Pith. sign in

REVIEW 3 major objections 8 minor 38 references

SIEDD: Shared-Implicit Encoder with Discrete Decoders

T0 review · 3 major / 8 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 A shared encoder trained briefly on a few anchor frames, then frozen, lets frame-group decoders train in parallel and cuts INR video encoding time by 20–30x at HD and 4K.

desk verdict A genuinely novel architecture for faster INR video encoding, but the headline 20-30X speedup is not yet established because the baseline encoding times are not quality- or budget-matched. read the letter →

arxiv 2506.23382 v1 pith:QRLXLIOG submitted 2025-06-29 cs.CV cs.AIcs.LG

Vikram Rangarajan , Shishira Maiya , Max Ehrlich , Abhinav Shrivastava This is my paper
classification cs.CVcs.AIcs.LG
keywords SIEDDimplicitneuralrepresentationvideocompressionINRcodecsharedencoderdiscretedecoderscoordinatesamplingencoding
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

Video codecs built on implicit neural representations (INRs) store a video as the weights of a neural network, giving high fidelity and flexible decoding, but encoding a clip takes hours of per-video gradient training. SIEDD attacks this bottleneck by splitting the network into a shared coordinate-based encoder and many small per-frame-group decoders. The encoder is trained briefly on sparse anchor frames and then frozen, at which point the independent decoders can be trained in parallel and with aggressive coordinate subsampling. The paper claims this yields a 20–30$ imes$ encoding speed-up over prior INR codecs on HD and 4K benchmarks while keeping reconstruction quality and bitrate competitive, and it keeps 2D coordinate input so a single model can decode at any resolution without transcoding. If right, this removes the main practical obstacle to INR-based video compression for real-world, high-resolution use.

What carries the argument

The central object is the two-stage SIEDD architecture: a shared encoder $f_\theta$, an MLP with sine activations that maps positionally-encoded 2D coordinates to a latent vector, followed by a set of discrete decoders $g_{\phi,i}$ (one per frame group) that map the latent to RGB. Stage one overfits the encoder together with a small number of frame-specific decoders on $N_s$ anchor frames—deliberately without full convergence—so the encoder captures low-frequency features that generalize. The load-bearing mechanism is the fast convergence and transferability of early INR layers, which justifies freezing the encoder and training only per-group decoders, each initialized from the nearest anchor decoder, in parallel. Aggressive coordinate-space sampling (uniform random sampling of roughly $H\cdot W/1024$ coordinates per minibatch) plus weight sharing within each group and post-training quantization (HQQ plus Huffman and LZMA coding) complete the encoding pipeline.

What would settle it

Encode the same clip twice with SIEDD, once with anchor frames taken from the beginning and once from the end, and compare per-group decoder training time and final PSNR as a function of temporal distance from the anchors; the transfer claim predicts direction-independence, so if later groups systematically take longer or reconstruct worse, the frozen-encoder premise is falsified. Equivalently, a clip with an abrupt scene change after the anchor segment should sharply increase group training time.

Watch

Extended reading notes

Core claim

SIEDD claims that the slow per-video optimization that makes INR codecs impractical is largely redundant: the early layers of the network, which encode global low-frequency structure, can be learned once on a handful of anchor frames and shared across the whole video. The paper's central discovery is that after this shared encoder is frozen, each frame group's lightweight decoder converges quickly on its own, and because the decoders are independent they train in parallel across devices. Aggressive uniform random sampling of the 2D coordinate grid—down to roughly one part in 1024 of the pixels—speeds this further with minimal quality loss. The result is a reported 20$ imes$ faster encoding on UVG-HD and over 30$ imes$ on UVG-4K relative to state-of-the-art INR codecs at comparable PSNR and bitrate, and because the shared encoder takes 2D coordinates as input, decoding resolution can be chosen after encoding, which removes the need to transcode for adaptive streaming.

Load-bearing premise

Everything rests on the premise that a network trained briefly on a few anchor frames keeps describing the rest of the video well enough that each later frame group can be encoded with only a small, quickly-trained decoder; if that transfer fails on harder content, the reported 20–30$ imes$ speedup shrinks toward ordinary per-frame INR training times.

Editorial extensions

If this is right

  • Encoding a 600-frame 4K video with a Video-INR drops from over an hour to tens of minutes; the paper reports the first INR codec to encode such a clip in under an hour on a single RTX A5000.
  • Because the shared encoder maps 2D coordinates to latents, one encoded model serves all output resolutions, so adaptive streaming no longer requires resolution-specific transcoding.
  • Decoder independence gives near-linear scaling of encoding time with GPU count: the paper reports SIEDD-L speeds up ${\sim}8\times$ from 1 to 8 GPUs, where monolithic baselines plateau.
  • The shared encoder transfers across datasets: weights trained on UVG-HD initialized DA VIS encoding with similar PSNR and roughly a third of the time, pointing toward a reusable video prior.
  • Simple MLP layers keep the architecture compatible with post-training quantization; 6-bit HQQ gives a negligible (under 0.2 dB) drop versus 8-bit, keeping bitrates competitive.

Reading between the lines

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

  • The strongest implicit consequence is that per-video INR encoding becomes time-competitive with cloud transcoding: at a 20–30$ imes$ speed-up, the roughly $3-per-minute training cost the paper quotes drops toward the ~$0.04 cost of conventional encoding, which is the actual threshold for deployment. The paper does not compute this, but it follows directly from its own cost numbers.
  • The coordinate-sampling floor of about 1/1024 of pixels suggests the required number of training samples per frame is tied to the representational capacity of the decoder rather than the pixel count; a testable extension is to check whether that bound scales with latent dimension or scene complexity rather than with resolution.
  • If the shared encoder really is a reusable video prior that transfers across datasets, a natural extension—which the paper only gestures at in its conclusion—is updating the prior once on a large corpus and then doing near-zero-shot INR encoding of new videos, which would change the economics of neural video compression more than any per-clip optimization.
  • The same freeze-the-encoder recipe could transfer to other per-signal INR tasks (audio, 3D scenes, medical volumes) where per-signal optimization time is the bottleneck, though the paper only demonstrates video.
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.

Referee Report

3 major / 8 minor

Summary. The paper introduces SIEDD, a two-stage video INR codec for fast encoding of HD and 4K video. Stage 1 trains a shared coordinate-based MLP encoder on a small set of anchor frames, and stage 2 freezes that encoder and trains lightweight frame-group decoders in parallel while aggressively subsampling the coordinate space. The authors report 20x faster encoding than INR baselines on UVG-HD and up to 30x on UVG-4K at competitive PSNR/SSIM and bitrate, together with continuous-resolution decoding, parallel GPU scaling, and a transferable shared encoder. The paper also presents ablations on coordinate sampling rate, shared-encoder iterations, model width/depth, group size, quantization, and a cross-dataset transfer experiment from UVG-HD to DAVIS.

Significance. If the speed and quality claims hold, SIEDD would be a meaningful step toward practical per-video INR encoding: the architecture is simple, parallelizable, and retains coordinate-level control, and the shared-encoder transfer result (Table 1) suggests a reusable video prior. The paper's ablations (e.g., Table 4(a) and Table 4(b)) provide useful evidence for the design choices, and the code release supports reproducibility. The central quantitative claim, however, is not yet established because the baseline timing and resolution protocols are inconsistent; the reported 20-30x speedup is currently an artifact of uncontrolled baseline times rather than a matched comparison.

major comments (3)
  1. [Section 5.3, Supplementary A, Table 3] The hard 60-minute encoding budget stated in Section 5.3 and repeated in Supplementary A is contradicted by Table 3, which reports 4K baseline times of 3,660 s (NeRV), 10,200 s (HiNeRV), and 75,600 s (Nirvana). The headline 20-30x speedups are computed against these uncontrolled times, so the denominator is not budget-matched or quality-matched. To support the central claim, the comparison should be time-to-target-quality (or time-to-target-bitrate) with all methods run to comparable PSNR/SSIM, or all methods capped at the same budget and the resulting quality reported. As written, the speedup factor is an artifact of the baseline protocol.
  2. [Supplementary A.2, A.4, Table 3] The baseline comparison is also not capacity- or resolution-matched. Supplementary A.2 states that the 4K HiNeRV baseline uses the small 'S' configuration with an additional block, and Supplementary A.4 states that HNeRV outputs at 960x1920 (12% fewer pixels) for UVG-HD and 3600x2160 (7% fewer) for UVG-4K. The 'competitive reconstruction quality' claim in Table 3 and Figure 2 is therefore not an equal-resolution or equal-capacity comparison. Please rerun the baselines at matched output resolution and model scale, or justify the chosen settings, before the quality comparison can be accepted.
  3. [Table 3, Section 5.4] All metrics appear to be single-run point estimates with no variance or repeated-seed information. In Table 3, SIEDD-L has higher PSNR than Nirvana (35.41 vs 35.18 dB) but markedly lower SSIM (0.85 vs 0.93); the claim of competitive quality therefore depends on which metric is emphasized, and the margins are small enough to fall within plausible run-to-run variation. Please report means and standard deviations over multiple runs, or provide a clear justification for why single runs suffice for the central claim.
minor comments (8)
  1. [Section 1] The word 'imporving' should be 'improving'.
  2. [Sections 5.4, 5.4.1, 5.8, Figure 3] The model name is spelled inconsistently as 'SEIDD' in several places; it should be 'SIEDD' throughout.
  3. [Figure 5 caption] The Figure 5 caption incorrectly repeats the Figure 2 caption ('Comparison of our method and baselines on UVG...') instead of describing the two subfigures, which show FPS versus resolution and encoding time versus number of GPUs.
  4. [Section 2.2, Section 5.1] 'DA VIS' should be written as 'DAVIS' without a space, and 'NIRV ANA' in Section 2.2 contains an unwanted space.
  5. [Section 5.4] The phrase 'present the results for 3 different model configurations in 3' should reference Table 3 rather than just 'in 3'.
  6. [Table 2, Section 5.8] The long-video comparison mixes quality and bitrate: SIEDD reaches 31.10 dB at 0.325 BPP while HNeRV reaches 24.78 dB at 0.035 BPP, so the 'outperforms' claim is not rate-distortion controlled; a matched-bitrate comparison or rate-distortion curve would be clearer.
  7. [Supplementary F] In Table 8, the abbreviation 'Post' is never defined, and the text contains the typo 'degregation' for 'degradation'.
  8. [Section 5.2, Table 4(b)] Section 5.2 states that shared-encoder training and frame training both use 20,000 iterations for the main results, but Table 4(b) sweeps shared-encoder iterations only up to 5,000; please clarify which setting corresponds to the main tables.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the speedup and quality claims rest on external empirical benchmarks and ablations, not on self-referential construction.

full rationale

SIEDD's central claims, including the 20-30X encoding speedup, competitive reconstruction quality, and compression ratios, are supported by measured encoding times and reconstruction metrics on UVG-HD, UVG-4K, DAVIS, and a long-video sequence. None of these quantities is derived from an assumed target by construction. The two-stage design, in which a shared encoder is trained on sparse anchor frames and then frozen while per-group decoders are trained, is an architectural proposal whose transfer behavior is tested empirically, including a cross-dataset shared-encoder transfer experiment (Table 1) and an ablation of shared-encoder training iterations (Table 4b). The cited works [12,13] for early-layer convergence are external works by Vyas et al. and Kim et al., not author self-citations, and the convergence assumption is additionally supported by the paper's own ablation. Prior author works such as HNeRV, NIRVANA, and NeRV are cited as baselines or architectural predecessors, and no uniqueness theorem, ansatz, or fitted parameter is imported from them to force the reported speedup. The conclusion's caveat that systematic cross-video generalization and zero-shot inference remain unexplored is a stated limitation, not a circular step. The skeptic concern about unequal baseline time budgets, such as the 75,600-second Nirvana time in Table 3 exceeding the stated 1-hour cap, is a legitimate correctness or experimental-fairness concern about baseline matching, but it is not circularity: no equation or fitted parameter in the paper reduces to the target result by definition. Because no specific circular reduction can be exhibited, the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central speed claim rests on several hand-chosen hyperparameters and on two content-transfer assumptions about INR layer behavior. None of these are fitted constants in a predictive law, but the coordinate sampling ratio and group size are tuned empirically and directly control the reported speedup.

free parameters (3)
  • Group size Ng = Ns = 20
    Chosen by ablation in Table 6; larger groups reduce bitrate and encoding time but degrade PSNR, while smaller groups improve quality at higher cost.
  • Coordinate sampling ratio = C = H*W/1024
    Chosen as an approximate lower limit from the ablation in Table 4a; below 1/1024 quality drops noticeably, and above it encoding time rises without quality gain.
  • Shared encoder training iterations = 20000
    Set equally with decoder iterations; Table 4b shows PSNR gains from 500 to 5000 iterations are marginal, so this value balances speed and quality.
assumptions (4)
  • domain assumption Early INR layers converge quickly and transfer well across frames
    Invoked in Section 4.2 and cited to references [12,13]; this is the premise that lets SIEDD freeze the shared encoder after training on a small set of anchor frames.
  • domain assumption A shared encoder trained on Ns anchor frames captures enough global low-frequency structure for all frame groups
    Section 4.2 and 4.3 rely on this; if the frozen encoder is not representative of later frames, decoder training would slow down or quality would drop.
  • ad hoc to paper Coordinate subsampling at roughly 1/1024 preserves reconstruction quality across content
    Table 4a demonstrates this only on UVG-HD with SIEDD-M; the same ratio is applied to 4K and DAVIS without per-content tuning or additional validation.
  • ad hoc to paper A hard 60-minute encoding budget is a fair comparison protocol for all baselines
    Appendix A applies the budget to all baselines; methods that converge slowly are penalized, which can inflate SIEDD's relative speedup if baselines were not given enough time to reach their final quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SIEDD: Shared-Implicit Encoder with Discrete Decoders." pith.science (2026). https://pith.science/paper/QRLXLIOG

@misc{pith2026250623382,
  author       = {Pith},
  title        = {Pith review of: SIEDD: Shared-Implicit Encoder with Discrete Decoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QRLXLIOG}},
  note         = {Machine review of arXiv:2506.23382}
}
read the original abstract

Implicit Neural Representations (INRs) offer exceptional fidelity for video compression by learning per-video optimized functions, but their adoption is crippled by impractically slow encoding times. Existing attempts to accelerate INR encoding often sacrifice reconstruction quality or crucial coordinate-level control essential for adaptive streaming and transcoding. We introduce SIEDD (Shared-Implicit Encoder with Discrete Decoders), a novel architecture that fundamentally accelerates INR encoding without these compromises. SIEDD first rapidly trains a shared, coordinate-based encoder on sparse anchor frames to efficiently capture global, low-frequency video features. This encoder is then frozen, enabling massively parallel training of lightweight, discrete decoders for individual frame groups, further expedited by aggressive coordinate-space sampling. This synergistic design delivers a remarkable 20-30X encoding speed-up over state-of-the-art INR codecs on HD and 4K benchmarks, while maintaining competitive reconstruction quality and compression ratios. Critically, SIEDD retains full coordinate-based control, enabling continuous resolution decoding and eliminating costly transcoding. Our approach significantly advances the practicality of high-fidelity neural video compression, demonstrating a scalable and efficient path towards real-world deployment. Our codebase is available at https://github.com/VikramRangarajan/SIEDD .

Figures

Figures reproduced from arXiv: 2506.23382 by the authors.

Figure 1
Figure 1. Overview of the SIEDD architecture. During the shared encoder training phase (left), a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison of our method and baselines on UVG. Left: rate–distortion; Right: [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Video Reconstruction Visualization. We compare the reconstructions of SEIDD with other [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Super resolution visualization between baseline methods (nearest, bilinear, bicubic) and [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of our method and baselines on UVG. Left: rate–distortion; Right: [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Ablation study on decoder architecture. (a) PSNR vs. BPP when varying decoder layer [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Visual comparison between the ground truth frame, patching output, and the default SIEDD [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: FLIP Visualization on UVG-HD YachtRide using SIEDD-L [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 28 canonical work pages

  1. [1]

    Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand

    Gary J. Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the h.264/avc video coding standard. IEEE Transactions on Circuits and Systems for Video Tech- nology, 13(7):560–576, 2004

    work page 2004

  2. [2]

    Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand

    Gary J. Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649–1668, 2012. 13

    work page 2012

  3. [3]

    Sullivan, and Ye-Kui Wang

    Benjamin Bross, Jianle Chen, Jens-Rainer Ohm, Gary J. Sullivan, and Ye-Kui Wang. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736–3764, 2021

    work page 2021

  4. [4]

    Deep contextual video compression

    Li Li, Dong Liu, and Shiqi Wang. Deep contextual video compression. In Advances in Neural Information Processing Systems, volume 34, pages 17572–17583, 2021

    work page 2021

  5. [5]

    Neural video compression with feature modulation

    Li Li, Dong Liu, and Shiqi Wang. Neural video compression with feature modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1–10, 2024

    work page 2024

  6. [6]

    Towards practical real-time neural video compression

    Zhaoyang Jia, Bin Li, Jiahao Li, Wenxuan Xie, Linfeng Qi, Houqiang Li, and Yan Lu. Towards practical real-time neural video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    work page 2025

  7. [7]

    On the out-of-distribution generalization of probabilistic image modelling

    Mingtian Zhang, Andi Zhang, and Steven McDonagh. On the out-of-distribution generalization of probabilistic image modelling. Advances in Neural Information Processing Systems , 34: 3811–3823, 2021

    work page 2021

  8. [8]

    Hnerv: A hybrid neural representation for videos

    Hao Chen, Matthew Gwilliam, Ser-Nam Lim, and Abhinav Shrivastava. Hnerv: A hybrid neural representation for videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10270–10279, 2023

    work page 2023

Show all 38 references
  1. [9]

    Implicit neural representations for image compression

    Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In European Conference on Computer Vision, pages 74–91. Springer, 2022

    2022

  2. [10]

    Meta-learning sparse implicit neural representations

    Jaeho Lee, Jihoon Tack, Namhoon Lee, and Jinwoo Shin. Meta-learning sparse implicit neural representations. Advances in Neural Information Processing Systems , 34:11769–11780, 2021

    2021

  3. [11]

    Fast encoding and decoding for implicit video representation

    Hao Chen, Saining Xie, Ser-Nam Lim, and Abhinav Shrivastava. Fast encoding and decoding for implicit video representation. In European Conference on Computer Vision, pages 402–418. Springer, 2024

    2024

  4. [12]

    Baraniuk, Ashok Veer- araghavan, and Guha Balakrishnan

    Kushal Vyas, Ahmed Imtiaz Humayun, Aniket Dashpute, Richard G. Baraniuk, Ashok Veer- araghavan, and Guha Balakrishnan. Learning transferable features for implicit neural represen- tations, 2025. URL https://arxiv.org/abs/2409.09566

    2025 arXiv

  5. [13]

    Generalizable implicit neural representations via instance pattern composers

    Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural representations via instance pattern composers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11808–11817, 2023

    2023

  6. [14]

    Nirvana: Neural implicit representations of videos with adaptive networks and autoregressive patch-wise modeling

    Shishira R Maiya, Sharath Girish, Max Ehrlich, Hanyu Wang, Kwot Sin Lee, Patrick Poirson, Pengxiang Wu, Chen Wang, and Abhinav Shrivastava. Nirvana: Neural implicit representations of videos with adaptive networks and autoregressive patch-wise modeling. In Proceedings of the I...

    2023

  7. [15]

    Half-quadratic quantization of large machine learning models, November 2023

    Hicham Badri and Appu Shaji. Half-quadratic quantization of large machine learning models, November 2023. URL https://mobiusml.github.io/hqq_blog/

    2023

  8. [16]

    Llm.int8(): 8-bit matrix multiplication for transformers at scale

    Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Llm.int8(): 8-bit matrix multiplication for transformers at scale. arXiv preprint arXiv:2208.07339, 2022

    2022 arXiv

  9. [17]

    Uvg dataset: 50/120fps 4k sequences for video codec analysis and development

    Alexandre Mercat, Marko Viitanen, and Jarno Vanne. Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM Multimedia Systems Conference, MMSys ’20, page 297–302, New York, NY , USA, 2020. Association for Computing Machinery...

    2020

  10. [18]

    Benchmarking conventional and learned video codecs with a low-delay configuration

    Siyue Teng, Yuxuan Jiang, Ge Gao, Fan Zhang, Thomas Davis, Zoe Liu, and David Bull. Benchmarking conventional and learned video codecs with a low-delay configuration. In 2024 IEEE International Conference on Visual Communications and Image Processing (VCIP), pages 1–5. IEEE, 2024. 14

    2024

  11. [19]

    Variational image compression with a scale hyperprior

    Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018

    2018 arXiv

  12. [20]

    Im- plicit neural representations with periodic activation functions

    Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Im- plicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462–7473, 2020

    2020

  13. [21]

    Nerv: Neural representations for videos

    Hao Chen, Bo He, Hanyu Wang, Yixuan Ren, Ser Nam Lim, and Abhinav Shrivastava. Nerv: Neural representations for videos. Advances in Neural Information Processing Systems , 34: 21557–21568, 2021

    2021

  14. [22]

    Ae- nerf: Audio enhanced neural radiance field for few shot talking head synthesis

    Dongze Li, Kang Zhao, Wei Wang, Bo Peng, Yingya Zhang, Jing Dong, and Tieniu Tan. Ae- nerf: Audio enhanced neural radiance field for few shot talking head synthesis. arXiv preprint arXiv:2312.10921, 2023

    2023 arXiv

  15. [23]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405–421. Springer, 2020

    2020

  16. [24]

    Hinerv: Video compres- sion with hierarchical encoding-based neural representation

    Ho Man Kwan, Ge Gao, Fan Zhang, Andrew Gower, and David Bull. Hinerv: Video compres- sion with hierarchical encoding-based neural representation. Advances in Neural Information Processing Systems, 36:72692–72704, 2023

    2023

  17. [25]

    Ffnerv: Flow-guided frame- wise neural representations for videos

    Joo Chan Lee, Daniel Rho, Jong Hwan Ko, and Eunbyung Park. Ffnerv: Flow-guided frame- wise neural representations for videos. InProceedings of the 31st ACM International Conference on Multimedia, pages 7859–7870, 2023

    2023

  18. [26]

    Tree-nerv: A tree-structured neural representation for efficient non-uniform video encoding

    Jiancheng Zhao, Yifan Zhan, Qingtian Zhu, Mingze Ma, Muyao Niu, Zunian Wan, Xiang Ji, and Yinqiang Zheng. Tree-nerv: A tree-structured neural representation for efficient non-uniform video encoding. arXiv preprint arXiv:2504.12899, 2025

    2025 arXiv

  19. [27]

    Ds-nerv: Im- plicit neural video representation with decomposed static and dynamic codes

    Hao Yan, Zhihui Ke, Xiaobo Zhou, Tie Qiu, Xidong Shi, and Dadong Jiang. Ds-nerv: Im- plicit neural video representation with decomposed static and dynamic codes. arXiv preprint arXiv:2403.15679, 2024

    2024 arXiv

  20. [28]

    Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding

    Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations (ICLR), 2016

    2016

  21. [29]

    The lottery ticket hypothesis: Finding sparse, trainable neural networks

    Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations (ICLR) , 2019

    2019

  22. [30]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference

    Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE Conference on Computer Vision an...

    2018

  23. [31]

    David A. Huffman. A method for the construction of minimum-redundancy codes. Proceedings of the IRE, 40(9):1098–1101, 1952

    1952

  24. [32]

    Perazzi, J

    F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016

    2016

  25. [33]

    Youtube-8m: A large-scale video classification benchmark, 2016

    Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark, 2016. URL https://arxiv.org/abs/1609.08675

    2016 arXiv

  26. [34]

    The road less scheduled, 2024

    Aaron Defazio, Xingyu Alice Yang, Harsh Mehta, Konstantin Mishchenko, Ahmed Khaled, and Ashok Cutkosky. The road less scheduled, 2024. URL https://arxiv.org/abs/2405. 15682. 15

    2024

  27. [35]

    Fairchild

    Pontus Andersson, Jim Nilsson, Tomas Akenine-Möller, Magnus Oskarsson, Kalle Åström, and Mark D. Fairchild. FLIP: A Difference Evaluator for Alternating Images. Proceedings of the ACM on Computer Graphics and Interactive Techniques , 3(2):15:1–15:23, 2020. doi: 10.1145/3406183

    2020 doi

  28. [36]

    Vmaf re-implementation on pytorch: Some experimental results, 2024

    Kirill Aistov and Maxim Koroteev. Vmaf re-implementation on pytorch: Some experimental results, 2024. URL https://arxiv.org/abs/2310.15578

    2024 arXiv

  29. [37]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. URL https://arxiv.org/abs/2106.09685

    2021 arXiv

  30. [38]

    Efficient learning with sine-activated low-rank matrices, 2025

    Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Efficient learning with sine-activated low-rank matrices, 2025. URL https://arxiv.org/abs/2403. 19243. 16

    2025

Pith tools

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