Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Q-VDiT shows video diffusion transformers can be quantized to 3-bit weights and 6-bit activations without losing scene coherence.

desk verdict The W3A6 scene-consistency jump (23.40 vs 12.04) looks real and makes the paper worth refereeing, but the entropy theorem is overclaimed as a foundation for the rank-1 estimator. read the letter →

arxiv 2505.22167 v1 pith:N5XZHOW7 submitted 2025-05-28 cs.CV

Weilun Feng , Chuanguang Yang , Haotong Qin , Xiangqi Li , Yu Wang , Zhulin An , Libo Huang , Boyu Diao
show 3 more authors
Zixiang Zhao Yongjun Xu Michele Magno
This is my paper
classification cs.CV
keywords videodiffusiontransformerpost-trainingquantizationlow-bitgenerationtoken-awaretemporalmaintenancedistillationrank-1errorcompensationsceneconsistencyW3A6
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

The paper is trying to establish that aggressive low-bit quantization (W3A6) of video-generation diffusion transformers does not have to destroy temporal and scene coherence. Existing quantization methods for image DiTs lose too much information because they ignore the extra token dimension that video frames introduce, and their MSE objective does not see cross-frame dependencies. The paper's two mechanisms are a Token-aware Quantization Estimator (TQE), which models the quantization error as a low-rank outer product scaled per-frame, and Temporal Maintenance Distillation (TMD), which aligns the full-precision and quantized models' distributions of frame-to-frame similarity. On VBench, the W3A6 model reaches scene consistency 23.40, nearly doubling the best prior result (12.04), and at W4A6 it is near lossless on several quality metrics. If the claim holds, video DiTs can run at a fraction of their memory and latency cost without the scene collapse seen in earlier PTQ methods.

What carries the argument

The Token-aware Quantization Estimator (TQE) is a rank-1 error compensation: after weights and activations are quantized, a correction $(M \odot \hat Q(X))\alpha\beta^{\top}$ is added, with $M$ initialized from the similarity between full-precision and quantized frame activations and from token saliency, and $\alpha, \beta$ trainable vectors with only $d_{\text{in}} + d_{\text{out}}$ parameters per layer. The Temporal Maintenance Distillation (TMD) builds a temporal relation distribution $D_i = \mathrm{softmax}([T_{i,1},\dots,T_{i,t}])$ from cosine similarities $T_{i,j}$ between frames and minimizes $\sum_i \mathrm{KL}(D_i^{FP}, D_i^Q)$, so the gradient for any one frame depends on every frame. The paper anchors the low-rank approach in Theorem 3.2, which states that the quantization error has entropy no larger than the original weight, $H(\Delta) \le H(W)$, so the error needs fewer parameters to represent than the weight itself.

What would settle it

Run the already-calibrated W3A6 Open-Sora model on a batch of video tokens, form the residual $R = X W^\top - \hat Q(X)\hat Q(W)^\top$ after the per-frame scaling $M$ is applied, and compute an SVD of $R$; if the top singular value accounts for less than about half of the squared Frobenius norm, the rank-1 assumption is empirically false. A cheaper companion check is to retrain TQE with rank 2 and rank 4: if scene consistency does not improve with rank, the model is not limited by a rank-1 error subspace.

Watch

Extended reading notes

Core claim

The central claim is that the quantization error of a video DiT is structured, not noise: it has lower information entropy than the original weights, so it can be absorbed by a small parameter budget. Concretely, the paper replaces the quantized matmul $X W^{\top} \approx \hat Q(X)\hat Q(W)^{\top} + (M \odot \hat Q(X))\alpha\beta^{\top}$, where $M$ is a per-frame scaling vector, $\alpha \in \mathbb{R}^{d_{\text{in}}}$, and $\beta \in \mathbb{R}^{d_{\text{out}}}$. This correction is meant to compensate information loss along both the token dimension (which frames suffer more from quantization) and the feature dimension. For optimization, standard MSE treats each frame independently, so the paper instead computes a distribution $D_i$ of cosine similarities between frame $i$ and every other frame and matches these distributions between full-precision and quantized models using KL divergence. With both components, the paper reports W3A6 scene consistency of 23.40 on VBench, versus 12.04 for the best baseline, and near-full-precision quality at W4A6.

Load-bearing premise

The error-correction story assumes that, after per-frame scaling, the activation quantization error can be written as one rank-1 outer product $X\alpha\beta^\top$; the entropy theorem only says the error has low information content, not that it is rank-1, so if the true residual has rich structure across tokens, TQE behaves as ordinary learned fine-tuning rather than as the error estimator the paper describes.

Editorial extensions

If this is right

  • W3A6 video DiTs no longer collapse: the 3-bit model keeps a scene-consistency score of 23.40, roughly double the previous best, making low-bit weights viable for text-to-video generation.
  • At W4A6 the quantized Open-Sora model matches or exceeds the FP16 model on VQA-Aesthetic (67.05 vs 66.91) and VQA-Technical (53.75 vs 53.49), so the quantized model is effectively lossless on those axes.
  • Deployment cost drops: the W4A8 quantized model uses 2.40x less GPU memory and runs 1.35x faster than FP16 in the reported setting.
  • TQE adds only $d_{\text{in}} + d_{\text{out}}$ parameters per layer and can be fused with the low-bit kernels, so the correction does not erase the memory and latency savings of quantization.
  • TMD is robust to its hyperparameter: varying $\gamma$ from 50 to 200 keeps the improvements large, so the method does not need fine per-task tuning.

Reading between the lines

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

  • The rank-1 outer-product form of TQE is the strong step in the argument: Theorem 3.2 guarantees only low entropy, not that the error is rank-1 after per-frame scaling, so part of the gain may come from the trained vectors acting as a general correction rather than from faithful estimation of the quantization error.
  • TMD optimizes relative frame relationships rather than absolute frame content, so the same loss could transfer to other video-level compression tasks, such as pruning or low-bit video autoencoders, where temporal coherence is the bottleneck.
  • Because the headline comparison centers on VBench scene consistency, the '1.9x' claim is benchmark-specific; the paper's own tables show smaller but consistent gains on imaging quality, subject consistency, and technical quality.
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 / 5 minor

Summary. The paper proposes Q-VDiT, a post-training quantization framework for video-generation diffusion transformers. Two components are introduced: a Token-aware Quantization Estimator (TQE), which uses a low-rank correction in the token and feature dimensions to compensate for quantization error, and Temporal Maintenance Distillation (TMD), which aligns the inter-frame similarity distributions of the quantized model with those of the full-precision model via KL divergence. The method is evaluated on Open-Sora and Latte at W4A6, W3A8, and W3A6 settings against several prior quantization approaches, with the headline result being a VBench scene-consistency score of 23.40 at W3A6, roughly 1.9x the best prior method. Ablations show that both TQE and TMD contribute to the reported gains.

Significance. If the empirical results hold, Q-VDiT would be a practically valuable step toward deploying low-bit video DiTs on resource-constrained devices, and the TMD temporal-distribution distillation idea is a sensible and relatively novel contribution to quantization-aware fine-tuning for video. The paper has real strengths: all baselines are rerun under a common protocol, the VBench headline result is measured on VBench prompt sets rather than on the calibration prompts, additional results are provided on Latte/UCF-101, and the ablation study isolates the contributions of the two components. However, the theoretical claim underpinning TQE is not supported by the supplied proof, and part of the multi-aspects evaluation is performed on the same prompts used for calibration. These issues affect how the contributions should be framed, but they do not by themselves overturn the main empirical result.

major comments (3)
  1. [Section 3.2, Theorem 3.2, and Eqs. (7)-(8)] The claim that Theorem 3.2 theoretically supports the TQE construction is not established. Theorem 3.2 proves only that the scalar quantization residual of each weight element has no larger entropy than the original element, via a deterministic-mapping argument. That is a statement about marginal per-element entropies; it does not imply that the matrix E = XW^T - Q(X)Q(W)^T is close to rank-1, nor that the correction (M K Q(X) alpha) beta^T can represent the dominant part of E. The TQE correction is exactly a rank-1 outer product in token/output space, so assuming it captures the quantization error is a modeling assumption, not a consequence of the entropy bound. In fact, as Appendix F acknowledges, the module is equivalent to a rank-1 LoRA branch, so it is better described as a low-rank fine-tuning correction than as an estimator whose accuracy follows from the theorem. The 'theoretically prove' bullet in the contributions should be removed or substantially softened, and the paper should either prove a low-rank property of E or present TQE as a heuristic. A concrete empirical diagnostic would be to report the singular-value spectrum of E on calibration layers; if the top singular value does not dominate, the stated link between the theorem and the method should be abandoned.
  2. [Section 4.1, Tables 2-3, and Appendix B] The multi-aspects evaluation is circular with respect to calibration. Section 4.1 states that 10 OpenSora prompts are used to generate videos for the multi-aspects metrics, and Appendix B states that the same 10 OpenSora prompts are used as the calibration dataset. Tables 2 and 3 therefore report performance on the calibration prompts, not generalization to unseen prompts. The claims in Section 4.2 that Q-VDiT achieves 'notable improvement across all metrics' and that W4A8 surpasses the full-precision model are weakened by this overlap. The VBench results in Table 1 use the VBench prompt sets and are not subject to this concern. Please evaluate the multi-aspects metrics on a disjoint prompt set, or explicitly label Tables 2 and 3 as calibration-set results and avoid claiming generalization from them.
  3. [Tables 1-3 and Section 4.2] All quantitative comparisons are based on a single generated video per prompt, with no error bars or repeated-seed statistics. Given that video generation and VBench evaluation are stochastic, the reported margins, especially the large Dynamic Degree and VQA improvements, would be more convincing with at least two or three seeds and a reported mean or variance. This is not a fatal flaw, but it should be addressed in a revision so the reader can assess whether the improvements exceed sampling noise.
minor comments (5)
  1. [Eq. (8)] The symbol K in Eq. (8) is never defined; the notation 'M_i K Q(X)[...]' should be clarified, for instance by specifying whether K is a per-frame scaling factor, a learned matrix, or a typo.
  2. [Appendix A, Eq. (23)] The second and third cases in Eq. (23) are not stated correctly: the first two cases have the same conclusion, and the third case repeats 'if Wi != Wj and fractional parts equal' in a way that is contradictory. The proof also should specify whether discrete or differential entropy is being used, since the scaling by s changes differential entropy.
  3. [Table 3] The rows for Q-DiT 6/6 and Q-DiT 4/8 appear misaligned or garbled; for example, the Q-DiT 4/8 row reports VQA-Aesthetic of 0.007 and VQA-Technical of 0.018, which are implausible relative to neighboring rows. Please reformat the table and verify each entry against the correct column.
  4. [Figure 4] The label 'L_temporal = 0' for the Q-Model without TMD is confusing; if TMD is not applied, the temporal loss is not computed, so reporting a value of 0 should be replaced with 'N/A' or an explicit note that the loss is not optimized.
  5. [Section 4.2, Table 3] The statement that Q-VDiT at W4A8 outperforms the full-precision model on VQA-Aesthetic (71.32 vs. 66.91) should be tempered, since the evaluation is on the calibration prompts (as noted in the second major comment) and could reflect calibration-set overfitting rather than genuine improvement.

Circularity Check

1 steps flagged · score 3.0 of 10

Headline VBench claim is clean, but Tables 2-3 evaluate on the same 10 prompts used for calibration; Theorem 3.2 does not actually support TQE's rank-1 form.

  1. fitted input called prediction [Sec. 4.1 (Experimental and Evaluation Settings) and Appendix B (Implementation Details); Tables 2 and 3]
    "For calibration dataset, we use 10 prompts provided by Open-Sora (HPC-AI, 2024) and uniformly select 50 steps as used in ViDiT-Q (Zhao et al., 2024). [Sec. 4.1:] use 10 prompts provided by OpenSora prompt sets to generate 10 videos for multi-aspects metrics evaluation."

    The TQE parameters (alpha, beta, M in Eq. 8) and the quantization parameters are fine-tuned on exactly the 10 OpenSora prompts described in Appendix B, and Section 4.1 then reports the 'multi-aspects metrics evaluation' on videos generated from the same 10 prompts, with Tables 2-3 presented as results. The fitted parameters directly determine the quantized outputs that are scored, so the gains in Tables 2-3 are partially forced by calibration-set overfitting rather than generalization. This is partial: all baselines share the same calibration set, the evaluation metrics are downstream of the MSE/KL training losses, and the headline VBench numbers in Table 1 use a disjoint prompt set, so the central claim is not wholly fitted.

full rationale

The central claim is not circular: calibration (Appendix B) uses 10 OpenSora prompts, while Table 1 evaluates on VBench's 93/72/86-prompt sets, so the headline Scene Consistency of 23.40 is a held-out measurement. The one construction-level circularity I can exhibit is the calibration/evaluation overlap: the same 10 OpenSora prompts are used both to calibrate TQE and quantization parameters and to produce Tables 2-3, so those tables should be read as calibration-set fit quality rather than generalization. This is partial because the baselines calibrate on the same prompts and the reported metrics are not the training losses; notably, Table 3's W4A8 Q-VDiT exceeds the FP teacher on calibration prompts (VQA-Aesthetic 71.32 vs 66.91), a signature of fitting. The gamma=100 choice in Eq. (19) is selected by ablation (Fig. 6) on the same Open-Sora/VBench configuration reported in the main tables, a mild test-set-tuning concern rather than a construction-level circularity. The claimed theoretical foundation does not carry the method: Theorem 3.2 proves only the scalar marginal entropy bound H(Delta) <= H(W) for elementwise quantization error (a standard data-processing fact, proved internally in Appendix A, though the proof of Lemma A.1 is itself not rigorous), whereas Eq. (8) asserts the full error is captured by a rank-1 per-frame correction; lower marginal entropy does not imply small matrix rank, so this is an unsupported foundational leap and a correctness risk, but not an equation-level circular reduction. No load-bearing self-citation appears: the authors' prior works (MPQ-DM, Relational Diffusion Distillation, Cross-image Relational KD) are background citations, and Appendix F transparently identifies TQE as equivalent to a rank-1 LoRA module. Overall: one partial fitted-input-called-prediction step, with otherwise independent empirical content.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method's concrete contribution rests on fitted low-rank correction vectors and a hand-tuned distillation weight. The information-theoretic theorem does not establish the rank-1 structure, so the rank-1 model is an ad hoc modeling assumption validated only empirically. No new particles, forces, conserved quantities, or architectural primitives are introduced beyond the two loss/forward-pass components.

free parameters (4)
  • TQE low-rank vectors alpha and beta (per layer) = not reported
    Eq. 7-8 model quantization error as X alpha then project with beta; both vectors are trained during calibration with learning rate 1e-5 (App. B).
  • Token-aware frame scaling M = not reported
    Eq. 8-9: per-frame scalar multipliers initialized by similarity and saliency, then updated during calibration; this is central to balancing quantization error across frames.
  • TMD weight gamma = 100
    Eq. 19: chosen by ablation in Fig. 6 among 50, 100, and 200; a hand-tuned hyperparameter, not derived.
  • Quantization clipping thresholds and scales = not reported
    Eq. 3: the paper states that post-training calibration fine-tunes quantization parameters, and App. B lists a learning rate of 1e-6 for weight quantization parameters, so effective scales are fitted rather than fixed a priori.
assumptions (5)
  • ad hoc to paper Discrete entropy reduction under quantization (Theorem 3.2)
    The proof in App. A treats quantization error as a deterministic mapping of weights and uses H(f(X)) <= H(X). It assumes discrete entropy and a surjection with no more outcomes, and then leaps from lower entropy to low-rank representability.
  • ad hoc to paper Rank-1 separable correction captures quantization error
    Eq. 7 approximates X(W^T - Q(W)^T) by X alpha beta^T. No theorem shows the residual is rank-1 in the token dimension; lower entropy does not imply this structural form.
  • domain assumption Cosine similarity between frame token sequences is the right temporal relation measure
    Eq. 13 uses rho(S_i, S_j) without definition; the gradient in Eq. 18 reveals rho is cosine similarity. Whether cosine similarity over averaged frame tokens preserves video semantics is assumed, not demonstrated.
  • domain assumption Matching full-precision inter-frame distributions improves quantized video quality
    TMD (Eq. 15) assumes KL alignment of frame-pair distributions transfers to human-perceived temporal coherence; no direct evidence isolates this effect beyond VBench metrics.
  • standard math Standard diffusion and quantization definitions
    Eqs. 1-4 restate the standard DDPM forward/reverse process and uniform quantization; these are background definitions used throughout and not contested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers." pith.science (2026). https://pith.science/paper/N5XZHOW7

@misc{pith2026250522167,
  author       = {Pith},
  title        = {Pith review of: Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N5XZHOW7}},
  note         = {Machine review of arXiv:2505.22167}
}
abstract

Diffusion transformers (DiT) have demonstrated exceptional performance in video generation. However, their large number of parameters and high computational complexity limit their deployment on edge devices. Quantization can reduce storage requirements and accelerate inference by lowering the bit-width of model parameters. Yet, existing quantization methods for image generation models do not generalize well to video generation tasks. We identify two primary challenges: the loss of information during quantization and the misalignment between optimization objectives and the unique requirements of video generation. To address these challenges, we present Q-VDiT, a quantization framework specifically designed for video DiT models. From the quantization perspective, we propose the Token-aware Quantization Estimator (TQE), which compensates for quantization errors in both the token and feature dimensions. From the optimization perspective, we introduce Temporal Maintenance Distillation (TMD), which preserves the spatiotemporal correlations between frames and enables the optimization of each frame with respect to the overall video context. Our W3A6 Q-VDiT achieves a scene consistency of 23.40, setting a new benchmark and outperforming current state-of-the-art quantization methods by 1.9$\times$. Code will be available at https://github.com/cantbebetter2/Q-VDiT.

Figures

Figures reproduced from arXiv: 2505.22167 by the authors.

Figure 1
Figure 1. Evaluation on VBench of different quantization methods under W3A6 setting. of generative tasks, including image generation (Ho et al., 2020; Rombach et al., 2022; Dhariwal & Nichol, 2021; Feng et al., 2024; Yang et al., 2025), image super-resolution (Lin et al., 2025; Wang et al., 2024b; Wu et al., 2024b), and video generation (HPC-AI, 2024; Ma et al., 2024). Diffusion Transformers (DiT) (Peebles & Xie, 2023) have e… view at source ↗
Figure 2
Figure 2. Overview of proposed Q-VDiT. The framework includes Token-aware Quantization Estimator (TQE) for forward process and Temporal Maintenance Distillation (TMD) for optimization. The middle part denotes the quantized forward process. ⊗ denotes matrix multiplication, ⊙ denotes token-wise multiplication. generation. Directly applying existing quantization meth￾ods (Li et al., 2023; Ashkboos et al., 2024; Chen et al., 2024… view at source ↗
Figure 3
Figure 3. An illustration of TQE in Q-VDiT accelerate the inference. For a floating vector xf , the quan￾tization process can be formulated as xˆq = Q(xf , s, z) = clip(⌊ xf s ⌉ + z, 0, 2 N − 1), s = u − l 2N − 1 , z = −⌊ l s ⌉, (3) where xˆq indicates quantized vector in integer, ⌊·⌉ is round fuction and clip(·) is function that clamps values into the range of [0, 2 N −1], s is a scale factor and z is a quantization zero poi… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: An illustration of TMD in Q-VDiT. We have enlarged the upper left and lower right corners additionally. output error between the full-precision (FP) model and the quantization model by Ltask = ||OF P (X;W), OQ(Qˆ(X); Qˆ(W))||2 , (10) where OF P (X;W) ∈ R n×d represents…
Figure 6
Figure 6. Figure 6: Ablation study on different γ in TMD. mance compared with baseline PTQ4DiT. This shows that our TMD is not sensitive to hyperparameter selection. In our practice, we use γ = 100 for balanced choice. 4.6. Training Resource Cost In Tab. 5, we present the training cost of…
Figure 5
Figure 5. Figure 5: Visualization of different frames in a single video. 4.4. Qualitative Comparison In [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: The qualitative results with prompt “A soaring drone footage captures the majestic beauty of a coastal cliff, its red and yellow stratified rock faces rich in color and against the vibrant turquoise of the sea. Seabirds can be seen taking flight around the cliff’s prec…
Figure 8
Figure 8. Figure 8: The qualitative results with prompt “The video captures the majestic beauty of a waterfall cascading down a cliff into a serene lake. The waterfall, with its powerful flow, is the central focus of the video. The surrounding landscape is lush and green, with trees and f…
Figure 9
Figure 9. Figure 9: The qualitative results with prompt “The vibrant beauty of a sunflower field. The sunflowers, with their bright yellow petals and dark brown centers, are in full bloom, creating a stunning contrast against the green leaves and stems. The sunflowers are arranged in neat…
Figure 10
Figure 10. Figure 10: The qualitative results with prompt “A serene underwater scene featuring a sea turtle swimming through a coral reef. The turtle, with its greenish-brown shell, is the main focus of the video, swimming gracefully towards the right side of the frame. The coral reef, tee…
Figure 11
Figure 11. Figure 11: The qualitative results with prompt “A vibrant underwater scene. A group of blue fish, with yellow fins, are swimming around a coral reef. The coral reef is a mix of brown and green, providing a natural habitat for the fish. The water is a deep blue, indicating a dept…
Figure 12
Figure 12. Figure 12: The qualitative results with prompt “A bustling city street at night, filled with the glow of car headlights and the ambient light of streetlights. The scene is a blur of motion, with cars speeding by and pedestrians navigating the crosswalks. The cityscape is a mix o…
Figure 13
Figure 13. Figure 13: The qualitative results with prompt “A snowy forest landscape with a dirt road running through it. The road is flanked by trees covered in snow, and the ground is also covered in snow. The sun is shining, creating a bright and serene atmosphere. The road appears to be…
Figure 14
Figure 14. Figure 14: The qualitative results with prompt “The dynamic movement of tall, wispy grasses swaying in the wind. The sky above is filled with clouds, creating a dramatic backdrop. The sunlight pierces through the clouds, casting a warm glow on the scene. The grasses are a mix of…
Figure 15
Figure 15. Figure 15: The qualitative results with prompt “A serene night scene in a forested area. The first frame shows a tranquil lake reflecting the star-filled sky above. The second frame reveals a beautiful sunset, casting a warm glow over the landscape. The third frame showcases the…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MPQ-DMv2: Flexible Residual Mixed Precision Quantization for Low-Bit Diffusion Models with Temporal Distillation

    cs.CV 2025-07 conditional novelty 6.0 of 10

    MPQ-DMv2 adds binary residual quantization, temporal relation distillation, and SVD-initialized LoRA to mixed-precision quantization, improving low-bit diffusion model generation quality.

  2. Charting the Future of Scholarly Knowledge with AI: A Community Perspective

    cs.DL 2025-08 unverdicted novelty 2.0 of 10

    A community perspective on how AI can support scholarly knowledge extraction, organization, and communication, with a proposed classification and ethical considerations.

Reference graph

Works this paper leans on

52 extracted references · 27 canonical work pages · cited by 2 Pith papers

  1. [1]

    L., Li, B., Cameron, P., Jaggi, M., Alistarh, D., Hoefler, T., and Hensman, J

    Ashkboos, S., Mohtashami, A., Croci, M. L., Li, B., Cameron, P., Jaggi, M., Alistarh, D., Hoefler, T., and Hensman, J. Quarot: Outlier-free 4-bit inference in rotated llms. arXiv preprint arXiv:2404.00456, 2024

    arXiv 2024

  2. [2]

    Flux.1, 2024

    Black-Forest-Labs. Flux.1, 2024. URL https://blackforestlabs.ai/. 2, 4, 7

    work page 2024

  3. [3]

    W., Fidler, S., and Kreis, K

    Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S. W., Fidler, S., and Kreis, K. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 22563--22575, 2023

    work page 2023

  4. [4]

    Q-dit: Accurate post-training quantization for diffusion transformers

    Chen, L., Meng, Y., Tang, C., Ma, X., Jiang, J., Wang, X., Wang, Z., and Zhu, W. Q-dit: Accurate post-training quantization for diffusion transformers. arXiv preprint arXiv:2406.17343, 2024

    arXiv 2024

  5. [5]

    T., Mittal, S., Emani, M., Vishwanath, V., and Somani, A

    Chitty-Venkata, K. T., Mittal, S., Emani, M., Vishwanath, V., and Somani, A. K. A survey of techniques for optimizing transformer inference. Journal of Systems Architecture, pp.\ 102990, 2023

    work page 2023

  6. [6]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34: 0 8780--8794, 2021

    2021

  7. [7]

    Reg-ptq: Regression-specialized post-training quantization for fully quantized object detector

    Ding, Y., Feng, W., Chen, C., Guo, J., and Liu, X. Reg-ptq: Regression-specialized post-training quantization for fully quantized object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 16174--16184, 2024

    work page 2024

  8. [8]

    Perceptual quality assessment of smartphone photography

    Fang, Y., Zhu, H., Zeng, Y., Ma, K., and Wang, Z. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3677--3686, 2020

    2020

Show all 52 references
  1. [9]

    Relational diffusion distillation for efficient image generation

    Feng, W., Yang, C., An, Z., Huang, L., Diao, B., Wang, F., and Xu, Y. Relational diffusion distillation for efficient image generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pp.\ 205--213, 2024

    2024

  2. [10]

    Mpq-dm: Mixed precision quantization for extremely low bit diffusion models

    Feng, W., Qin, H., Yang, C., An, Z., Huang, L., Diao, B., Wang, F., Tao, R., Xu, Y., and Magno, M. Mpq-dm: Mixed precision quantization for extremely low bit diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 16595--16603, 2025

    2025

  3. [11]

    W., and Keutzer, K

    Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M. W., and Keutzer, K. A survey of quantization methods for efficient neural network inference. In Low-Power Computer Vision, pp.\ 291--326. Chapman and Hall/CRC, 2022

    2022

  4. [12]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    He, K., Zhang, X., Ren, S., and Sun, J. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pp.\ 1026--1034, 2015

    2015

  5. [13]

    Efficientdm: Efficient quantization-aware fine-tuning of low-bit diffusion models

    He, Y., Liu, J., Wu, W., Zhou, H., and Zhuang, B. Efficientdm: Efficient quantization-aware fine-tuning of low-bit diffusion models. arXiv preprint arXiv:2310.03270, 2023

    2023 arXiv

  6. [14]

    Ptqd: Accurate post-training quantization for diffusion models

    He, Y., Liu, L., Liu, J., Wu, W., Zhou, H., and Zhuang, B. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024

    2024

  7. [15]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  8. [16]

    Open-sora, 2024

    HPC-AI . Open-sora, 2024. URL https://github.com/hpcaiitech/Open-Sora

    2024

  9. [17]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  10. [18]

    Tfmq-dm: Temporal feature maintenance quantization for diffusion models

    Huang, Y., Gong, R., Liu, J., Chen, T., and Liu, X. Tfmq-dm: Temporal feature maintenance quantization for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 7362--7371, 2024 a

    2024

  11. [19]

    Vbench: Comprehensive benchmark suite for video generative models

    Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., Jiang, Y., Zhang, Y., Wu, T., Jin, Q., Chanpaisit, N., et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 21807--2181...

    2024

  12. [20]

    Musiq: Multi-scale image quality transformer

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., and Yang, F. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 5148--5157, 2021

    2021

  13. [21]

    Aesthetic-predictor, 2022

    LAION-AI. Aesthetic-predictor, 2022. URL https://github.com/LAION-AI/aesthetic-predictor. Accessed: 2022-04-16

    2022

  14. [22]

    Svdqunat: Absorbing outliers by low-rank components for 4-bit diffusion models

    Li, M., Lin, Y., Zhang, Z., Cai, T., Li, X., Guo, J., Xie, E., Meng, C., Zhu, J.-Y., and Han, S. Svdqunat: Absorbing outliers by low-rank components for 4-bit diffusion models. arXiv preprint arXiv:2411.05007, 2024 a

    2024

  15. [23]

    Q-diffusion: Quantizing diffusion models

    Li, X., Liu, Y., Lian, L., Yang, H., Dong, Z., Kang, D., Zhang, S., and Keutzer, K. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 17535--17545, 2023

    2023

  16. [24]

    Q-dm: An efficient low-bit quantized diffusion model

    Li, Y., Xu, S., Cao, X., Sun, X., and Zhang, B. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024 b

    2024

  17. [25]

    Diffbir: Toward blind image restoration with generative diffusion prior

    Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., Qiao, Y., Ouyang, W., and Dong, C. Diffbir: Toward blind image restoration with generative diffusion prior. In European Conference on Computer Vision, pp.\ 430--448. Springer, 2025

    2025

  18. [26]

    Evalcrafter: Benchmarking and evaluating large video generation models

    Liu, Y., Cun, X., Liu, X., Wang, X., Zhang, Y., Chen, H., Liu, Y., Zeng, T., Chan, R., and Shan, Y. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 22139--22149, 2024

    2024

  19. [27]

    Reactnet: Towards precise binary neural network with generalized activation functions

    Liu, Z., Shen, Z., Savvides, M., and Cheng, K.-T. Reactnet: Towards precise binary neural network with generalized activation functions. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XIV 16, pp.\ 143--159. Springer, 2020

    2020

  20. [28]

    Terdit: Ternary diffusion models with transformers

    Lu, X., Zhou, A., Lin, Z., Liu, Q., Xu, Y., Zhang, R., Wen, Y., Ren, S., Gao, P., Yan, J., et al. Terdit: Ternary diffusion models with transformers. arXiv preprint arXiv:2405.14854, 2024

    2024 arXiv

  21. [29]

    Latte: Latent diffusion transformer for video generation

    Ma, X., Wang, Y., Jia, G., Chen, X., Liu, Z., Li, Y.-F., Chen, C., and Qiao, Y. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024

    2024 arXiv

  22. [30]

    and Xie, S

    Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 4195--4205, 2023

    2023

  23. [31]

    Compression of convolutional neural networks: A short survey

    Pilipovi \'c , R., Buli \'c , P., and Risojevi \'c , V. Compression of convolutional neural networks: A short survey. In 2018 17th International Symposium INFOTEH-JAHORINA (INFOTEH), pp.\ 1--6. IEEE, 2018

    2018

  24. [32]

    Forward and backward information retention for accurate binary neural networks

    Qin, H., Gong, R., Liu, X., Shen, M., Wei, Z., Yu, F., and Song, J. Forward and backward information retention for accurate binary neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 2250--2259, 2020

    2020

  25. [33]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PMLR, 2021

    2021

  26. [34]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  27. [35]

    Post-training quantization on diffusion models

    Shang, Y., Yuan, Z., Xie, B., Wu, B., and Yan, Y. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1972--1981, 2023

    1972

  28. [36]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  29. [37]

    Ucf101: A dataset of 101 human actions classes from videos in the wild

    Soomro, K. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012

    2012 arXiv

  30. [38]

    and Deng, J

    Teed, Z. and Deng, J. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part II 16, pp.\ 402--419. Springer, 2020

    2020

  31. [39]

    Quest: Low-bit diffusion model quantization via efficient selective finetuning

    Wang, H., Shang, Y., Yuan, Z., Wu, J., and Yan, Y. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024 a

    2024 arXiv

  32. [40]

    C., and Loy, C

    Wang, J., Yue, Z., Zhou, S., Chan, K. C., and Loy, C. C. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pp.\ 1--21, 2024 b

    2024

  33. [41]

    Internvid: A large-scale video-text dataset for multimodal understanding and generation

    Wang, Y., He, Y., Li, Y., Li, K., Yu, J., Ma, X., Li, X., Chen, G., Chen, X., Wang, Y., et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023

    2023 arXiv

  34. [42]

    Exploring video quality assessment on user generated contents from aesthetic and technical perspectives

    Wu, H., Zhang, E., Liao, L., Chen, C., Hou, J., Wang, A., Sun, W., Yan, Q., and Lin, W. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 201...

    2023

  35. [43]

    Ptq4dit: Post-training quantization for diffusion transformers

    Wu, J., Wang, H., Shang, Y., Shah, M., and Yan, Y. Ptq4dit: Post-training quantization for diffusion transformers. arXiv preprint arXiv:2405.16005, 2024 a

    2024 arXiv

  36. [44]

    Seesr: Towards semantics-aware real-world image super-resolution

    Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., and Zhang, L. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 25456--25467, 2024 b

    2024

  37. [45]

    Smoothquant: Accurate and efficient post-training quantization for large language models

    Xiao, G., Lin, J., Seznec, M., Wu, H., Demouth, J., and Han, S. Smoothquant: Accurate and efficient post-training quantization for large language models. In International Conference on Machine Learning, pp.\ 38087--38099. PMLR, 2023

    2023

  38. [46]

    Cross-image relational knowledge distillation for semantic segmentation

    Yang, C., Zhou, H., An, Z., Jiang, X., Xu, Y., and Zhang, Q. Cross-image relational knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12319--12328, 2022

    2022

  39. [47]

    Multi-party collaborative attention control for image customization

    Yang, H., Yang, C., Wang, Q., An, Z., Feng, W., Huang, L., and Xu, Y. Multi-party collaborative attention control for image customization. arXiv preprint arXiv:2505.01428, 2025

    2025 arXiv

  40. [48]

    Vidit-q: Efficient and accurate quantization of diffusion transformers for image and video generation

    Zhao, T., Fang, T., Liu, E., Rui, W., Soedarmadji, W., Li, S., Lin, Z., Dai, G., Yan, S., Yang, H., et al. Vidit-q: Efficient and accurate quantization of diffusion transformers for image and video generation. arXiv preprint arXiv:2406.02540, 2024

    2024 arXiv

  41. [49]

    Mixdq: Memory-efficient few-step text-to-image diffusion models with metric-decoupled mixed precision quantization

    Zhao, T., Ning, X., Fang, T., Liu, E., Huang, G., Lin, Z., Yan, S., Dai, G., and Wang, Y. Mixdq: Memory-efficient few-step text-to-image diffusion models with metric-decoupled mixed precision quantization. In European Conference on Computer Vision, pp.\ 285--302. Springer, 2025

    2025

  42. [50]

    Bidm: Pushing the limit of quantization for diffusion models

    Zheng, X., Liu, X., Bian, Y., Ma, X., Zhang, Y., Wang, J., Guo, J., and Qin, H. Bidm: Pushing the limit of quantization for diffusion models. arXiv preprint arXiv:2412.05926, 2024 a

    2024 arXiv

  43. [51]

    Binarydm: Towards accurate binarization of diffusion model

    Zheng, X., Qin, H., Ma, X., Zhang, M., Hao, H., Wang, J., Zhao, Z., Guo, J., and Liu, X. Binarydm: Towards accurate binarization of diffusion model. arXiv preprint arXiv:2404.05662, 2024 b

    2024 arXiv

  44. [52]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.