Memory-Efficient Personalization of Text-to-Image Diffusion Models via Selective Optimization Strategies

Hyoungwoo Park; Jeongho Kim; Seokeon Choi; Sunghyun Park; Sungrack Yun

REVIEW 3 major objections 6 minor 18 references

Memory-Efficient Personalization of Text-to-Image Diffusion Models via Selective Optimization Strategies

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Selective routing between low-resolution backpropagation and high-resolution zeroth-order optimization, gated by a timestep-aware probability, matches full-resolution fine-tuning quality on diffusion models while cutting peak memory by up…

desk verdict Novel scheduling idea that deserves a serious referee, but the missing component ablations and overclaimed 'all metrics' result make the central quality claim unproven as written. read the letter →

arxiv 2507.10029 v2 pith:MFFN7YOV submitted 2025-07-14 cs.CV cs.LG

Seokeon Choi , Sunghyun Park , Hyoungwoo Park , Jeongho Kim , Sungrack Yun This is my paper · ORCID

classification cs.CVcs.LG

keywords text-to-imagediffusionpersonalizationmemory-efficientfine-tuningzeroth-orderoptimizationMeZOlow-resolutionbackpropagationtimestep-awareschedulingon-devicegeneration

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

This paper proposes a way to fine-tune text-to-image diffusion models on devices with limited memory by never running full-resolution backpropagation. It alternates between backpropagation on low-resolution images, which injects subject-specific features cheaply, and zeroth-order optimization on high-resolution images, which refines global structure without storing activations. A dynamic timestep-aware probability function decides which strategy runs at each training step, reserving high-timestep steps for the zeroth-order pass. The authors report that this selective scheme matches or beats full-resolution backpropagation on subject fidelity and text alignment across SD V1.5, SD V2.1, SDXL, and SSD-1B, while cutting peak memory by up to 33.69% on SDXL. If correct, this makes private, on-device personalization practical without changing the inference pipeline.

What carries the argument

The load-bearing mechanism is the dynamic timestep-aware probability (DTAP), a logistic function $p^{\mathrm{ZO}}_{i,t} = 1/(1+e^{-k(t-t_{\mathrm{dyn}}(i))})$ whose midpoint $t_{\mathrm{dyn}}(i)$ slides linearly from $t_{\mathrm{max}}$ to $2t_{\mathrm{mid}}-t_{\mathrm{max}}$ over training. At each step it draws on the diffusion timestep $t$ and training progress $i$ to route the update: low-resolution backpropagation captures subject features, while zeroth-order MeZO on full-resolution images, using accumulated gradient estimates over $N$ random perturbations, repairs structure at high timesteps. The sliding midpoint encodes the paper’s two observations: high timesteps carry global structure, and zeroth-order gradients only become reliable after partial adaptation.

What would settle it

Run the identical personalization protocol with the DTAP curve reversed—negative $k$ so ZO-high is selected early and BP-low late—under the same memory budget. If quality does not degrade, the timestep-aware ordering is not what produces the reported gains. A second check: grid-search $t_{\mathrm{mid}}$ on a held-out subject set; if no $t_{\mathrm{mid}}$ changes the outcome, the structural-timestep assumption plays no load-bearing role.

Watch

Extended reading notes

Core claim

The central claim is that the two weaknesses—BP-low’s structural distortion at high timesteps and ZO-high’s slow convergence from scratch—compensate for each other when orchestrated by timestep and training progress. After BP-low has supplied target-specific cues, ZO-high becomes effective at high timesteps, repairing structure; meanwhile the dynamic probability suppresses BP-low exactly where it overfits. The paper demonstrates this on the DreamBooth benchmark with LoRA rank-4 fine-tuning, showing comparable or better CLIP-I, CLIP-T, and DINO scores than full-resolution backpropagation while reducing peak memory, with the largest reduction on SDXL.

Load-bearing premise

The schedule assumes a fixed relationship between diffusion timestep and structural information, along with fixed hyperparameters ($k=0.05$, $t_{\mathrm{mid}}=750$, $i_{\mathrm{max}}=1000$) that were not validated across the full range of subjects, prompts, models, or resolutions.

Editorial extensions

If this is right

Peak memory during personalization becomes the larger of the two low-memory paths instead of full backpropagation, dropping up to 33.69% on SDXL while keeping quality.
Because the sampling process is untouched, the fine-tuned model runs at inference with no added latency.
The framework allows smaller resize ratios for BP-low with little or no quality loss, giving users a memory-quality dial.
On-device fine-tuning becomes feasible for privacy-sensitive applications, since personal images need not leave the device.
The same DTAP selection transfers across SD V1.5, SD V2.1, SDXL, and SSD-1B without changing hyperparameters, per the paper’s experiments.

Reading between the lines

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

If the timestep–structure correspondence is as general as assumed, the same selective BP-low/ZO-high routing could apply to other diffusion tasks such as video or 3D generation, where activation memory is an even larger bottleneck.
The fixed sigmoid schedule leaves an untested handle: adapting $t_{\mathrm{dyn}}(i)$ online from measured structural drift, rather than a linear schedule, could remove the need for hand-set $k$ and $t_{\mathrm{mid}}$.
The memory ceiling may drop further by also applying zeroth-order steps to the text encoder or by shrinking the perturbation count $N$ late in training, neither of which the paper explores.
A direct implication the authors do not state: the effectiveness of MeZO on diffusion models depends on a warm start, so any memory-first personalization method should spend its backpropagation budget early.

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

Let me give you the short version: the paper has a genuinely new scheduling idea that is well motivated, but the experiments as presented cannot support the strong memory/quality claim because the key ablations are missing. Worth a serious referee, not a desk reject.

What's new: combining low-resolution backprop (BP-low) with high-resolution zeroth-order MeZO, and routing between them via a timestep-dependent probability that shifts over training (DTAP). The two observations in Fig. 1 are plausible, and the sigmoid schedule is a reasonable translation of them. The math for the MeZO gradient estimate is standard and correctly stated. The memory numbers in Table 1 look real: peak memory is max of BP and ZO, and reductions up to 33.69% on SDXL are consistent with the architecture.

The soft spots are real and concentrated in the experiments. Table 1 compares only Ours (at three resize ratios) against BP-high. There is no BP-low-only row, no ZO-high-only row, and no TAP-versus-DTAP comparison. That means the table cannot tell you whether the memory saving comes entirely from BP-low or whether ZO-high and the dynamic schedule actually buy the quality. The qualitative Fig. 3 is not a substitute. Also, no error bars or seeds are reported; the performance deltas are within a few thousandths of a point, so "matches or outperforms" is not established. Worse, the abstract's claim of "improvements across all metrics" on SDXL is contradicted by the table itself: at r=0.750 DINO and CLIP-I are lower than BP-high, and at the other ratios CLIP-T is lower. That overclaim needs to be fixed.

The DTAP hyperparameters (k, tmid) are chosen from observations and not swept; that is a minor issue since it is a design choice rather than a fit to the test set. Implementation details are deferred to an unavailable supplementary, and no code is released; that is a practical limitation for reproduction.

I don't think the core idea is broken. The schedule is a reasonable heuristic, and the combination is novel, so the paper deserves peer review. But the authors need to add the missing ablations, report variance, and tone down the "across all metrics" language before the central claim is believable.

Referee Report

3 major / 6 minor

Summary. This paper proposes a memory-efficient personalization method for text-to-image diffusion models. It alternates between low-resolution backpropagation (BP-low) and high-resolution zeroth-order optimization (ZO-high, via MeZO), where the choice at each training step is governed by a dynamic timestep-aware probability (DTAP) that is a sigmoid function of the diffusion timestep with a midpoint that shifts during training. The authors argue that BP-low injects target-specific features but distorts global structure at high timesteps, while ZO-high preserves structure but converges slowly from scratch, so DTAP uses BP-low early and at low timesteps and ZO-high increasingly at high timesteps as training progresses. Experiments on SD V1.5, V2.1, SDXL, and SSD-1B report DINO/CLIP-I/CLIP-T scores and memory usage, showing comparable or better metrics than full-resolution backpropagation with up to 33.7% lower peak memory on SDXL.

Significance. The core idea is interesting and the paper is clearly written. The MeZO gradient estimator (Eq. 2) and the DTAP schedule (Eq. 5) are stated precisely, and the memory measurements in Table 1 are concrete and consistent with the intended resource savings. If the missing ablations confirm that ZO-high and DTAP contribute beyond BP-low, this would be a solid contribution to on-device personalization. However, the current manuscript's main table compares only Ours against BP-High, and the absence of component ablations, error bars, and sensitivity analysis means the incremental value of the proposed mechanism is not yet established. The strength of the paper is its formulation and memory benchmark; the weakness is the statistical and ablative support.

major comments (3)

[§3.2, Table 1] The quantitative evaluation includes only BP-High and Ours, so no row isolates the effect of ZO-high or of the DTAP schedule. Because the memory savings come entirely from the BP-low branch (the ZO memory column is constant for a given model) and the reported differences in DINO and CLIP-I between Ours and BP-High are as small as 0.001-0.005, Table 1 cannot distinguish whether BP-low alone already achieves the reported quality, whether ZO-high adds nothing, or whether DTAP is the ingredient that preserves structure. Please add quantitative ablations for BP-low alone, ZO-high alone, and fixed TAP, with means and error bars over multiple seeds.
[§3.2, Table 1] No variance or significance information is reported. For instance, on SD V1.5 the DINO score for Ours (r=0.750) is 0.6403 versus 0.6434 for BP-High, and on SDXL the DINO scores across resize ratios span 0.7269 to 0.7373; without seed variance or a paired test it is impossible to claim that 'our method consistently matches or outperforms BP-high' or that SDXL achieves 'improvements across all metrics.' Report at least three seeds with standard deviations and a significance test (e.g., paired t-test or Wilcoxon) for each metric.
[§2.4, Eqs. (3)-(5); §3.3] The DTAP schedule is the central novelty, but its hyperparameters (k=0.05, tmid=750, tstart=tmax, tend=2*tmid-tmax, imax=1000) are justified only by qualitative observations (Fig. 1) and qualitative comparisons (Fig. 3). The last paragraph of §3.3 states that ablation studies and hyperparameter selection are deferred to the supplementary material, but that supplementary document is not part of the submission. The main text therefore lacks any sensitivity analysis for k, tmid, and imax, and lacks a quantitative comparison of DTAP against fixed TAP, random selection, and reversed scheduling. The revision should include these experiments in the main paper (or submit the supplementary with the manuscript), because the claim that DTAP is preferable to simpler schedules is load-bearing.

minor comments (6)

[§2.1] The prompt example 'a scs dog' appears to be a typo; presumably the DreamBooth identifier is 'sks' or a unique token.
[Table 1] The text says peak memory is underlined, but the table has no underlined peak-memory column; please report the max of BP and ZO explicitly.
[§2.3] The loss L in Eq. (2) is not defined; specify that it is the same diffusion reconstruction loss used in Eq. (1), and define that loss (e.g., the standard noise-prediction objective).
[Fig. 2] The caption would benefit from also showing the DTAP curve at i = 0.5*imax, since the text notes that DTAP aligns with TAP at this point.
[Algorithm 1] Line 4 says 'Compute pzo_{i,t} using Eq. 5' but the pseudocode would be clearer if the thresholding rule for random() was specified (e.g., 'if random() > pzo_{i,t} then').
[§3.3] The qualitative comparisons in Fig. 3 do not state whether all strategies use the same number of training steps and same LoRA rank; please make the experimental setting explicit for a fair visual comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DTAP is a fixed design schedule and all reported metrics are external to the schedule's construction.

full rationale

The paper's central claim—combining BP-low and ZO-high under a timestep-aware probabilistic schedule matches BP-high quality with lower memory—does not reduce to its inputs. The DTAP probability in Eq. 5 is defined as a logistic function of the diffusion timestep t and a linearly interpolated midpoint t_dyn(i); its constants (k=0.05, tmid=750, imax=1000) are fixed hyperparameters chosen from the qualitative Observations 1 and 2 (Fig. 1), not fitted to the CLIP-I/DINO/CLIP-T scores in Table 1. No reported evaluation quantity is used to define or re-estimate pzo_i,t, and no load-bearing argument relies on a self-citation. The lack of per-component ablations in Table 1 is a valid evidence-quality criticism—the table cannot isolate the contribution of ZO-high or DTAP—but an omitted ablation is not a circular derivation. There is no equation in which a claimed prediction coincides by construction with a fitted parameter or with the evaluation target, so the correct finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on several hand-chosen hyperparameters (k, tmid, imax, r, N, learning rates) and on three domain assumptions about diffusion timestep semantics and the effectiveness of zeroth-order optimization. No new physical entities or exotic constructs are introduced. The DTAP scheduling function is a heuristic, not derived from first principles.

free parameters (6)

k (sigmoid steepness) = 0.05
Controls the steepness of the TAP/DTAP sigmoid in Eq. (3) and Eq. (5); chosen by hand, no sensitivity analysis in the main text.
tmid (sigmoid midpoint) = 750
Midpoint timestep for the selection probability at mid-training; set from Observation 2, not tuned or justified across models.
imax (total training steps) = 1000
Used to define t_dyn(i) in Eq. (4); fixed for all experiments, no analysis of sensitivity to training length.
resize ratio r = 0.5, 0.625, 0.75
Downsampling factor for BP-low; swept over three values, but results vary by metric and model, and no single value is shown to be universally best.
number of ZO perturbations N = not stated in main text
Number of perturbation directions in MeZO (Algorithm 1), directly affects gradient quality and memory; value deferred to supplementary.
learning rates eta and alpha = not stated in main text
Learning rates for BP-low and ZO-high updates; not reported, deferred to supplementary, and their impact on convergence is not analyzed.

assumptions (3)

domain assumption High timesteps in the diffusion process carry structural/global information, while low timesteps carry fine details.
Invoked in Section 2.4 to route ZO-high toward high timesteps; based on a citation to perception prioritized training [3] rather than a direct measurement in this paper.
domain assumption MeZO gradient estimates become effective only after partial fine-tuning has injected target-specific information.
Observation 1 (Fig. 1a) is the basis for DTAP's dynamic midpoint moving to lower timesteps over training; supported only by a qualitative illustration, not quantitative evidence.
domain assumption BP-low causes structural distortions concentrated at high timesteps.
Observation 2 (Fig. 1b) motivates the entire TAP design; shown on a few qualitative examples, with no quantitative measurement of distortion as a function of timestep.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Memory-Efficient Personalization of Text-to-Image Diffusion Models via Selective Optimization Strategies." pith.science (2026). https://pith.science/paper/MFFN7YOV

@misc{pith2026250710029,
  author       = {Pith},
  title        = {Pith review of: Memory-Efficient Personalization of Text-to-Image Diffusion Models via Selective Optimization Strategies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MFFN7YOV}},
  note         = {Machine review of arXiv:2507.10029}
}

read the original abstract

Memory-efficient personalization is critical for adapting text-to-image diffusion models while preserving user privacy and operating within the limited computational resources of edge devices. To this end, we propose a selective optimization framework that adaptively chooses between backpropagation on low-resolution images (BP-low) and zeroth-order optimization on high-resolution images (ZO-high), guided by the characteristics of the diffusion process. As observed in our experiments, BP-low efficiently adapts the model to target-specific features, but suffers from structural distortions due to resolution mismatch. Conversely, ZO-high refines high-resolution details with minimal memory overhead but faces slow convergence when applied without prior adaptation. By complementing both methods, our framework leverages BP-low for effective personalization while using ZO-high to maintain structural consistency, achieving memory-efficient and high-quality fine-tuning. To maximize the efficacy of both BP-low and ZO-high, we introduce a timestep-aware probabilistic function that dynamically selects the appropriate optimization strategy based on diffusion timesteps. This function mitigates the overfitting from BP-low at high timesteps, where structural information is critical, while ensuring ZO-high is applied more effectively as training progresses. Experimental results demonstrate that our method achieves competitive performance while significantly reducing memory consumption, enabling scalable, high-quality on-device personalization without increasing inference latency.

Figures

Figures reproduced from arXiv: 2507.10029 by the authors.

**Figure 2.** Comparison of TAP and DTAP over training steps [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Qualitative comparison of optimization methods. Overfitting (BP-high), distortions (BP-low), poor personalization (ZO-high), [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 11 canonical work pages

[1]

Emerg- ing properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the International Conference on Computer Vi- sion (ICCV), 2021. 3

work page 2021
[2]

InstructBooth: Instruction-following Personalized Text-to-Image Generation

Daewon Chae, Nokyung Park, Jinkyu Kim, and Kimin Lee. Instructbooth: Instruction-following personalized text-to- image generation. arXiv preprint arXiv:2312.03011, 2023. 1

work page Pith review arXiv 2023
[3]

Perception pri- oritized training of diffusion models

Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception pri- oritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472–11481, 2022. 3

work page 2022
[4]

An image is worth one word: Personalizing text-to- image generation using textual inversion

Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patash- nik, Amit H Bermano, Gal Chechik, and Daniel Cohen- Or. An image is worth one word: Personalizing text-to- image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1

arXiv 2022
[5]

Progressive knowledge dis- tillation of stable diffusion xl using layer level loss

Yatharth Gupta, Vishnu V Jaddipal, Harish Prabhala, Sayak Paul, and Patrick V on Platen. Progressive knowledge dis- tillation of stable diffusion xl using layer level loss. arXiv preprint arXiv:2401.02677, 2024. 3

arXiv 2024
[6]

Lora: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2

2022
[7]

Emojidiff: Advanced facial expression con- trol with high identity preservation in portrait generation

Liangwei Jiang, Ruida Li, Zhifeng Zhang, Shuo Fang, and Chenguang Ma. Emojidiff: Advanced facial expression con- trol with high identity preservation in portrait generation. arXiv preprint arXiv:2412.01254, 2024. 1

work page arXiv 2024
[8]

Multi-concept customization of text-to-image diffusion

Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 1931–1941, 2023. 1

work page 1931

Show all 18 references

[9]

Blip-diffusion: Pre- trained subject representation for controllable text-to-image generation and editing

Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pre- trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Pro- cessing Systems, 36:30146–30166, 2023. 1

2023
[10]

Fine- tuning language models with just forward passes

Sadhika Malladi, Tianyu Gao, Eshaan Nichani, Alex Damian, Jason D Lee, Danqi Chen, and Sanjeev Arora. Fine- tuning language models with just forward passes. Advances in Neural Information Processing Systems, 36:53038–53075,
[11]

Sdxl: Improving latent diffusion mod- els 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 mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3

2023 arXiv
[12]

Learning transferable visual models from natural language supervi- sion

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

2021
[13]

High-resolution image synthesis with latent diffusion models

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 3

2022
[14]

Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2250...

2023
[15]

In- stantbooth: Personalized text-to-image generation without test-time finetuning

Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. In- stantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 8543–8552, 2024. 1

2024
[16]

Elite: Encoding visual con- cepts into textual embeddings for customized text-to-image generation

Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual con- cepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943–15953, 2023. 1

2023
[17]

Avatarbooth: High-quality and customizable 3d human avatar generation

Yifei Zeng, Yuanxun Lu, Xinya Ji, Yao Yao, Hao Zhu, and Xun Cao. Avatarbooth: High-quality and customizable 3d human avatar generation. arXiv preprint arXiv:2306.09864,

arXiv
[18]

Jedi: Joint- image diffusion models for finetuning-free personalized text- to-image generation

Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting- Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint- image diffusion models for finetuning-free personalized text- to-image generation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recogni...

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Emerg- ing properties in self-supervised vision transformers

Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the International Conference on Computer Vi- sion (ICCV), 2021. 3

work page 2021

[2] [2]

InstructBooth: Instruction-following Personalized Text-to-Image Generation

Daewon Chae, Nokyung Park, Jinkyu Kim, and Kimin Lee. Instructbooth: Instruction-following personalized text-to- image generation. arXiv preprint arXiv:2312.03011, 2023. 1

work page Pith review arXiv 2023

[3] [3]

Perception pri- oritized training of diffusion models

Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception pri- oritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472–11481, 2022. 3

work page 2022

[4] [4]

An image is worth one word: Personalizing text-to- image generation using textual inversion

Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patash- nik, Amit H Bermano, Gal Chechik, and Daniel Cohen- Or. An image is worth one word: Personalizing text-to- image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1

arXiv 2022

[5] [5]

Progressive knowledge dis- tillation of stable diffusion xl using layer level loss

Yatharth Gupta, Vishnu V Jaddipal, Harish Prabhala, Sayak Paul, and Patrick V on Platen. Progressive knowledge dis- tillation of stable diffusion xl using layer level loss. arXiv preprint arXiv:2401.02677, 2024. 3

arXiv 2024

[6] [6]

Lora: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2

2022

[7] [7]

Emojidiff: Advanced facial expression con- trol with high identity preservation in portrait generation

Liangwei Jiang, Ruida Li, Zhifeng Zhang, Shuo Fang, and Chenguang Ma. Emojidiff: Advanced facial expression con- trol with high identity preservation in portrait generation. arXiv preprint arXiv:2412.01254, 2024. 1

work page arXiv 2024

[8] [8]

Multi-concept customization of text-to-image diffusion

Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 1931–1941, 2023. 1

work page 1931