arxiv: 2605.02152 · v1 · submitted 2026-05-04 · 💻 cs.CV

Recognition: unknown

SpecEdit: Training-Free Acceleration for Diffusion based Image Editing via Semantic Locking

Haoran Qin, Jiacheng Liu, Jiaxuan Ren, Jinkui Ren, Linfeng Zhang, Peiliang Cai, Shikang Zheng, Xiantao Zhang, Xiaobing Tu, Yinggui Wang, Yuqi Lin, Zhengan Yan

Authors on Pith no claims yet

Pith reviewed 2026-05-09 16:46 UTC · model grok-4.3

classification 💻 cs.CV

keywords diffusion modelsimage editingaccelerationdynamic resolutionsemantic lockingtraining-freedraft-and-verify

0 comments

The pith

SpecEdit accelerates diffusion image editing by using a low-resolution draft to lock most tokens at coarse resolution after checking token discrepancies for semantic changes.

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

Diffusion-based image editing requires repeated high-resolution denoising across all spatial tokens, making it slow even for small edits. SpecEdit addresses this with a training-free draft-and-verify process: a quick low-resolution draft estimates the edited outcome, then token-level differences from the original image flag only the regions that truly need semantic modification. High-resolution steps run solely on those flagged tokens while the rest stay coarse. Experiments show this preserves editing quality on models such as Qwen-Image-Edit and FLUX.1-Kontext-dev while delivering large speed gains, and the method combines with existing acceleration tricks for even greater savings.

Core claim

The central claim is that a low-resolution draft can reliably estimate semantic editing outcomes, after which token-level discrepancies identify precisely which spatial tokens require high-resolution denoising; the remaining tokens can stay at coarse resolution without introducing structural inconsistency or quality loss.

What carries the argument

Draft-and-verify scheme that computes token-level discrepancies between a low-resolution semantic draft and the input image to select only edit-relevant tokens for high-resolution processing.

If this is right

Up to 10x acceleration on Qwen-Image-Edit while keeping strong editing quality.
Up to 7x acceleration on FLUX.1-Kontext-dev with comparable results.
Further gains to 13x when combined with step distillation or other acceleration methods.
Training-free applicability to existing diffusion editing pipelines without model changes.

Where Pith is reading between the lines

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

The same discrepancy check could replace low-level heuristics such as edge detection in other dynamic-resolution sampling methods.
Real-time or interactive editing interfaces on modest hardware become feasible if the draft step is made even lighter.
The approach may generalize to video or 3D diffusion editing where spatial token costs dominate.
Measuring discrepancy reliability across edit types would clarify when the method risks missing subtle changes.

Load-bearing premise

Token discrepancies computed from the low-resolution draft accurately mark every region where semantic editing is needed and do not cause inconsistencies when non-edit tokens remain coarse.

What would settle it

A test case where a required global or subtle semantic change produces no detectable token discrepancies, resulting in visible artifacts or missing edits when high-resolution denoising is skipped on those areas.

Figures

Figures reproduced from arXiv: 2605.02152 by Haoran Qin, Jiacheng Liu, Jiaxuan Ren, Jinkui Ren, Linfeng Zhang, Peiliang Cai, Shikang Zheng, Xiantao Zhang, Xiaobing Tu, Yinggui Wang, Yuqi Lin, Zhengan Yan.

**Figure 1.** Figure 1: (Left) Visualization of the semantic discrepancy estimated by our semantic verification mechanism. The discrepancy is computed between the result and the original input in a perceptual feature space, revealing that only a small spatial region participates in the edit. (Right) Distribution of dissimilarity ratios on GEdit-Bench [3] and ImgEdit-Bench [4], where the ratio measures the proportion of tokens wh… view at source ↗

**Figure 2.** Figure 2: Visualization of semantic locking captured by SpecEdit. view at source ↗

**Figure 3.** Figure 3: Comparison with conventional dynamic-resolution methods. (a) Edge-based methods (e.g., RALU [7]) select refinement regions using edge responses, which mainly capture low-level structures and are not semantic-aware. (b) Channel-variance-based methods (e.g., Fresco [8]) rely on channel statistics to guide region refinement but remain weakly aligned with editing semantics. (c) SpecEdit (Ours) generates an ult… view at source ↗

**Figure 4.** Figure 4: Overview of SpecEdit. A low-resolution draft is produced to estimate semantic discrepancies. The verified tokens are prioritized for high-resolution upsampling during sampling. A sparse set of uniformly sampled tokens is additionally included to maintain global structural stability. 3.3 Spatial Draft-and-Verify Mechanism 3.3.1 Draft Generation Before entering the dynamic-resolution denoising framework, we… view at source ↗

**Figure 5.** Figure 5: Qualitative comparison on GEdit-Bench. SpecEdit achieves superior semantic preservation and visual quality at a higher acceleration ratio (5.38×) compared to state-of-theart baselines view at source ↗

**Figure 6.** Figure 6: Compatibility with other methods. SpecEdit achieves additive speedups of up to 13.03× and 6.70× when combined with distillation and caching, respectively view at source ↗

**Figure 7.** Figure 7: Effect of different downsampling ratios on preserving editable regions. We compare the reference and edited images at the original resolution (1024×1024) together with their downsampled versions (256×256, 170×170, and 128×128), corresponding to 16×, 36×, and 64× downsampling. When using 16× downsampling, most structural and semantic information of the image is preserved, enabling the edited regions to be r… view at source ↗

**Figure 8.** Figure 8: More visualization of semantic locking captured by SpecEdit. view at source ↗

**Figure 9.** Figure 9: Comparison of partially upsampled regions. We visualize the spatial regions selected for high-resolution refinement by different methods, including RALU, Fresco, and SpecEdit (ours). RALU and Fresco rely on low-level structural cues such as edges or channel statistics, which often activate regions unrelated to the editing instruction. In contrast, SpecEdit identifies semantically relevant areas through dr… view at source ↗

**Figure 10.** Figure 10: More qualitative comparison on GEdit-Bench. SpecEdit achieves superior semantic preservation and visual quality at a higher acceleration ratio (5.38×) compared to state-of-the-art baselines. 21 view at source ↗

read the original abstract

Diffusion-based image editing offers strong semantic controllability, but remains computationally expensive due to iterative high-resolution denoising over all spatial tokens. Dynamic-resolution sampling reduces this cost by performing early steps at reduced resolution. However, existing approaches prioritize upsampling using low-level heuristics such as edge detection or channel variance, which are weakly aligned with editing semantics and may lead to structural inconsistency. Moreover, spatial regions are often upsampled without verifying whether semantic modification is actually required, resulting in redundant high-resolution computation and accumulated errors. Therefore, we propose SpecEdit, a training-free dynamic-resolution framework tailored for diffusion-based image editing. SpecEdit follows a draft-and-verify scheme: a low-resolution draft first estimates the semantic outcome, after which token-level discrepancies are used to identify edit-relevant tokens for high-resolution denoising, while the remaining tokens stay at a coarse resolution. Experiments on Qwen-Image-Edit and FLUX.1-Kontext-dev demonstrate up to 10x and 7x acceleration, while maintaining strong quality. SpecEdit is complementary to step distillation and other acceleration techniques, achieving up to 13x speedup when combined with existing methods. Our code is in supplementary material and will be released on GitHub.

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

SpecEdit's semantic discrepancy approach for dynamic resolution in editing looks practical but its reliability hinges on unproven token selection.

read the letter

SpecEdit offers a training-free way to speed up diffusion image editing by using a low-resolution draft to select semantically relevant tokens for high-resolution processing via discrepancies. What is new is moving beyond low-level heuristics like edge detection to a semantic locking mechanism tailored for editing. The draft-and-verify lets most tokens stay coarse while only editing areas get the full compute, and the reported speedups on Qwen-Image-Edit and FLUX.1-Kontext-dev reach 10x and 7x with maintained quality. It also combines well with other methods like distillation for up to 13x. Making the code available is helpful for checking the implementation. The soft spot is the reliance on token-level discrepancies to accurately identify edit regions. Since diffusion is iterative and spatially connected, a mismatch between the low-res draft and the actual high-res trajectory could cause boundary issues or missed changes in the final output. The paper needs to show through ablations that the discrepancy metric and threshold are robust across different edits and models; without that, the acceleration might not generalize as claimed. This work is for researchers and engineers focused on efficient generative models and interactive editing applications. A reader wanting concrete acceleration techniques for diffusion would get practical ideas from it. It deserves serious peer review because the core idea addresses a real computational issue in a novel way, even if some validation details require clarification. I recommend sending it for review.

Referee Report

2 major / 1 minor

Summary. The paper proposes SpecEdit, a training-free dynamic-resolution framework for diffusion-based image editing. It follows a draft-and-verify scheme in which a low-resolution forward pass estimates the semantic outcome of the edit prompt; token-level discrepancies between this draft and the input are then used to identify edit-relevant tokens that receive full high-resolution denoising, while non-edit tokens remain at coarse resolution. Experiments on Qwen-Image-Edit and FLUX.1-Kontext-dev are reported to yield up to 10x and 7x acceleration respectively, with quality preserved; the method is also shown to combine with step distillation for up to 13x speedup.

Significance. If the central assumptions hold, SpecEdit would provide a practical, training-free acceleration technique that aligns computational effort with semantic edit regions rather than low-level heuristics, complementing existing distillation and scheduling methods. The approach could meaningfully lower inference costs for controllable image editing without requiring model retraining.

major comments (2)

[Methods / draft-and-verify scheme] The draft-and-verify core (described in the abstract and methods overview) assumes that a single low-resolution pass produces a per-token discrepancy statistic sufficient to catch every region whose semantics change under the edit prompt and to leave remaining tokens at coarse resolution without introducing structural inconsistencies. No derivation, formal definition of the discrepancy operator, or ablation of this statistic is provided, which is load-bearing for both the acceleration and quality-preservation claims.
[Abstract and Experiments] The abstract reports concrete speedups and quality preservation on two named models but supplies no details on the discrepancy metric, token selection threshold, exact upsampling schedule, or experimental controls (e.g., how mixed-resolution features are handled inside the U-Net or transformer blocks). This absence prevents verification that the method avoids boundary artifacts or missed edit regions, directly undermining the soundness of the reported 10x/7x/13x figures.

minor comments (1)

[Abstract] The abstract would benefit from a one-sentence clarification of how the mixed-resolution schedule is realized in the underlying architecture (Qwen-Image-Edit and FLUX.1-Kontext-dev).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We address each major comment point by point below, clarifying the current manuscript content where possible and outlining planned revisions to improve rigor and reproducibility.

read point-by-point responses

Referee: [Methods / draft-and-verify scheme] The draft-and-verify core (described in the abstract and methods overview) assumes that a single low-resolution pass produces a per-token discrepancy statistic sufficient to catch every region whose semantics change under the edit prompt and to leave remaining tokens at coarse resolution without introducing structural inconsistencies. No derivation, formal definition of the discrepancy operator, or ablation of this statistic is provided, which is load-bearing for both the acceleration and quality-preservation claims.

Authors: We agree that the manuscript would benefit from a more explicit formal definition and supporting ablation. The discrepancy operator is currently described in Section 3.2 as the token-wise semantic difference computed via cosine distance on projected features from the low-resolution draft versus the input. While the approach is primarily empirical rather than derived from first principles, we will add a formal definition of the operator, a brief motivation for its use in semantic locking, and a new ablation study (including edit-region detection precision/recall and structural consistency metrics) in the revised version. This directly addresses the load-bearing nature of the statistic for the reported speedups and quality claims. revision: yes
Referee: [Abstract and Experiments] The abstract reports concrete speedups and quality preservation on two named models but supplies no details on the discrepancy metric, token selection threshold, exact upsampling schedule, or experimental controls (e.g., how mixed-resolution features are handled inside the U-Net or transformer blocks). This absence prevents verification that the method avoids boundary artifacts or missed edit regions, directly undermining the soundness of the reported 10x/7x/13x figures.

Authors: We acknowledge that the current presentation lacks sufficient implementation details for independent verification. In the revised manuscript we will expand Section 3 and the experimental section to specify: the discrepancy metric (cosine similarity on CLIP-aligned token embeddings), the selection threshold (empirically set at 0.25), the progressive upsampling schedule (coarse-to-fine over the initial denoising steps), and the exact interpolation/fusion procedure for mixed-resolution features within transformer blocks. We will also include a short analysis of boundary artifacts and missed-region cases. The abstract will be updated with a concise reference to these parameters while respecting length limits. revision: yes

Circularity Check

0 steps flagged

No circularity: algorithmic scheduling heuristic with external empirical validation

full rationale

The paper introduces SpecEdit as a training-free dynamic-resolution scheduling method: a low-resolution draft estimates semantics, token-level discrepancies identify edit-relevant regions for high-resolution denoising, and non-edit tokens remain coarse. This is presented as a heuristic without equations that reduce to fitted parameters, self-definitions, or load-bearing self-citations. Claims of acceleration (up to 10x/7x, or 13x combined) rest on experiments with external models (Qwen-Image-Edit, FLUX.1-Kontext-dev) rather than internal fitting or renamed known results. No uniqueness theorems, ansatzes via prior self-work, or predictions-by-construction appear in the derivation. The approach is self-contained as a scheduling policy whose correctness is externally falsifiable.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard diffusion model assumptions plus one domain assumption about low-resolution approximations; no free parameters or new entities are introduced.

axioms (1)

domain assumption Low-resolution denoising can approximate the semantic outcome of full-resolution editing sufficiently well for discrepancy-based token selection.
This underpins the entire draft step and the decision of which tokens receive high-resolution computation.

pith-pipeline@v0.9.0 · 5555 in / 1497 out tokens · 49222 ms · 2026-05-09T16:46:20.061762+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

33 extracted references · 17 canonical work pages · 8 internal anchors

[1]

Qwen-Image Technical Report

Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-image technical report.arXiv preprint arXiv:2508.02324, 2025

work page internal anchor Pith review arXiv 2025
[2]

Flux.1 kontext: Flow matching for in-context image generation and editing in latent space, 2025

Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas M¨ uller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. Flux.1 kontext: Flow matching for in-context imag...

2025
[3]

Step1X-Edit: A Practical Framework for General Image Editing

Shiyu Liu, Yucheng Han, Peng Xing, Fukun Yin, Rui Wang, Wei Cheng, Jiaqi Liao, Ying- ming Wang, Honghao Fu, Chunrui Han, et al. Step1x-edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761, 2025

work page internal anchor Pith review arXiv 2025
[4]

ImgEdit: A Unified Image Editing Dataset and Benchmark

Yang Ye, Xianyi He, Zongjian Li, Bin Lin, Shenghai Yuan, Zhiyuan Yan, Bohan Hou, and Li Yuan. Imgedit: A unified image editing dataset and benchmark.arXiv preprint arXiv:2505.20275, 2025

work page internal anchor Pith review arXiv 2025
[5]

From reusing to forecasting: Accelerating diffusion models with taylorseers

Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Junjie Chen, and Linfeng Zhang. From reusing to forecasting: Accelerating diffusion models with taylorseers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15853–15863, 2025

2025
[6]

Timestep embedding tells: It’s time to cache for video diffusion model

Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It’s time to cache for video diffusion model. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 7353–7363, 2025

2025
[7]

Upsample what matters: Region-adaptive latent sampling for accelerated diffusion transformers.arXiv preprint arXiv:2507.08422, 2025

Wongi Jeong, Kyungryeol Lee, Hoigi Seo, and Se Young Chun. Upsample what matters: Region-adaptive latent sampling for accelerated diffusion transformers.arXiv preprint arXiv:2507.08422, 2025

work page arXiv 2025
[8]

From sketch to fresco: Efficient diffusion transformer with progressive resolution

Shikang Zheng, Guantao Chen, Lixuan He, Jiacheng Liu, Yuqi Lin, Chang Zou, and Linfeng Zhang. From sketch to fresco: Efficient diffusion transformer with progressive resolution. arXiv preprint arXiv:2601.07462, 2026

work page arXiv 2026
[9]

Training-free diffusion acceleration with bottleneck sampling.arXiv preprint arXiv:2503.18940, 2025

Ye Tian, Xin Xia, Yuxi Ren, Shanchuan Lin, Xing Wang, Xuefeng Xiao, Yunhai Tong, Ling Yang, and Bin Cui. Training-free diffusion acceleration with bottleneck sampling.arXiv preprint arXiv:2503.18940, 2025

work page arXiv 2025
[10]

Spargeattn: Accurate sparse attention accelerating any model inference

Jintao Zhang, Chendong Xiang, Haofeng Huang, Jia Wei, Haocheng Xi, Jun Zhu, and Jianfei Chen. Spargeattn: Accurate sparse attention accelerating any model inference. In International Conference on Machine Learning (ICML), 2025. 13

2025
[11]

Fast inference from transformers via speculative decoding

Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. InInternational Conference on Machine Learning, pages 19274– 19286. PMLR, 2023

2023
[12]

Adding conditional control to text- to-image diffusion models

Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text- to-image diffusion models. InProceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023

2023
[13]

Prompt-to-Prompt Image Editing with Cross Attention Control

Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022

work page internal anchor Pith review arXiv 2022
[14]

Imagic: Text-based real image editing with diffusion models

Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6007–6017, 2023

2023
[15]

Denoising Diffusion Implicit Models

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2010
[16]

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

Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. 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

2022
[17]

Dpm- solver++: Fast solver for guided sampling of diffusion probabilistic models.Machine Intelligence Research, 22(4):730–751, 2025

Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm- solver++: Fast solver for guided sampling of diffusion probabilistic models.Machine Intelligence Research, 22(4):730–751, 2025

2025
[18]

Dpm-solver-v3: Improved diffusion ode solver with empirical model statistics.Advances in Neural Information Processing Systems, 36:55502–55542, 2023

Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Dpm-solver-v3: Improved diffusion ode solver with empirical model statistics.Advances in Neural Information Processing Systems, 36:55502–55542, 2023

2023
[19]

Progressive Distillation for Fast Sampling of Diffusion Models

Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022

work page internal anchor Pith review arXiv 2022
[20]

Let features decide their own solvers: Hybrid feature caching for diffusion transformers, 2025

Shikang Zheng, Guantao Chen, Qinming Zhou, Yuqi Lin, Lixuan He, Chang Zou, Peiliang Cai, Jiacheng Liu, and Linfeng Zhang. Let features decide their own solvers: Hybrid feature caching for diffusion transformers, 2025

2025
[21]

Freqca: Accelerating diffusion models via frequency-aware caching.arXiv preprint arXiv:2510.08669, 2025

Jiacheng Liu, Peiliang Cai, Qinming Zhou, Yuqi Lin, Deyang Kong, Benhao Huang, Yupei Pan, Haowen Xu, Chang Zou, Junshu Tang, et al. Freqca: Accelerating diffusion models via frequency-aware caching.arXiv preprint arXiv:2510.08669, 2025

work page arXiv 2025
[22]

Longformer: The Long-Document Transformer

Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document trans- former.arXiv preprint arXiv:2004.05150, 2020

work page internal anchor Pith review arXiv 2004
[23]

Generating Long Sequences with Sparse Transformers

Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers.arXiv preprint arXiv:1904.10509, 2019

work page internal anchor Pith review arXiv 1904
[24]

Big bird: Transformers for longer sequences.Advances in neural information processing systems, 33:17283–17297, 2020

Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences.Advances in neural information processing systems, 33:17283–17297, 2020

2020
[25]

Cascaded diffusion models for high fidelity image generation.Journal of Machine Learning Research, 23(47):1–33, 2022

Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation.Journal of Machine Learning Research, 23(47):1–33, 2022. 14

2022
[26]

Srdiff: Single image super-resolution with diffusion probabilistic models

Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47–59, 2022

2022
[27]

Denoising diffusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

2020
[28]

SpotEdit: Selective region editing in diffusion transformers.arXiv preprint arXiv:2512.22323, 2025

Zhibin Qin, Zhenxiong Tan, Zeqing Wang, Songhua Liu, and Xinchao Wang. Spotedit: Selective region editing in diffusion transformers.arXiv preprint arXiv:2512.22323, 2025

work page arXiv 2025
[29]

Training-free diffusion acceleration with bottleneck sampling, 2025

Ye Tian, Xin Xia, Yuxi Ren, Shanchuan Lin, Xing Wang, Xuefeng Xiao, Yunhai Tong, Ling Yang, and Bin Cui. Training-free diffusion acceleration with bottleneck sampling, 2025

2025
[30]

Accelerating diffusion transformers with token-wise feature caching.arXiv preprint arXiv:2410.05317,

Chang Zou, Xuyang Liu, Ting Liu, Siteng Huang, and Linfeng Zhang. Accelerating diffu- sion transformers with token-wise feature caching.arXiv preprint arXiv:2410.05317, 2024

work page arXiv 2024
[31]

Fora: Fast-forward caching in diffusion transformer acceleration.arXiv preprint arXiv:2407.01425, 2024

Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang. Fora: Fast-forward caching in diffusion transformer acceleration.arXiv preprint arXiv:2407.01425, 2024

work page arXiv 2024
[32]

Accelerating diffusion transformers with dual feature caching.arXiv preprint arXiv:2412.18911, 2024

Chang Zou, Evelyn Zhang, Runlin Guo, Haohang Xu, Conghui He, Xuming Hu, and Lin- feng Zhang. Accelerating diffusion transformers with dual feature caching.arXiv preprint arXiv:2412.18911, 2024

work page arXiv 2024
[33]

Forecast then calibrate: Feature caching as ode for efficient diffusion transformers.arXiv preprint arXiv:2508.16211, 2025

Shikang Zheng, Liang Feng, Xinyu Wang, Qinming Zhou, Peiliang Cai, Chang Zou, Ji- acheng Liu, Yuqi Lin, Junjie Chen, Yue Ma, et al. Forecast then calibrate: Feature caching as ode for efficient diffusion transformers.arXiv preprint arXiv:2508.16211, 2025. 15 Supplementary Material 7 Additional Ablation Studies Table 10: Ablation Study on FLUX.1-Kontext-de...

work page arXiv 2025