'triangle' markers, but these are not defined in the text or figure legend.

Circularity Check

2 steps flagged

DiffIER's central derivation is self-referential: the optimal reverse distribution is defined as an expectation over itself, and the final objective minimizes distance to a freshly sampled Gaussian rather than to the true score, so the claimed error reduction is not derived from the stated mathematics.

specific steps
  1. self definitional [Supplementary A.2, Eq. (22); also main text Eq. (10)]
    "Then we reach the conclusion with the Lagrange Multiplier Method: pθ∗ (xt−1|xt) = Epθ∗ q(xt−1|x0:T −1/t−1, xT)"

    The optimal predicted distribution pθ* is defined as an expectation taken under pθ* itself. The right-hand side therefore requires the very distribution being solved for, making Eq. 22 a fixed-point equation rather than a determination of pθ*. The paper never solves this self-consistency; it immediately moves to Eq. 23, so the claimed mathematical derivation of the optimization target is not actually completed.

  2. other [Supplementary A.2, Eq. (23); cf. Algorithm 1, lines 8-9]
    "Considering the setting from (Song et al. 2020), the convergence target of every step should be: L = || ϵcf g θ,ω (xt, c) − ϵ ||^2, here ϵcf g θ,ω (xt, c) means the weight score function with CFG and ϵ ∼ N(0, I)."

    During inference there is no known forward-noise realization ε* relating x_t to x_0; the true error e_t,c = ∇ log q(xt|c) − sθ(xt,c) is independent of the fresh draw ε∼N(0,I). Minimizing ||ε_cfg − ε||^2 therefore drives ε_cfg toward an arbitrary random vector (or, with the +η∇L sign in Algorithm 1 line 10, away from it), not toward ∇ log q(xt|c). The step from Eq. 22 to Eq. 23 is asserted via 'Considering the setting from (Song et al. 2020)', not derived, so the claimed 'iterative error reduction' reduces an objective whose target was defined by the algorithm itself, rather than the training-inference gap defined earlier.

full rationale

The paper's empirical comparisons are against external baselines and do not rely on self-citation, so this is not a self-citation-chain case. However, the derivation chain that is supposed to establish the method is circular/unsupported at two load-bearing points. Eq. 22 defines the optimal reverse transition as an expectation under the same distribution being optimized, a self-referential fixed point. The subsequent simplification to Eq. 12/23 replaces the score error with a distance to an independently sampled Gaussian ε; no valid chain connects the KL/mean-field derivation to this objective. Because the target ε is drawn by the algorithm itself, the reduction of L is by construction relative to that chosen target, not to the theoretical gradient of the log-likelihood. This warrants a 6: partial circularity in the central derivation, while the paper retains some independent empirical content (though the algorithm as written also has a sign error, a correctness concern rather than a circularity).

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The 'training-inference gap' is a re-framing of the known exposure bias / score approximation error; no new physical entity is introduced and the gap has no independent falsifiable handle beyond the paper's own measurements. The central claim rests on the ad hoc equivalence between the proposed objective and KL divergence reduction, which is not established.

free parameters (3)
  • gradient step size eta = 5e-2
    Chosen by hand in Algorithm 1; no sensitivity analysis is provided.
  • convergence threshold = 1e-3
    Chosen by hand in Algorithm 1; the inner-loop termination criterion is not precisely specified.
  • CFG guidance weight omega = not specified per task
    Still an input to DiffIER; the method does not eliminate the need to select omega, it only optimizes the CFG prediction for a given omega.
axioms (4)
  • domain assumption The score network s_theta(xt) is a sufficiently accurate approximation to the gradient of the log-likelihood that the residual error e_{t,c} is well-defined and small.
    Standard assumption in diffusion models, invoked when defining e_{t,c} in Section 4.1 and Eq. 9.
  • ad hoc to paper The optimization objective L = ||epsilon_cfg - epsilon||^2 with epsilon ~ N(0,I) corresponds to minimizing the KL divergence between the predicted and true reverse processes.
    This equivalence is asserted after Eq. 22 in Section 4.2 and Supplementary A.2, but the derivation is circular and the target epsilon is a fresh random sample, not the true noise.
  • ad hoc to paper A mean-field decomposition allows optimizing each time step independently in the KL objective.
    Invoked in Supplementary A.2 via the 'philosophy of mean-field approximation' with no justification for its validity in this setting.
  • domain assumption Reducing the accumulated score error at each step improves final sample quality monotonically.
    Claimed in the introduction and supported only by correlation experiments, not proven or bounded.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffIER: Optimizing Diffusion Models with Iterative Error Reduction." pith.science (2026). https://pith.science/paper/E7PYUFBN

@misc{pith2026250813628,
  author       = {Pith},
  title        = {Pith review of: DiffIER: Optimizing Diffusion Models with Iterative Error Reduction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E7PYUFBN}},
  note         = {Machine review of arXiv:2508.13628}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diffusion models have demonstrated remarkable capabilities in generating high-quality samples and enhancing performance across diverse domains through Classifier-Free Guidance (CFG). However, the quality of generated samples is highly sensitive to the selection of the guidance weight. In this work, we identify a critical ``training-inference gap'' and we argue that it is the presence of this gap that undermines the performance of conditional generation and renders outputs highly sensitive to the guidance weight. We quantify this gap by measuring the accumulated error during the inference stage and establish a correlation between the selection of guidance weight and minimizing this gap. Furthermore, to mitigate this gap, we propose DiffIER, an optimization-based method for high-quality generation. We demonstrate that the accumulated error can be effectively reduced by an iterative error minimization at each step during inference. By introducing this novel plug-and-play optimization framework, we enable the optimization of errors at every single inference step and enhance generation quality. Empirical results demonstrate that our proposed method outperforms baseline approaches in conditional generation tasks. Furthermore, the method achieves consistent success in text-to-image generation, image super-resolution, and text-to-speech generation, underscoring its versatility and potential for broad applications in future research.

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

23 extracted references · 6 canonical work pages · 2 internal anchors

  1. [3]

    training-inference gap

    The proposed algorithm is expressed as: Additional quantitative resuts We also assess our approach for unconditional sampling with the Latent Diffusion Model (LDM) trained on datasets in- cluding CelebA (Liu et al. 2015), LSUN-Church, and LSUN- Bedroom (Yu et al. 2015). We extend our analysis using the Predictor-Corrector sampling strategy from (Song et a...

    work page 2015

  2. [6]

    arXiv preprint arXiv:2406.08070

    CFG++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070. Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei- Fei, L

    Pith/arXiv arXiv

  3. [8]

    Text-to-3D Generation with Bidirectional Diffusion using both 2D and 3D priors

    Text-to-3D Generation with Bidirectional Diffusion using both 2D and 3D priors. arXiv:2312.04963. Dinh, L.; Sohl-Dickstein, J.; and Bengio, S

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [10]

    arXiv preprint arXiv:2207.12598

    Classifier-free diffusion guid- ance. arXiv preprint arXiv:2207.12598. Jiang, Y .; Zhang, Z.; Xue, T.; and Gu, J

    Pith/arXiv arXiv

  5. [11]

    arXiv:2310.10123

    AutoDIR: Au- tomatic All-in-One Image Restoration with Latent Diffusion. arXiv:2310.10123. Karras, T.; Aittala, M.; Kynk¨a¨anniemi, T.; Lehtinen, J.; Aila, T.; and Laine, S

    Pith/arXiv arXiv

  6. [12]

    arXiv preprint arXiv:2305.15583

    Al- leviating exposure bias in diffusion models through sampling with shifted time steps. arXiv preprint arXiv:2305.15583. Liu, R.; Wu, R.; Van Hoorick, B.; Tokmakov, P.; Zakharov, S.; and V ondrick, C. 2023a. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, 9298–9309. Liu, X.; Zhan...

    Pith/arXiv arXiv

  7. [13]

    (sθ(xt, c) − sθ(xt, ∅)) (∇xt log q(xt|c) − sθ(xt, ∅)) (sθ(xt, c) − sθ(xt, ∅))2 # = Ext

    The derivation of this esti- mation lies in the expression: L(ω) ≜ scf g θ,ω (xt, c) − ∇xt log q(xt|c) 2 = ∥ω · sθ(xt, c) + (1 − ω) · sθ(xt, ∅) − ∇xt log q(xt|c)∥2 . (14) By treating ω as a variable, we can obtain the minimum value of L(ω) as follows: ω∗ = arg min ω L(ω) = Ext " (sθ(xt, c) − sθ(xt, ∅)) (∇xt log q(xt|c) − sθ(xt, ∅)) (sθ(xt, c) − sθ(xt, ∅))...

    work page 2020

  8. [14]

    arXiv preprint arXiv:2209.14988

    Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988. Popov, V .; V ovk, I.; Gogoryan, V .; Sadekova, T.; and Kudinov, M

    Pith/arXiv arXiv

  9. [16]

    Gradient-Free Classifier Guidance for Diffusion Model Sampling

    Gradient-Free Classifier Guidance for Diffusion Model Sampling. arXiv preprint arXiv:2411.15393. Sohl-Dickstein, J.; Weiss, E.; Maheswaranathan, N.; and Gan- guli, S. 2015a. Deep unsupervised learning using nonequi- librium thermodynamics. In International conference on machine learning, 2256–2265. pmlr. Sohl-Dickstein, J.; Weiss, E. A.; Maheswaranathan, ...

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [18]

    arXiv preprint arXiv:2011.13456

    Score-based generative model- ing through stochastic differential equations. arXiv preprint arXiv:2011.13456. Text2Speech.org

    Pith/arXiv arXiv 2011

  11. [19]

    In Proceedings of the AAAI conference on artificial intelligence, volume 37, 2555–2563

    Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 2555–2563. Wang, J.; Yue, Z.; Zhou, S.; Chan, K. C.; and Loy, C. C. 2024a. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12): 5929–5949. Wang, X.; Dufour, N...

    Pith/arXiv arXiv

  12. [20]

    arXiv preprint arXiv:2206.02262

    Diffusion-gan: Training gans with diffusion. arXiv preprint arXiv:2206.02262. Wu, J. Z.; Ge, Y .; Wang, X.; Lei, S. W.; Gu, Y .; Shi, Y .; Hsu, W.; Shan, Y .; Qie, X.; and Shou, M. Z

    Pith/arXiv arXiv

  13. [2009]

    In 2009 IEEE conference on computer vision and pattern recognition, 248–255

    Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, 248–255. Ieee. Dhariwal, P.; and Nichol, A

    work page 2009

  14. [2015]

    arXiv preprint arXiv:1506.03365

    Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365. Yuzhe, Y .; Chen, J.; Huang, Z.; Lin, H.; Wang, M.; Dai, G.; and Wang, J. ???? Manifold Constraint Reduces Exposure Bias in Accelerated Diffusion Sampling. In The Thirteenth International Conference on Learning Representations. T...

    Pith/arXiv arXiv 2022

  15. [2016]

    arXiv preprint arXiv:1605.08803

    Density estimation using real nvp. arXiv preprint arXiv:1605.08803. Dou, Z.; and Song, Y

    Pith/arXiv arXiv

  16. [2017]

    In Pro- ceedings of the IEEE conference on computer vision and pattern recognition workshops, 126–135

    Ntire 2017 challenge on single image super-resolution: Dataset and study. In Pro- ceedings of the IEEE conference on computer vision and pattern recognition workshops, 126–135. Arjovsky, M.; Chintala, S.; and Bottou, L

    work page 2017

  17. [2018]

    arXiv preprint arXiv:1801.01973

    A note on the inception score. arXiv preprint arXiv:1801.01973. Bradley, A.; and Nakkiran, P

    Pith/arXiv arXiv

  18. [2019]

    arXiv preprint arXiv:1910.00292

    Generalization in generation: A closer look at exposure bias. arXiv preprint arXiv:1910.00292. Shenoy, R.; Pan, Z.; Balakrishnan, K.; Cheng, Q.; Jeon, Y .; Yang, H.; and Kim, J

    Pith/arXiv arXiv 1910

  19. [2020]

    arXiv preprint arXiv:2010.02502

    Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502. Song, Y .; Sohl-Dickstein, J.; Kingma, D. P.; Kumar, A.; Er- mon, S.; and Poole, B

    Pith/arXiv arXiv 2010

  20. [2022]

    arXiv preprint arXiv:2201.06503

    Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503. Barratt, S.; and Sharma, R

    Pith/arXiv arXiv

  21. [2023]

    arXiv:2211.09800

    Instruct- Pix2Pix: Learning to Follow Image Editing Instructions. arXiv:2211.09800. Chung, H.; Kim, J.; Park, G. Y .; Nam, H.; and Ye, J. C

    Pith/arXiv arXiv

  22. [2024]

    arXiv preprint arXiv:2408.09000

    Classifier-free guidance is a predictor-corrector. arXiv preprint arXiv:2408.09000. Brooks, T.; Holynski, A.; and Efros, A. A

    Pith/arXiv arXiv

  23. [2025]

    Ning, M.; Li, M.; Su, J.; Salah, A

    Luvvoice - Free Text to Speech with AI V oices https://luvvoice.com. Ning, M.; Li, M.; Su, J.; Salah, A. A.; and Ertugrul, I. O. 2023a. Elucidating the exposure bias in diffusion models. arXiv preprint arXiv:2308.15321. Ning, M.; Sangineto, E.; Porrello, A.; Calderara, S.; and Cucchiara, R. 2023b. Input perturbation reduces exposure bias in diffusion mode...

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.