Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Fast cached diffusion inference can be both faster and more accurate by forwarding the exact features a verification step already computed, instead of discarding them.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 00:25 UTC pith:YVMSRJOG

load-bearing objection A small, well-controlled idea — reuse the exact feature a paid check already computed — with credible matched-compute evidence, but its benefit rests on an unguaranteed local-exactness premise and needs artifacts. the 3 major comments →

arxiv 2607.27842 v1 pith:YVMSRJOG submitted 2026-07-30 cs.CV cs.LG

FeatFix: Reuse What You Verify through Local Exact-Feature Correction for Faster Cached Diffusion Inference

Hanshuai Cui , Zhiqing Tang , Zhi Yao , Qianli Ma , Fanshuai Meng , Weijia Jia This is my paper
classification cs.CV cs.LG
keywords diffusion transformersinference accelerationfeature cachingverificationexact-feature correctiontraining-free accelerationlatency reductionimage and video generation
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.

This paper claims that when a cached diffusion accelerator pays for an exact block computation to verify a draft feature, the resulting exact tensor should be forwarded downstream rather than discarded. Doing so removes the local draft residual by construction at zero extra compute. The paper reports that this correction reduces downstream feature error and improves quality metrics at matched latency. The claim is tested across four image and video backbones, with speedups up to 6.70× over full inference while preserving competitive output quality.

Core claim

FeatFix's central claim is that the exact block output already computed at a paid verification site is a better forward signal than the draft it verifies. At each fixed layer–timestep site, FeatFix replaces the complete draft block output with the exact output computed from the same incoming state; because the verification pass already computed it, the swap adds no block FLOPs. The paper argues whole-block replacement is the compute-consistent unit: partial token or channel replacement leaves residual error, and full-timestep recomputation repeats the transformer stack. Under matched compute, FeatFix outperforms its Verify control on every reported quality metric and reduces downstream relat

What carries the argument

The correction rule is Eq. (5): at sites in a frozen schedule S, the outgoing block feature is the exact output h^f (computed from the same incoming state), otherwise it is the draft. The mechanism works because the verification site already evaluates the full block B(u;t,c), so forwarding the whole [B,N,C] tensor costs no extra exact-block FLOPs. The paper uses relative-L1 discrepancy between draft and exact features, plus downstream feature traces, to measure the local reset and its propagation.

Load-bearing premise

The exact feature is exact only relative to the drifted incoming state; the paper assumes that starting the next block from this exact output is better than starting it from the draft, i.e., local exactness outweighs global trajectory drift. If a drifted state makes the exact output incompatible with downstream blocks, the swap could hurt.

What would settle it

Find a layer–timestep site where forwarding the exact block output increases downstream feature error or worsens an end metric relative to forwarding the draft. The paper's own Fig 6a shows late-stage sites (s25–26, L28) with only ~50% positive effect; a systematic sweep that identifies sites with negative propagation would falsify the claim that forwarding exact features is always beneficial at matched compute.

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

If this is right

  • Any verification-based accelerator can convert its audit signal into a free correction by forwarding the exact tensor instead of discarding it.
  • Quality at a given speed improves: in the controlled FLUX study, FeatFix beats the matched Verify control on every reported metric at essentially the same latency.
  • The correction propagates downstream: feature traces show reduced relative-L1 error through later layers and timesteps, with an overall reduction of 2.54×10⁻³.
  • The method transfers across image and video diffusion transformers, achieving up to 6.70× speedup on FLUX, 3.39× on DiT-XL/2, and similar gains on Qwen-Image and HunyuanVideo-1.5.
  • Partial replacement (token- or channel-level) is suboptimal; the whole block is the efficient unit for removing the verified residual.

Where Pith is reading between the lines

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

  • The principle suggests that verification and correction should be unified: every time you pay for an exact feature, you should use it as the output. This could reshape how caching schedules are designed, choosing paid sites for their correction value rather than solely for error auditing.
  • The paper's own Fig 6 shows that late-stage corrections exhibit decaying positive effects (about 50% at s25–26 for L28). An adaptive policy might skip corrections where drift has saturated, saving the cost of checks that no longer help.
  • A natural extension is to apply the same 'reuse what you verify' rule to other block types or to forecast-based draft policies, resetting drift periodically with exact features.
  • The correction could interact with budget-aware scheduling: since each forwarded exact feature is free at verification sites, the marginal cost of correction is zero, so the optimal verification schedule may be denser than one designed purely for monitoring.

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

3 major / 5 minor

Summary. FeatFix proposes a training-free correction mechanism for cached diffusion inference. When a verification-based accelerator already computes an exact block output at a paid layer–timestep site, FeatFix forwards that exact output downstream instead of discarding it. The paper argues that replacing the draft block output with the exact block output computed from the same incoming state eliminates the local same-input residual by construction (Eq. 5), without adding exact-block FLOPs relative to a matched Verify control. The method is evaluated on FLUX.1-dev, Qwen-Image, HunyuanVideo-1.5, and DiT-XL/2, reporting speedups up to 6.70× over Vanilla with competitive quality, and a controlled matched-compute comparison where FeatFix beats Verify on every reported metric at essentially the same latency (Table 2). Mechanism analyses show downstream feature error reduction with bootstrap CIs, and sensitivity sweeps over threshold and paid layer are provided.

Significance. If the results hold, this is a simple and practical insight: paid verification features should be reused for correction rather than discarded. The matched-compute Verify control is a strong causal design, and the paired bootstrap confidence intervals for downstream error reduction (Fig. 2b) and the four-backbone evaluation give the empirical claims unusual solidity. The paper is honest about limitations, explicitly acknowledging that the remaining trajectory may still differ from Vanilla and that site selection is currently frozen after a development-stage audit. The contribution is incremental but useful, and the experimental methodology sets a good standard for this type of training-free acceleration work.

major comments (3)
  1. [Method, after Eq. (5)] The paper never formally specifies the underlying draft policy D, the paid-site schedule S, or the threshold τ that triggers paid checks. The text says S is 'fixed before evaluation by a development-stage audit', but no audit procedure, algorithm, or pseudocode is given. Since all reported results depend on this schedule, the method is not reproducible as stated. Please specify the base caching/forecasting policy used in each experiment, how S is selected, and how τ enters the paid-check decision.
  2. [Mechanism Analysis, Fig. 6] The propagation analysis shows positive downstream effects for most stages, but the latest measured stage (s25–26 for layer L28) has only a 50% positive rate, i.e., no measurable benefit. The paper's claim that forwarding the exact feature 'reduces downstream feature error' is therefore not universal, even within the tested configuration. The manuscript should discuss this decay and state conditions (e.g., drift magnitude, layer depth, or denoising stage) under which correction is expected to help. As written, the central premise—local exactness dominates global trajectory drift—is supported empirically but only through a majority of stages, and the paper should qualify the claim accordingly.
  3. [Experiments, Tables 1–5] It is not stated which base draft policy FeatFix is built on for each backbone. For example, Table 1's FeatFix rows have identical FLOPs to SpeCa at N=20 and N=28, but the text does not say whether FeatFix is applied on top of the SpeCa draft policy, a custom fallback, or something else. Without this information, the reader cannot attribute the quality gains to the correction versus the underlying caching/forecasting behavior. Please make the base policy for each experiment explicit, and describe how the paid-site schedule is instantiated for each backbone.
minor comments (5)
  1. [Abstract and Introduction] The phrase 'resets the local draft residual' is accurate only with respect to the drifted incoming state u, not with respect to the Vanilla trajectory. The paper does acknowledge this later, but the abstract could be slightly more careful to distinguish local exactness from global trajectory fidelity.
  2. [Table 1] In the header, 'FLUX.1-dev on DrawBench 200' is slightly awkward; suggest 'FLUX.1-dev on DrawBench (200 prompts)'. Also, the horizontal rules that delimit speed regions are helpful, but the criteria for grouping are not stated; a sentence in the text would help.
  3. [Figure 2] Panel (b) reports '24/24 pointwise 95% CIs > 0', but the figure does not show the bootstrap intervals themselves. Consider adding error bars or a table with the intervals, as these are central to the claim.
  4. [Human Preference Study, Fig. 9] The study uses 18 participants × 10 questions. The confidence intervals are shown, but the text should state the exact counts (e.g., number of wins vs. losses) and whether the comparisons to SpeCa and TaylorSeer are statistically significant after multiple-comparison correction.
  5. [Limitations] The Limitations section is brief and could also mention that the method's benefit diminishes at late denoising stages, as shown in Fig. 6a. This would make the limitation statement more concrete.

Circularity Check

0 steps flagged

No significant circularity: the local correction is an explicit identity, and the load-bearing downstream claims are tested against a matched external control.

full rationale

The derivation is self-contained. The only 'by construction' statement is Eq. (5): at selected sites S, h_out = h_f, so the local same-input residual h_f - h_f is zero by definition. The paper explicitly labels this an identity ('Thus replacement eliminates the local same-input draft residual by construction') and does not use it as evidence of quality gain. The load-bearing claims—that forwarding the paid exact feature reduces downstream feature error and improves evaluator metrics—are tested against the matched Verify control (Table 2; Fig. 2b), where Verify computes the same exact tensor and discards it. Nothing in Eq. (5) forces ImageReward 1.318 vs 0.911 or the positive downstream relative-L1 intervals; the paper even reports decay at late stages (Fig. 6a), which is inconsistent with a purely definitional benefit. Site schedules are frozen by a development-stage audit, but the paper does not derive the reported gains from that audit by any equation, and it lists automatic site selection as future work rather than claiming a derivation. No load-bearing self-citation or imported uniqueness theorem appears; references to TaylorSeer/SpeCa are background and describe a different forwarding decision. The result is therefore an empirical intervention with an explicit control, not a fit or renaming.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

FeatFix introduces one new moving part — where and when to forward the verified exact feature — and inherits everything else from the base Taylor/cache draft policy and pretrained checkpoints. The free parameters are concentrated in the frozen site schedule and thresholds, tuned by a disclosed but under-specified development-stage audit. No new physical or architectural entities are postulated.

free parameters (4)
  • paid site schedule S (layer–timestep set) = Frozen after development-stage audit; e.g., FLUX paid layers L28–L37, per-speed-region timestep sites
    Chosen on development data; the reported operating points depend on this selection. Disclosed ('fixed before evaluation by a development-stage audit') but the selection rule is not fully specified; sensitivity shown in Fig 7.
  • base threshold τ for triggering paid checks = Per speed region; swept in Fig 7(a) as 'base threshold'
    Controls which sites become paid sites; tuned on the development set; directly sets the speed–quality operating point reported in Tables 1–5.
  • paid-block layer index = Deep FLUX blocks (e.g., L34/L35 anchors); swept across L28–L37 in Fig 7(b)
    Layer at which the correction is applied; the paper shows some layers give a better speed–quality tradeoff, an audited empirical choice rather than a derived one.
  • N, the number of denoising steps (speed region) = N=50, 28, 20 for FLUX; corresponding step/frame settings for Qwen-Image, HunyuanVideo-1.5, DiT-XL/2
    Inherited from the base draft policy (TaylorSeer/SpeCa-style). The aggressive 6.70× point is a tuned operating point of the composed system, not of the correction alone.
axioms (4)
  • standard math Taylor-series forecast of future DiT features from past exact refreshes (Eq. 2) is an adequate draft.
    Inherited from TaylorSeer/SpeCa (cited); FeatFix needs a forecasting/reuse draft to exist but does not depend on its specific accuracy.
  • domain assumption Intermediate features in diffusion transformers are temporally redundant, so cached or forecast features can substitute for exact computation at most sites.
    The shared premise of DeepCache, FORA, ToCa, TeaCache, HiCache etc. (cited); FeatFix inherits it unchanged.
  • ad hoc to paper Forwarding the exact block output h^f computed from a drift-affected incoming state u is no worse than forwarding the draft h^d from the same u.
    The paper's core premise. Empirically supported by Fig 2(b) and Fig 6 but not derived; the paper concedes 'the remaining trajectory may still differ from Vanilla' (Method).
  • domain assumption The base policy already performs paid verification (computes exact features) at the audited sites.
    FeatFix is scoped by the Limitations section to accelerators that 'already evaluate occasional exact block features'; without paid checks, correction costs FLOPs.

pith-pipeline@v1.3.0-daily-deepseek · 13113 in / 16527 out tokens · 156834 ms · 2026-08-01T00:25:59.154960+00:00 · methodology

0 comments
read the original abstract

Diffusion models are widely used to generate high-quality images and videos, but their iterative denoising process remains computationally intensive. A growing class of training-free accelerators reduces this cost by reusing cached intermediate features or forecasting future ones. To control draft drift, these methods sometimes compute an exact block feature for verification. Yet the resulting exact feature is typically used only to measure discrepancy or guide a later decision and is then discarded. We find that this previously computed feature can instead be reused for correction. Forwarding it at the verification site resets the local draft residual and reduces downstream feature error. Based on this observation, we introduce FeatFix, a local exact-feature correction method for cached diffusion inference. FeatFix operates at a fixed sparse set of layer--timestep sites. At each selected site, it replaces the complete draft block output with the exact output computed from the same incoming state, avoiding token- or channel-level partial replacement and full-timestep recomputation. Experiments across four image and video backbones show that FeatFix consistently accelerates generation, achieving a speedup of up to $6.70\times$ over Vanilla while maintaining competitive output quality.

Figures

Figures reproduced from arXiv: 2607.27842 by Fanshuai Meng, Hanshuai Cui, Qianli Ma, Weijia Jia, Zhiqing Tang, Zhi Yao.

Figure 1
Figure 1. Figure 1: Cross-backbone comparisons. Each column pairs [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Whole-block correction and downstream propaga [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FeatFix framework. At sites defined by a frozen layer–timestep schedule, FeatFix forwards the already-computed exact [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: FLUX runtime-matched reference and paired ef [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: FLUX.1-dev visual comparison. Arm Lat. (s)↓ CLIP↑ Pick↑ IR↑ PSNR↑ SSIM↑LPIPS↓ Vanilla 23.93 32.54 23.77 1.361 ∞ 1.0000 0.0000 Verify 3.60 31.48 22.80 0.911 14.29 0.5854 0.5381 FeatFix 3.57 33.03 23.56 1.318 16.13 0.6166 0.4492 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: FLUX threshold and paid-layer sensitivity. [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FLUX latency and peak-memory scaling across [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 11
Figure 11. Figure 11: DiT-XL/2 visual comparison. block features. This focus enables a clean matched-compute study while keeping the correction mechanism training-free, local, and auditable. For the reported settings, we freeze the sites after development-stage audits and keep them fixed within each evaluation. A natural next step is to extend this principle to automatic budget-aware site selection and addi￾tional inference st… view at source ↗

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

40 extracted references · 3 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume =

    Denoising Diffusion Probabilistic Models , author =. Advances in Neural Information Processing Systems , volume =. 2020 , url =

    2020

  2. [2]

    Diffusion Models Beat

    Dhariwal, Prafulla and Nichol, Alexander , booktitle =. Diffusion Models Beat. 2021 , url =

    2021

  3. [3]

    Advances in Neural Information Processing Systems , volume =

    Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding , author =. Advances in Neural Information Processing Systems , volume =. 2022 , url =

    2022

  4. [4]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    High-Resolution Image Synthesis with Latent Diffusion Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2022 , url =

    2022

  5. [5]

    International Conference on Learning Representations , year =

    Score-Based Generative Modeling through Stochastic Differential Equations , author =. International Conference on Learning Representations , year =

  6. [6]

    Advances in Neural Information Processing Systems , volume =

    Elucidating the Design Space of Diffusion-Based Generative Models , author =. Advances in Neural Information Processing Systems , volume =. 2022 , url =

    2022

  7. [7]

    2022 , url =

    Lu, Cheng and Zhou, Yuhao and Bao, Fan and Chen, Jianfei and Li, Chongxuan and Zhu, Jun , booktitle =. 2022 , url =

    2022

  8. [8]

    International Conference on Learning Representations , year =

    Flow Matching for Generative Modeling , author =. International Conference on Learning Representations , year =

  9. [9]

    International Conference on Learning Representations , year =

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author =. International Conference on Learning Representations , year =

  10. [10]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Scalable Diffusion Models with Transformers , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2023 , url =

    2023

  11. [11]

    Proceedings of the 41st International Conference on Machine Learning , series =

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author =. Proceedings of the 41st International Conference on Machine Learning , series =. 2024 , url =

    2024

  12. [12]

    International Conference on Learning Representations , year =

    Progressive Distillation for Fast Sampling of Diffusion Models , author =. International Conference on Learning Representations , year =

  13. [13]

    Proceedings of the 40th International Conference on Machine Learning , series =

    Consistency Models , author =. Proceedings of the 40th International Conference on Machine Learning , series =. 2023 , url =

    2023

  14. [14]

    2024 , howpublished =

    2024

  15. [15]

    2508.02324 , archivePrefix =

    Wu, Chenfei and Li, Jiahao and Zhou, Jingren and Lin, Junyang and Gao, Kaiyuan and Yan, Kun and Yin, Sheng-ming and Bai, Shuai and Xu, Xiao and Chen, Yilei and Chen, Yuxiang and Tang, Zecheng and Zhang, Zekai and Wang, Zhengyi and Yang, An and Yu, Bowen and Cheng, Chen and Liu, Dayiheng and Li, Deqing and Zhang, Hang and Meng, Hao and Wei, Hu and Ni, Jing...

    Pith/arXiv arXiv

  16. [16]

    2025 , eprint =

    Wu, Bing and Zou, Chang and Li, Changlin and Huang, Duojun and Yang, Fang and Tan, Hao and Peng, Jack and Wu, Jianbing and Xiong, Jiangfeng and Jiang, Jie and. 2025 , eprint =

    2025

  17. [17]

    2024 , url =

    Ma, Xinyin and Fang, Gongfan and Wang, Xinchao , booktitle =. 2024 , url =

    2024

  18. [18]

    2407.01425 , archivePrefix =

    Selvaraju, Pratheba and Ding, Tianyu and Chen, Tianyi and Zharkov, Ilya and Liang, Luming , year =. 2407.01425 , archivePrefix =

    Pith/arXiv arXiv

  19. [19]

    2406.01125 , archivePrefix =

    Chen, Pengtao and Shen, Mingzhu and Ye, Peng and Cao, Jianjian and Tu, Chongjun and Bouganis, Christos-Savvas and Zhao, Yiren and Chen, Tao , year =. 2406.01125 , archivePrefix =

    Pith/arXiv arXiv

  20. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2025 , url =

    2025

  21. [21]

    International Conference on Learning Representations , year =

    Accelerating Diffusion Transformers with Token-wise Feature Caching , author =. International Conference on Learning Representations , year =

  22. [22]

    From Reusing to Forecasting: Accelerating Diffusion Models with

    Liu, Jiacheng and Zou, Chang and Lyu, Yuanhuiyi and Chen, Junjie and Zhang, Linfeng , booktitle =. From Reusing to Forecasting: Accelerating Diffusion Models with. 2025 , url =

    2025

  23. [23]

    2025 , doi =

    Liu, Jiacheng and Zou, Chang and Lyu, Yuanhuiyi and Ren, Fei and Wang, Shaobo and Li, Kaixin and Zhang, Linfeng , booktitle =. 2025 , doi =

    2025

  24. [24]

    International Conference on Learning Representations , year =

    Real-Time Video Generation with Pyramid Attention Broadcast , author =. International Conference on Learning Representations , year =

  25. [25]

    , booktitle =

    Lyu, Zhengyao and Si, Chenyang and Song, Junhao and Yang, Zhenyu and Qiao, Yu and Liu, Ziwei and Wong, Kwan-Yee K. , booktitle =. 2025 , url =

    2025

  26. [26]

    2025 , url =

    Ma, Zehong and Wei, Longhui and Wang, Feng and Zhang, Shiliang and Tian, Qi , booktitle =. 2025 , url =

    2025

  27. [27]

    2026 , url =

    Feng, Liang and Zheng, Shikang and Liu, Jiacheng and Lin, Yuqi and Zhou, Qinming and Cai, Peiliang and Wang, Xinyu and Chen, Junjie and Zou, Chang and Ma, Yue and Zhang, Linfeng , booktitle =. 2026 , url =

    2026

  28. [28]

    2026 , url =

    Gu, Lihui and He, Jingbin and Su, Lianghao and He, Kang and Wang, Wenxiao and Liu, Yuliang , booktitle =. 2026 , url =

    2026

  29. [29]

    2026 , url =

    Bu, Jiazi and Ling, Pengyang and Zhou, Yujie and Wang, Yibin and Zang, Yuhang and Wu, Tong and Lin, Dahua and Wang, Jiaqi , booktitle =. 2026 , url =

    2026

  30. [30]

    2025 , howpublished =

    2025

  31. [31]

    2023 , url =

    Ghosh, Dhruba and Hajishirzi, Hannaneh and Schmidt, Ludwig , booktitle =. 2023 , url =

    2023

  32. [32]

    Microsoft

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll. Microsoft. Computer Vision -- ECCV 2014 , series =. 2014 , doi =

    2014

  33. [33]

    2009 , doi =

    Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li , booktitle =. 2009 , doi =

    2009

  34. [34]

    2024 , url =

    Huang, Ziqi and He, Yinan and Yu, Jiashuo and Zhang, Fan and Si, Chenyang and Jiang, Yuming and Zhang, Yuanhan and Wu, Tianxing and Jin, Qingyang and Chanpaisit, Nattapol and Wang, Yaohui and Chen, Xinyuan and Wang, Limin and Lin, Dahua and Qiao, Yu and Liu, Ziwei , booktitle =. 2024 , url =

    2024

  35. [35]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =. 2018 , url =

    2018

  36. [36]

    IEEE Transactions on Image Processing , volume =

    Image Quality Assessment: From Error Visibility to Structural Similarity , author =. IEEE Transactions on Image Processing , volume =. 2004 , doi =

    2004

  37. [37]

    2017 , url =

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , booktitle =. 2017 , url =

    2017

  38. [38]

    2021 , doi =

    Hessel, Jack and Holtzman, Ari and Forbes, Maxwell and Le Bras, Ronan and Choi, Yejin , booktitle =. 2021 , doi =

    2021

  39. [39]

    Advances in Neural Information Processing Systems , volume =

    Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation , author =. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    2023

  40. [40]

    2023 , url =

    Xu, Jiazheng and Liu, Xiao and Wu, Yuchen and Tong, Yuxuan and Li, Qinkai and Ding, Ming and Tang, Jie and Dong, Yuxiao , booktitle =. 2023 , url =

    2023