arxiv: 2604.05634 · v1 · submitted 2026-04-07 · 💻 cs.AI

Recognition: 2 theorem links

· Lean Theorem

PECKER: A Precisely Efficient Critical Knowledge Erasure Recipe For Machine Unlearning in Diffusion Models

Zhiyong Ma , Zhitao Deng , Huan Tang , Jialin Chen , Zhijun Zheng , Zhengping Li , Qingyuan Chuai

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:22 UTC · model grok-4.3

classification 💻 cs.AI

keywords machine unlearningdiffusion modelssaliency maskdistillation frameworkclass forgettingconcept forgettinggradient updates

0 comments

The pith

A saliency mask in PECKER focuses gradient updates on key parameters to erase targeted data from diffusion models faster.

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

Machine unlearning removes specific data or concepts from generative models to meet safety and privacy needs. Existing methods for diffusion models often run slowly because their gradient updates spread effort across many parameters that contribute little to forgetting. PECKER places a saliency mask inside a distillation framework so that only the most relevant parameters receive priority updates. This change shortens training while still producing images that match the desired distribution after class or concept removal. A reader would care because practical unlearning becomes feasible for large models without repeated full retraining.

Core claim

Within a distillation framework, PECKER introduces a saliency mask to prioritize updates to parameters that contribute most to forgetting the targeted data, thereby reducing unnecessary gradient computation and shortening overall training time without sacrificing unlearning efficacy. Our method generates samples that unlearn related class or concept more quickly, while closely aligning with the true image distribution on CIFAR-10 and STL-10 datasets, achieving shorter training times for both class forgetting and concept forgetting.

What carries the argument

A saliency mask that selects and prioritizes parameter updates during distillation-based unlearning in diffusion models.

If this is right

Shorter overall training times for both class forgetting and concept forgetting.
Faster generation of samples that have unlearned the targeted class or concept.
Close alignment with the true image distribution on CIFAR-10 and STL-10 after unlearning.
Performance that matches or exceeds that of existing machine unlearning methods.

Where Pith is reading between the lines

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

The same mask idea could be tested in unlearning tasks for non-diffusion generative models.
Energy use for repeated unlearning operations on large models might drop if the mask approach scales.
Combining the mask with adaptive thresholds during training could yield further efficiency gains.

Load-bearing premise

The saliency mask reliably identifies the parameters most responsible for the targeted knowledge without missing essential ones or introducing instability that harms model performance.

What would settle it

Running the unlearning process with identical compute budget but replacing the saliency mask with uniform random parameter selection, then checking whether forgetting quality and image distribution match remain the same on CIFAR-10 and STL-10.

Figures

Figures reproduced from arXiv: 2604.05634 by Huan Tang, Jialin Chen, Qingyuan Chuai, Zhengping Li, Zhijun Zheng, Zhitao Deng, Zhiyong Ma.

**Figure 1.** Figure 1: Overview of PECKER. (1) A pseudo-score network sψ provides datafree supervision; gradients are stopped through sψ during generator updates. (2) We compute saliency gs and a mask ms from a temporarily frozen generator (Eq. 9–10), where red marks masked parameters. (3) Retain batches use full updates, while forget batches update only masked parameters. (4) The resulting generator produces concept-erased out… view at source ↗

**Figure 2.** Figure 2: Sample images during PECKER training. Top: forgetting trajectory for class 0 (plane), gradually replaced by class 1 (automobile in CIFAR-10 or bird in STL-10) across checkpoints. Bottom: 2×5 grids of samples from retained classes at different training steps, where class labels are ordered from 1 to 9 (left-to-right, top-to-bottom) and random seeds are fixed across all grids for consistency. success rate of… view at source ↗

**Figure 3.** Figure 3: The curves of UA and FID for PECKER and SFD in 50,000 training steps on CIFAR-10. The blue lines represent PECKER, while the orange lines correspond to SFD. 5.4 Concept Forgetting Concept forgetting refers to the process of ensuring that a model forgets specific concepts as thoroughly as possible, such that even under prompts associated with the forgotten concepts, it fails to generate the corresponding im… view at source ↗

**Figure 4.** Figure 4: Celebrity forgetting process. Each column uses the same prompt and random seed at different checkpoints (0/20/40/60/80/100k imgs). forgetting tasks to quantitatively assess its performance. Our objective is to evaluate whether our method can achieve faster and more effective results in some extent. 5.4.1 Celebrity Forgetting Evaluation As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: Generated images using various text-to-image diffusion models with prompts formulated as ‘A photo of a <nudity keyword> <human subject>.‘ Sensitive parts are manually censored after generation [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

**Figure 6.** Figure 6: The process of Nudity Forgetting between PECKER and SFD. Due to the difficulty in displaying the entire nudity forgetting process over 300k images, and to balance image quality with the demonstration of the forgetting effects of our method, we selected 0, 20, 60, 80, 100k images to visually compare the entire process. Sensitive parts are manually censored after generation. 5.4.2 Nudity forgetting In the or… view at source ↗

read the original abstract

Machine unlearning (MU) has become a critical technique for GenAI models' safe and compliant operation. While existing MU methods are effective, most impose prohibitive training time and computational overhead. Our analysis suggests the root cause lies in poorly directed gradient updates, which reduce training efficiency and destabilize convergence. To mitigate these issues, we propose PECKER, an efficient MU approach that matches or outperforms prevailing methods. Within a distillation framework, PECKER introduces a saliency mask to prioritize updates to parameters that contribute most to forgetting the targeted data, thereby reducing unnecessary gradient computation and shortening overall training time without sacrificing unlearning efficacy. Our method generates samples that unlearn related class or concept more quickly, while closely aligning with the true image distribution on CIFAR-10 and STL-10 datasets, achieving shorter training times for both class forgetting and concept forgetting.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PECKER adds a saliency mask inside distillation-based unlearning for diffusion models and reports faster runtimes on CIFAR-10 and STL-10, but the claimed efficiency edge rests on an unproven assumption about gradient costs.

read the letter

The core contribution is a named recipe that inserts a saliency mask into an existing distillation unlearning loop so that only the most relevant parameters get updated when erasing a class or concept. On the two small image datasets the authors show shorter wall-clock times than prior methods while keeping generated samples close to the original distribution. That is the practical angle worth noting: it targets a real pain point in machine unlearning for generative models where full retraining or heavy fine-tuning is too slow for compliance use cases. The mask idea itself is not brand new, but packaging it specifically for diffusion unlearning with reported timing numbers is the incremental step the paper actually delivers. The experiments appear to use standard benchmarks and at least compare against a few baselines, which is better than many unlearning abstracts. The main soft spot is the efficiency story. Computing a useful saliency mask typically requires a full backward pass on the forget set; if that step is repeated or uses the same loss as the main update, the dominant cost does not disappear and the net saving shrinks to the optimizer step alone. The abstract gives no detail on whether the mask is a cheap static pre-computation or recomputed each iteration, so the headline claim of “reducing unnecessary gradient computation” needs explicit verification in the full text and code. The datasets are also small, so it remains open whether the same mask strategy holds its advantage on larger diffusion models or more complex concepts. This is the sort of targeted engineering paper that a practitioner in safe generative AI might want to try, but it is not foundational. I would bring the method section to a reading group to see the exact mask formula and timing breakdowns. It is solid enough to send out for peer review so referees can check the gradient accounting and run the numbers themselves.

Referee Report

3 major / 2 minor

Summary. The paper proposes PECKER, a machine unlearning method for diffusion models that operates within a distillation framework. It introduces a saliency mask to prioritize gradient updates on parameters most relevant to the forgetting objective, with the goal of reducing unnecessary computation and shortening training time while preserving unlearning performance. Experiments are claimed to show that the approach matches or exceeds prior methods on CIFAR-10 and STL-10 for both class-level and concept-level forgetting, producing samples that align closely with the target distribution.

Significance. If the efficiency claims hold under rigorous measurement, the work could meaningfully lower the barrier to deploying compliant unlearning in generative models, where training overhead has been a practical obstacle. The saliency-driven selective update idea is a natural extension of importance sampling techniques and, if shown to be both cheap and stable, would be a useful addition to the unlearning toolkit.

major comments (3)

[Abstract] Abstract: the central efficiency claim—that the saliency mask 'reduc[es] unnecessary gradient computation'—is load-bearing for the paper’s contribution, yet the abstract supplies no quantitative metrics, wall-clock timings, baseline comparisons, or error bars. Without these, it is impossible to determine whether the reported shorter training times are real or merely an artifact of unreported implementation details.
[Abstract] Abstract (and implied method section): the description of the saliency mask does not specify its computational cost. Standard saliency scores (gradient magnitude or Fisher information w.r.t. the forgetting loss) require a full backward pass on the targeted data; if the mask is recomputed each step or shares the same loss as the main update, the dominant cost remains unchanged and only the optimizer step is sparsified, which is typically negligible. A concrete accounting of FLOPs or measured wall-clock time before/after masking is required to substantiate the efficiency gain.
[Abstract] Abstract: the claim that unlearning 'closely align[s] with the true image distribution' is stated without reference to any quantitative metric (FID, precision/recall, or membership inference attack success rate). Because the central promise is that efficacy is not sacrificed, the absence of these numbers in the summary of results undermines the ability to evaluate the trade-off.

minor comments (2)

[Abstract] The acronym 'PECKER' is introduced without expansion on first use; a parenthetical definition would improve readability.
[Abstract] The abstract refers to 'prevailing methods' without naming them; a brief parenthetical list of the strongest baselines (e.g., 'compared with [method A] and [method B]') would help readers situate the contribution.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. The comments highlight important aspects of how we present our efficiency and efficacy claims. We have revised the abstract and method sections to provide more quantitative details and clarifications as detailed below.

read point-by-point responses

Referee: [Abstract] Abstract: the central efficiency claim—that the saliency mask 'reduc[es] unnecessary gradient computation'—is load-bearing for the paper’s contribution, yet the abstract supplies no quantitative metrics, wall-clock timings, baseline comparisons, or error bars. Without these, it is impossible to determine whether the reported shorter training times are real or merely an artifact of unreported implementation details.

Authors: We agree with this observation. The abstract in the original submission was concise but lacked specific numbers. In the revised manuscript, we have incorporated key quantitative metrics from our experiments, including wall-clock timings, baseline comparisons, and error bars. These details are now summarized in the abstract, with full results and statistical analysis in Section 4. revision: yes
Referee: [Abstract] Abstract (and implied method section): the description of the saliency mask does not specify its computational cost. Standard saliency scores (gradient magnitude or Fisher information w.r.t. the forgetting loss) require a full backward pass on the targeted data; if the mask is recomputed each step or shares the same loss as the main update, the dominant cost remains unchanged and only the optimizer step is sparsified, which is typically negligible. A concrete accounting of FLOPs or measured wall-clock time before/after masking is required to substantiate the efficiency gain.

Authors: This is a fair critique, and we appreciate the referee pointing out the need for transparency on the mask's overhead. Upon review, the method section would benefit from a more explicit description of the saliency mask computation. In the revised version, we have clarified that the mask is computed once using a single forward-backward pass on a representative subset of the forgetting data and remains fixed thereafter. We have also added a concrete accounting of the computational cost, including FLOPs estimates and wall-clock time measurements before and after applying the mask, demonstrating the overall efficiency gain. This is included in the updated method section. revision: yes
Referee: [Abstract] Abstract: the claim that unlearning 'closely align[s] with the true image distribution' is stated without reference to any quantitative metric (FID, precision/recall, or membership inference attack success rate). Because the central promise is that efficacy is not sacrificed, the absence of these numbers in the summary of results undermines the ability to evaluate the trade-off.

Authors: We concur that referencing quantitative metrics in the abstract would strengthen the summary. Our experiments include metrics such as FID scores, precision/recall, and membership inference attack success rates to demonstrate that unlearned models align well with the target distribution without sacrificing performance. We have updated the abstract to reference these key metrics explicitly, directing readers to the detailed results and comparisons in the experimental section. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper proposes PECKER as an efficient unlearning method via a saliency mask in a distillation framework. The abstract and context contain no equations, derivations, fitted parameters presented as predictions, or self-citations that serve as load-bearing justifications. The efficiency claim rests on the mask's prioritization of updates, but without any reduction of a result to its own inputs by construction, the approach is self-contained. No steps match the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no mathematical formulation, so no free parameters, axioms, or invented entities can be identified.

pith-pipeline@v0.9.0 · 5467 in / 1088 out tokens · 49943 ms · 2026-05-10T19:22:59.382492+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Saliency Scoring and Masking. We temporarily freeze the generator gθ compute gradient saliency under the standard noise-prediction objective and derive a binary mask that pinpoints parameters encoding critical “forget” semantics.
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

PECKER introduces a saliency mask to prioritize updates to parameters that contribute most to forgetting the targeted data

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

33 extracted references · 4 canonical work pages · 1 internal anchor

[1]

In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16

Bau, D., Liu, S., Wang, T., Zhu, J.Y., Torralba, A.: Rewriting a deep generative model. In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16. p. 351–369 (2020)

2020
[2]

In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP 2021)

Cao, N.D., Aziz, W., Titov, I.: Editing factual knowledge in language models. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP 2021). pp. 6491–6506. Association for Computational Lin- guistics (2021)

2021
[3]

Chattopadhay, A., Sarkar, A., Howlader, P., Balasubramanian, V.N.: Gradcam++: Generalizedgradient-basedvisualexplanationsfordeepconvolutionalnetworks.In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 839–847. IEEE (2018)

2018
[4]

In: The Thirteenth Interna- tional Conference on Learning Representations (2025)

Chen, T., Zhang, S., Zhou, M.: Score forgetting distillation: A swift, data-free method for machine unlearning in diffusion models. In: The Thirteenth Interna- tional Conference on Learning Representations (2025)

2025
[5]

In: Proceedings of the 60th Annual Meeting of the As- sociation for Computational Linguistics

Dai, D., Dong, L., Hao, Y., Sui, Z., Chang, B., Wei, F.: Knowledge neurons in pretrained transformers. In: Proceedings of the 60th Annual Meeting of the As- sociation for Computational Linguistics. vol. 1, pp. 8493–8502. Association for Computational Linguistics (2022)

2022
[6]

In: The Twelfth International Conference on Learning Representations (2024)

Fan, C., Liu, J., Zhang, Y., Wong, E., Wei, D., Liu., S.: Salun: Empowering ma- chineunlearningvia gradient-basedweight saliency in both image classification and generation. In: The Twelfth International Conference on Learning Representations (2024)

2024
[7]

In: Proceedings of the International Conference on Learning Rep- resentations (ICLR) (2019)

Frankle, J., Carbin, M.: The lottery ticket hypothesis: Finding sparse, trainable neural networks. In: Proceedings of the International Conference on Learning Rep- resentations (ICLR) (2019)

2019
[8]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2426–2436 (2023) 14 Z. Ma et al

2023
[9]

In: Proceedings of the International Conference on Learning Representations (ICLR) (2016)

Han, S., Mao, H., Dally, W.J.: Deep compression: Compressing deep neural net- works with pruning, trained quantization and huffman coding. In: Proceedings of the International Conference on Learning Representations (ICLR) (2016)

2016
[10]

In: Advances in Neural Information Processing Systems

Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. In: Advances in Neural Information Processing Systems. vol. 36 (2024)

2024
[11]

In: Advances in Neural Information Processing Systems

Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems. p. 6626–6637 (2017)

2017
[12]

In: Advances in Neural Information Processing Systems

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Advances in Neural Information Processing Systems. vol. 33 (2020)

2020
[13]

In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021)

Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021)

2021
[14]

In: 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML)

Kong, Z., Chaudhuri, K.: Data redaction from pre-trained gans. In: 2023 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML). p. 638–677 (2023)

2023
[15]

In: Advances in Neural Informa- tion Processing Systems

Kynkäänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. In: Advances in Neural Informa- tion Processing Systems. vol. 32 (2019)

2019
[16]

In: Advances in Neural Information Processing Systems

Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual as- sociations in gpt. In: Advances in Neural Information Processing Systems. vol. 35, pp. 17359–17372. NeurIPS (2022)

2022
[17]

In: Proceedings of the Interna- tional Conference on Learning Representations (ICLR) (2024)

Patil, V., Hase, P., Bansal, M.: Can sensitive information be deleted from llms? objectives for defending against extraction attacks. In: Proceedings of the Interna- tional Conference on Learning Representations (ICLR) (2024)

2024
[18]

right to be forgotten

Politou, E., Alepis, E., Virvou, M., Patsakis, C.: The "right to be forgotten" in the gdpr: Implementation challenges and potential solutions. Learning and Analytics in Intelligent Systems pp. 41–68 (2022)

2022
[19]

Hierarchical Text-Conditional Image Generation with CLIP Latents

Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text- conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)

work page internal anchor Pith review Pith/arXiv arXiv 2022
[20]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 10684–10695 (2022)

2022
[21]

In: Neural Information Processing Systems

Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Lopes, R.G., Ayan, B.K., Tim Salimans, e.a.: Photorealistic text-toimage diffu- sion models with deep language understanding. In: Neural Information Processing Systems. vol. 35, p. 36479–36494 (2022)

2022
[22]

In: Advances in Neural Information Processing Systems

Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X.: Im- proved techniques for training gans. In: Advances in Neural Information Processing Systems. p. 2234–2242 (2016)

2016
[23]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522– 22531 (2023)

2023
[24]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 22522–22531 (2023)

2023
[25]

SmoothGrad: removing noise by adding noise

Smilkov, D., Nikhil Thorat, B.K., Viegas, F., Wattenberg, M.: Smoothgrad: re- moving noise by adding noise. arXiv preprint arXiv:1706.03825 (2017) PECKER 15

work page Pith review arXiv 2017
[26]

Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

Vedaldi,K.S.A.,Zisserman,A.:Deepinsideconvolutionalnetworks:Visualisingim- age classification models and saliency maps. arXiv preprint arXiv:1312.6034 (2013)

work page Pith review arXiv 2013
[27]

2023 , issue_date =

Yang,L.,Zhang,Z.,Song,Y.,Hong,S.,Xu,R.,Zhao,Y.,Zhang,W.,Cui,B.,Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. ACM Comput. Surv.56(4) (Nov 2023).https://doi.org/10.1145/3626235, https://doi.org/10.1145/3626235

work page doi:10.1145/3626235 2023
[28]

In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Yin, T., Gharbi, M., Zhang, R., Shechtman, E., Durand, F., Freeman, W.T., Park, T.: One-step diffusion with distribution matching distillation. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 6613–6623 (2024)

2024
[29]

In: European conference on computer vision

Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: European conference on computer vision. p. 818–833. Springer (2014)

2014
[30]

ACM Computing Surveys56(4), 1–36 (2023)

Zhang, L., Chen, J., Lee, D.: Machine unlearning for generative ai: A survey and taxonomy. ACM Computing Surveys56(4), 1–36 (2023)

2023
[31]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep fea- tures for discriminative localization. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2921–2929. IEEE (2016)

2016
[32]

In: International Conference on Learn- ing Representations (2025)

Zhou, M., Wang, Z., Zheng, H., Huang, H.: Guided score identity distillation for data-free one-step text-to-image generation. In: International Conference on Learn- ing Representations (2025)

2025
[33]

In: Forty-first International Conference on Machine Learning (2024), https://openreview.net/forum?id=QhqQJqe0Wq

Zhou, M., Zheng, H., Wang, Z., Yin, M., Huang, H.: Score identity distilla- tion: Exponentially fast distillation of pretrained diffusion models for one-step generation. In: Forty-first International Conference on Machine Learning (2024), https://openreview.net/forum?id=QhqQJqe0Wq

2024