Pith. sign in

REVIEW 4 major objections 5 minor 97 references

Parallel Decoding Distillation predicts multiple denoising steps in one network evaluation, reproducing teacher trajectories with 4-8 function evaluations for large-scale image, video, and audio generation.

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 00:52 UTC pith:DGC3BQOB

load-bearing objection A genuinely simpler parallel-decoding distillation idea with real results, but the theory is proved for the ideal objective while the headline experiments use data-free rollouts and approximate targets, and the 'SOTA' claim is stronger than the tables. the 4 major comments →

arxiv 2607.26004 v1 pith:DGC3BQOB submitted 2026-07-28 cs.CV cs.LG

Parallel Decoding Distillation for Fast Image and Video Generation

Neta Shaul , Chao Liu , Arash Vahdat , Julius Berner This is my paper
classification cs.CV cs.LG
keywords parallel decoding distillationtrajectory-based distillationflow matchingdiffusion modelsfew-step generationvideo generationtext-to-image generationgeneration diversity
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.

The paper sets out to show that a single network evaluation can replace many denoising steps in diffusion and flow models. Instead of distilling a one-step jump or aligning distributions, PDD trains a parallel decoder to predict the average velocity over each sub-interval of a block, then sums those velocities to advance the state across the whole block. The authors argue that this plain regression objective, which matches teacher Runge-Kutta steps, is sufficient to match or beat distribution-based distillation on large text-to-image, text-to-video, and text-to-audio models. If correct, it would make high-resolution video generation practical at 4-8 evaluations while avoiding the mode collapse and training instability associated with adversarial and score-distillation losses.

Core claim

The central claim is that a student model sharing the teacher's backbone, with the final linear layer repeated once per grid interval, can learn to output all mean velocities in a block from a single forward pass. The parallel-decoding loss regresses each sampled interval's output against a Runge-Kutta approximation of the teacher's mean velocity at the student's on-policy state. Proposition 1 states that a minimizer of this loss satisfies the parallel-decoder condition and, up to the Runge-Kutta approximation error, produces exactly the teacher's trajectory. On this basis, PDD reports state-of-the-art few-step quality on text-to-video, text-to-audio, and text-to-image models at 4-8 function

What carries the argument

The load-bearing object is the interval's mean velocity: the time-averaged teacher velocity over a discretization interval, which integrates the flow ODE exactly in one step. The parallel decoder is the same backbone as the teacher but with N final linear heads, one per grid interval, so one forward pass yields L velocities. The PD loss is a mean-squared regression of a randomly sampled head's output against a teacher mean-velocity target computed by an Euler or Midpoint step on the student's own parallelized trajectory. At generation, the block-step rule sums the heads, and layer fusion combines them into a single weighted-average linear layer, so advancing L intervals costs one evaluation

Load-bearing premise

The PD loss is assumed to be realizable with zero global minimum: the shared backbone, conditioned only on the block-start time, can represent all per-interval teacher mean velocities, and training on one randomly sampled interval per block transfers to all intervals at inference.

What would settle it

Train PDD on a simple curved flow with known closed-form trajectories and a fixed block of, say, four intervals; then compare the student's predicted mean velocity against the teacher's on the same states for each intra-block position, and compare a block-step rollout against the exact trajectory. If errors grow with the number of intervals skipped from the block start, or if a zero-loss fit on the training objective still fails on the last interval, the realizability and transfer assumption is false.

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

If this is right

  • High-resolution image, video, and audio generation becomes practical at 4-8 network evaluations instead of hundreds, while matching or beating teacher quality on standard benchmarks.
  • A single trained parallel decoder supports multiple inference budgets: varying the block size at generation time yields 2, 4, or 8 NFE from one checkpoint without retraining.
  • Trajectory regression is sufficient at scale: no variational-score, adversarial, or derivative-based losses are needed, removing a known source of mode collapse and training instability.
  • Because generation uses only the fused weighted-average head, the enlarged prediction head is essentially free at inference time.
  • PDD's loss needs only one or two teacher evaluations per training step, making distillation tractable on large models without expensive Jacobian-vector products or finite differences.

Where Pith is reading between the lines

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

  • If the block-parallel idea transfers to autoregressive models, a block of tokens could be predicted per forward pass, using a discrete analogue of the mean velocity as the target; the paper notes this as future work but does not develop it.
  • The data-free training variant suggests that a teacher checkpoint plus a prompt set is enough to distill a model even when the original training data is unavailable, widening the practical use of distillation.
  • An adaptive verifier choosing block size per sample would let compute track local trajectory difficulty; the paper flags this possibility but does not implement it, leaving a testable extension.
  • Because guidance is folded into the teacher's mean-velocity target, PDD may serve as a general acceleration wrapper for guided ODE sampling, not only for generative-model distillation.

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

4 major / 5 minor

Summary. The paper introduces Parallel Decoding Distillation (PDD), a trajectory-based distillation method for diffusion and flow models. PDD augments a pretrained teacher's backbone with N parallel linear heads so that, for a block of L time intervals, a single forward pass predicts all per-interval mean velocities. Training is a simple regression objective (the PD loss, Eq. (11)) against one-step Euler or Midpoint approximations of the teacher's mean velocity, computed on student rollouts. Inference advances the state by a fused linear layer (Eqs. (14)-(15)), so L steps are taken per network evaluation, giving variable NFE by varying L. Experiments cover ImageNet-256 (SiT-XL+REPA), Qwen-Image 20B, Wan2.1 1.3B/14B, and LTX-2.3 22B text-to-video/audio. The central claims are that PDD achieves state-of-the-art few-step generation at 4-8 NFE on Qwen-Image and Wan2.1, matches or outperforms the official distilled LTX-2.3 model, preserves diversity better than distribution-based baselines, and that its PD-loss minimizer reproduces teacher trajectories (Proposition 1).

Significance. If the empirical claims hold, PDD is a genuinely useful contribution: it shows that a pure trajectory-regression objective, without VSD, GAN losses, JVPs, or finite differences, can be scaled to 14B-22B video/audio models and produce competitive few-step generation with variable NFE. The architecture is simple—N repeated final linear layers on a frozen-pretrained backbone—and the fused-layer identity in Eqs. (14)-(15) is a clean and practical trick for avoiding extra inference cost. The paper also provides reproducible pseudocode for training and inference, and it makes a concrete falsifiable prediction: the PD-loss minimizer matches teacher trajectories. The diversity-focused evaluation (pairwise V-JEPA 2 / VideoMAE V2 distances) is a useful counterpoint to the known mode-collapse of distribution-based methods. However, the central theoretical guarantee, as stated, applies only to an idealized objective and does not cover the actual training algorithm used for the headline results; the empirical SOTA claim is also somewhat stronger than the tables support. With those gaps addressed, the contribution would be solid and significant for the fast-generation community.

major comments (4)
  1. [Section 3, Eq. (11), Appendix D] Proposition 1's proof assumes (i) the PD objective is realizable with global minimum zero, (ii) X_n is distributed as p_tn via the interpolant, and (iii) the teacher mean velocity u_k is exact. None of these hold for the algorithm that produced the main results. In Algorithms 2 and 3 the target is a one-step Euler or Midpoint approximation of u_k, so a zero-loss student would reproduce the numerical solver's trajectory, not the teacher ODE; the phrase 'up to a controllable Runge-Kutta approximation error' is not quantified. The proposition needs to be stated for the approximate objective, and an error-accumulation bound over N/L block steps should be provided, or the paper should clearly separate the idealized theoretical statement from a separate, empirically justified claim for the RK-approximated loss.
  2. [Section 3, 'Data-free training', Algorithm 3] All large-scale T2I/T2V experiments use data-free training (Algorithm 3), where the block starting state X_n is not sampled from p_tn but is generated by the student's own block-step rule from previous iterations. Proposition 1's induction starts from X_n ~ p_tn; for student-rollout states the induction breaks, and the paper gives no fixed-point or convergence analysis for the data-free loop. The conclusion even states that data-free settings beyond ImageNet-256 remain future work, but the abstract's SOTA results rely entirely on this unanalyzed data-free regime. This is a load-bearing gap: it should be either formally addressed or the theoretical claims should be explicitly restricted to the data-dependent setting, with empirical evidence carrying the data-free results.
  3. [Tables 4 and 5 vs. Abstract] The abstract and introduction state 'SOTA performance' and 'significant improvement in generated video diversity.' The tables are more nuanced: on Wan2.1 14B, AnyFlow has higher VBench Overall than PDD-short at both 4 NFE (84.95 vs. 84.92) and 8 NFE (85.08 vs. 84.96); on HPSv2 and PickScore, DMD2 remains ahead at 4 and 8 NFE (e.g., HPSv2 32.34 vs. 31.33 at 4 NFE). The diversity comparisons are also reported on pairwise feature distances without confidence intervals or seed variability, and the 'significant improvement' is primarily qualitative. Please temper the claims to match the evidence, e.g., 'competitive or state-of-the-art on specific benchmarks' and 'improved diversity relative to distribution-based baselines in our measurements.'
  4. [Section 3, Eq. (13), Appendix B.3, Figure 17] The architecture assumes that a single shared backbone feature H_tn(X_n), linearly decoded by N heads, can represent the teacher per-interval mean velocities u_k(X_k) for k > n. This is a capacity/realizability assumption that is never demonstrated. Figure 17 reports average intra-block curvature over 10 trajectories, which is a weak proxy: it does not compare the student's per-interval predictions against the teacher's u_k at the corresponding states, nor does it measure representational error. Some statement about capacity (e.g., an ablation of the number of heads or a per-interval prediction-error plot) is needed to support the architecture choice, particularly for long blocks.
minor comments (5)
  1. [Appendix B.2, checkpoint selection] The Qwen-Image models are evaluated at the iteration that achieves the best average on the three benchmarks (1250 for Euler, 2250 for Midpoint). This is effectively checkpoint selection on the test set; it should be disclosed more prominently, and ideally the variance across nearby checkpoints should be reported.
  2. [Table 5 and Figure 22] The LTX-2.3 comparison in Figure 22 is based on a Gemini 3.1 Pro Preview judge on 100 prompts, with PDD winning 142, tying 35, and losing 123 of 300 paired comparisons (mean score 2.62 vs. 2.59). This is a very small aggregate margin; the text should not describe this as clearly 'on par or better' without also reporting the per-axis margins and the judge's reliability.
  3. [Throughout] Several tables have formatting artifacts ('T able 3', 'T able 5'), and the author line contains a stray space in 'Arash V ahdat'. Please proofread.
  4. [Section 5, Tables 2-5] All benchmark numbers are reported without confidence intervals or repeated-run variability. Given that differences between top methods are often small (e.g., 84.92 vs. 84.95 on VBench), error bars or at least a statement of evaluation budget would help assess whether the differences are meaningful.
  5. [Figure 17] The figure measures 'average curvature' but the definition of curvature is not given. Please specify how curvature is computed, over what time interval, and why intra-block curvature is the relevant quantity for validating the representational capacity of the parallel decoder.

Circularity Check

0 steps flagged

No circularity: the PD loss is a standard student-teacher regression onto frozen teacher targets; Proposition 1 is a zero-loss characterization, not a tautology.

full rationale

The core derivation chain is a conventional student-teacher distillation. The student map in Eq. (13) learns N linear heads over a shared backbone, and the PD loss (11) regresses those predictions onto the frozen pretrained teacher's Runge-Kutta approximation of the mean velocity. The target u_k(sg(\bar X_k)) depends on student states only through a stop-gradient on-policy rollout, and the teacher itself is never trained or fit to the student. Proposition 1 (Appendix D) is not circular: it assumes a global minimizer with zero loss and proves by induction on the block that the student state \bar X_k equals the teacher state X_k, converting the zero-loss equality (22) into the parallel-decoder condition (8). This is a mathematical characterization of a zero-loss fixed point, not an identity imposed by definition. The fused-layer identity (14)-(15) is an algebraic reparameterization of the block update and introduces no circularity. Self-citations in the paper (FastGen [38] as a baseline implementation, Bespoke non-stationary solvers [48] for the shift time transformation) are background or evaluation tools and are not used to define the PDD objective or to prove the central claim. The principal weakness is a correctness gap rather than circularity: the large-scale text-to-image and video experiments use the data-free Algorithm 3, which advances states by the student's own block-step rule instead of sampling X_n ~ p_{t_n} as assumed in Proposition 1, and the practical targets are Euler/Midpoint approximations rather than exact mean velocities. These are unproven assumptions about distribution shift and error accumulation, but they do not amount to the derivation reducing to its own inputs. Because the method is a regression onto externally computed, frozen-teacher targets and is benchmarked against independent checkpoints and metrics, the paper is self-contained with respect to circularity and receives score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

PDD introduces an architectural pattern (parallel linear heads, fused linear layer) but no new physical or mathematical entity. The method's load-bearing assumptions are optimization/capacity and teacher-fidelity assumptions rather than invented postulates. The free parameters listed are task-specific training and evaluation choices, several of which are tuned or selected post hoc.

free parameters (6)
  • Grid size N = 64/128 for ImageNet; 128/256 for text-to-image/video; 256 for LTX Euler
    Determines time discretization and the number of parallel linear heads; chosen per task, not derived.
  • Block-size range L_min..L_max = Chosen so inference NFE is 1/2/4/8 for ImageNet and 2/4/8 or 4/8 for others
    Controls variable NFE support; a design choice tuned to target NFE sets.
  • Shift scale s in time reparameterization (16) = 5 for Qwen, 6 for Wan, 10 for LTX
    Time-grid transformation tuned per model; no principled prediction.
  • CFG guidance scale w = 2.7/2.9/3.2 for ImageNet, 4 for Qwen, 5 for Wan
    Guidance scale used in teacher target (17); tuned per model and reported for best results.
  • CFG unconditional layer skip = layer 10 for Wan 1.3B, layer 12 for Wan 14B
    Skipping a specific backbone layer for the unconditional term improves performance; an ad hoc per-model choice.
  • Checkpoint selection iteration = Qwen: 1250 (Euler), 2250 (Midpoint); Wan 1.3B: 25/225; Wan 14B: 200/3000; LTX: 250
    Checkpoints selected by best evaluated benchmark scores during training; post hoc selection inflates reported numbers.
axioms (5)
  • domain assumption The generator is a deterministic flow with well-defined velocity field v_t and source distribution p_0 (Eq. 1), and pretrained models use the linear interpolant (Eq. 2).
    Needed to define mean velocity (4) and sample training states X_n via the interpolant.
  • domain assumption The pretrained teacher flow model is frozen and treated as the ground-truth generator to match.
    The entire distillation objective (11) is defined by teacher trajectories; no teacher error is modeled.
  • ad hoc to paper The PD objective is realizable and its global minimum is zero; the teacher mean velocity is approximated with one Euler or Midpoint step.
    Proposition 1 assumes a zero global minimum; in practice targets are Runge-Kutta approximations, so exact teacher-trajectory recovery is not guaranteed.
  • ad hoc to paper The student backbone, with block-start time conditioning and N linear heads, has enough capacity to represent all per-interval mean velocities u_k(X_k) from X_n alone.
    Central to Eq. (8) and architecture (13); supported empirically by curvature comparison (Fig. 17) but not proven.
  • domain assumption Data-free on-policy training (Algorithm 3), with stop-gradient state advancement, provides a training distribution that covers the generation-time distribution.
    Standard on-policy assumption; no convergence or coverage proof is given.

pith-pipeline@v1.3.0-alltime-deepseek · 30658 in / 11732 out tokens · 114217 ms · 2026-08-01T00:52:07.327603+00:00 · methodology

0 comments
read the original abstract

Generation in video diffusion or flow models is computationally expensive due to the slow and iterative sampling process. Current state-of-the-art (SOTA) acceleration methods heavily rely on variational score distillation (VSD) and adversarial losses to distill diffusion models into few-step generators. Albeit achieving high-quality video generation, these training losses are notoriously hard to optimize and suffer from mode collapse, leading to loss of video diversity and lack of motion. In this paper, we introduce Parallel Decoding Distillation (PDD), a simplified and scalable trajectory-based distillation method for fast inference of diffusion and flow matching models. Our architecture and training procedure are compatible with any pre-trained model and support sampling with a varying number of function evaluations (NFE). PDD accelerates generation by predicting multiple denoising steps per network evaluation. Conceptually, it learns a representation of the mean velocity without regressing its derivative using JVPs or finite-difference approximations. Our method achieves SOTA performance with 4-8 NFE on LTX-2.3 Text-to-Video/Audio, Wan 14B Text-to-Video, and Qwen-Image Text-to-Image. Moreover, PDD presents a significant improvement in generated video diversity.

Figures

Figures reproduced from arXiv: 2607.26004 by Arash Vahdat, Chao Liu, Julius Berner, Neta Shaul.

Figure 1
Figure 1. Figure 1: (top) 10s 720p videos with audio from LTX-2.3 teacher (4 × 30 NFE), Parallel Decoding Distillation (PDD) (8 NFE), and the official distilled model (8 NFE). PDD can follow the teacher more closely. (bottom) PDD vs. DMD2 and AnyFlow on the Wan2.1 Text-to-Video 14B model. All methods use 4 NFE. Notably, PDD demonstrates high video quality while preserving better generation diversity across distinct initial no… view at source ↗
Figure 2
Figure 2. Figure 2: The sampling trajectory is discretized into 𝑁 intervals, which are grouped into blocks of size 𝐿. The parallel decoder predicts the mean velocities for all intervals within a block using a single evaluation. number of function evaluations (NFEs) during infer￾ence. PDD is conceptually related to Pi-Flow [7], but provides a simplified training algorithm that removes the need for an additional policy head and… view at source ↗
Figure 3
Figure 3. Figure 3: (left) The PDD student approximates the mean velocity across multiple consecutive intervals in a single evaluation. The pre-trained flow model (teacher) provides the mean velocity of a single interval using an ODE solver step. (right) Illustrate estimation of the PD loss given an initial state 𝑋𝑛 ∼ 𝑝𝑡𝑛 and a block size 𝐿 = 4; i) student predicts the mean velocities 𝑢¯ 𝜃 𝑛 (·|𝑋𝑛) ∈ 𝒳 𝐿 ; ii) Following the s… view at source ↗
Figure 4
Figure 4. Figure 4: Architecture of the parallel decoder (b) vs. the pre-trained flow model (a). Notably, the parallel decoder utilizes the same backbone, but with 𝑁 times the final linear layer. (c) At generation, instead of applying a linear layer per intra-block step (9), we can fuse the layers into a single linear layer that outputs the block’s average velocity, which advances the state across the entire block (14). Impor… view at source ↗
Figure 5
Figure 5. Figure 5: PDD - Midpoint (ours) vs. DMD2, and AnyFlow on the Wan Text-to-Video 14B model. The three methods use 4 NFE. Notably, PDD achieves high-quality video while preserving better diversity. While PDD could potentially be extended to a self-distillation framework, such methods require a completely different pipeline, including larger datasets and significantly more compute. Distribution-based distillation Instea… view at source ↗
Figure 6
Figure 6. Figure 6: FID vs. NFE of PDD with Euler and Midpoint methods for approximating the mean velocity (4) on ImageNet-256 using SiT-XL+REPA model as teacher with guidance scale 2.9. models, each with a different grid size, Runge-Kutta method, and block-size range. Additional training details are provided in Appendix B. For class-conditional image generation, we set 𝑁 = 128 for the Euler method and 𝑁 = 64 for the midpoint… view at source ↗
Figure 7
Figure 7. Figure 7: LTX-2.3 teacher with 4 × 30 NFE compared to PDD (ours) and the official distilled LTX-2.3 model, both using 8 NFE. within each block. For the teacher, this leads to 4×30 NFE per generation. The comparisons in [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FID vs. NFE of PDD with Euler and Midpoint methods for approximating the mean velocity (4) on Repa-ImageNet-256. FID vs. Training iteration We evaluate FID every 10K training iteration with NFE= 1, 2, 4, 8 and report the results in [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FID vs. Training iteration of PDD with Euler and Midpoint methods for approximating the mean velocity (4) on Repa-ImageNet-256. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Overall metrics of OneIG, DPG-Bench, and GenEval vs. Training iteration of the Qwen-Image PDD model. Additional results We provide all dimensions of our considered benchmarks in tables 6, 7, and 8. Moreover, we provide additional comparisons between the Qwen-Image teacher, PDD, and DMD2 (Lightning-v2) in Figures 11 to 16. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Teacher uses Euler method with 2×50 NFE, PDD and DMD2 (Lightning-v2) use 4 NFE. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Teacher uses Euler method with 2×50 NFE, PDD and DMD2 (Lightning-v2) use 4 NFE. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Teacher uses Euler method with 2×50 NFE, PDD and DMD2 (Lightning-v2) use 4 NFE. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Teacher uses Euler method with 2×50 NFE, PDD and DMD2 (Lightning-v2) use 4 NFE. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Teacher uses Euler method with 2×50 NFE, PDD and DMD2 (Lightning-v2) use 4 NFE. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Teacher uses Euler method with 2×50 NFE, PDD and DMD2 (Lightning-v2) use 4 NFE. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Curvature of PDD vs. Teacher on the Wan2.1 14B model. We report the averaged curvature over 10 trajectories. Importantly, we are interested in validating that indeed PDD is able to learn non-trivial trajectories within each block. Thus, for PDD we report only intra-block curvature. Subject Consistency Background Consistency Temporal Flickering Motion Smoothness Dynamic Degree Aesthetic Quality Imaging Qua… view at source ↗
Figure 18
Figure 18. Figure 18: Full VBench dimensions on the Wan2.1 1.3B and 14B models. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Wan2.1 14B: PDD - Midpoint (ours) vs. DMD2 (FastGen) and AnyFlow with 4 NFE. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Wan2.1 14B: PDD - Midpoint (ours) vs. DMD2 (FastGen) and AnyFlow with 4 NFE. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Wan2.1 14B: PDD - Midpoint (ours) vs. DMD2 (FastGen) and AnyFlow with 4 NFE. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Per-axis judge preference between PDD and the official distilled LTX-2.3 model, both with 8 NFE, as scored by Gemini 3.1 Pro Preview. Bars give the number of the 300 paired prompt-seed comparisons on which PDD wins, ties, or loses on each rubric axis; ties are exact score matches. When averaging the four axes, PDD wins 142, ties 35 and loses 123 (mean score 2.62 against 2.59). 30 [PITH_FULL_IMAGE:figures… view at source ↗
Figure 23
Figure 23. Figure 23: LTX-2.3 teacher (4×30 NFE) vs. PDD (8 NFE) vs. official distilled model (8 NFE). 31 [PITH_FULL_IMAGE:figures/full_fig_p031_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: LTX-2.3 teacher (4×30 NFE) vs. PDD (8 NFE) vs. official distilled model (8 NFE). 32 [PITH_FULL_IMAGE:figures/full_fig_p032_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: LTX-2.3 teacher (4×30 NFE) vs. PDD (8 NFE) vs. official distilled model (8 NFE). 33 [PITH_FULL_IMAGE:figures/full_fig_p033_25.png] 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

97 extracted references · 39 linked inside Pith

  1. [1]

    Albergo, Nicholas M

    Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions,

  2. [2]

    V-jepa 2: Self-supervised video models enable un- derstanding, prediction and planning, 2025

    MidoAssran, AdrienBardes, DavidFan, Quentin Garrido, Russell Howes, Mojtaba, Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, Sergio Arnaud, Abha Gejji, Ada Martin, Francois Robert Hogan, Daniel Dugas, Piotr Bojanowski, Vasil Khali- dov, Patrick Labatut, Francisco Massa, Marc Szafraniec, Kapil Krishnakumar, Yong Li, Xi- ...

    Pith/arXiv arXiv 2025

  3. [3]

    Boffi, Michael S

    Nicholas M. Boffi, Michael S. Albergo, and Eric Vanden-Eijnden. Flow map matching with stochastic interpolants: A mathematical frame- work for consistency models, 2025. URLhttps: //arxiv.org/abs/2406.07507

    Pith/arXiv arXiv 2025

  4. [4]

    Boffi, Michael S

    Nicholas M. Boffi, Michael S. Albergo, and Eric Vanden-Eijnden. How to build a consistency model: Learning flow maps via self-distillation,

  5. [5]

    Shortcutting pre-trained flow matching diffusion models is almost free lunch, 2025

    Xu Cai, Yang Wu, Qianli Chen, Haoran Wu, Lichuan Xiang, and Hongkai Wen. Shortcutting pre-trained flow matching diffusion models is almost free lunch, 2025. URLhttps://arxiv. org/abs/2510.17858

    arXiv 2025

  6. [6]

    URL https://arxiv.org/abs/2505. 18825

  7. [7]

    pi-flow: Policy-based few-step generation via imitation distillation, 2025

    Hansheng Chen, Kai Zhang, Hao Tan, Leonidas Guibas, Gordon Wetzstein, and Sai Bi. pi-flow: Policy-based few-step generation via imitation distillation, 2025. URLhttps://arxiv.org/ abs/2510.14974

    arXiv 2025

  8. [8]

    Oneig-bench: Omni- dimensional nuanced evaluation for image gener- ation, 2025

    Jingjing Chang, Yixiao Fang, Peng Xing, Shuhan Wu, Wei Cheng, Rui Wang, Xianfang Zeng, Gang Yu, and Hai-Bao Chen. Oneig-bench: Omni- dimensional nuanced evaluation for image gener- ation, 2025. URLhttps://arxiv.org/abs/ 2506.07977

    Pith/arXiv arXiv 2025

  9. [9]

    Lightx2v: Light video generation inference framework

    LightX2V Contributors. Lightx2v: Light video generation inference framework. https:// github.com/ModelTC/lightx2v, 2025

    2025

  10. [10]

    Twinflow: Realizing one-step gener- ation on large models with self-adversarial flows,

    Zhenglin Cheng, Peng Sun, Jianguo Li, and Tao Lin. Twinflow: Realizing one-step gener- ation on large models with self-adversarial flows,

  11. [11]

    Phased dmd: Few-step distribution matching distil- lation via score matching within subintervals,

    Xiangyu Fan, Zesong Qiu, Zhuguanyu Wu, Fanzhou Wang, Zhiqian Lin, Tianxiang Ren, Dahua Lin, Ruihao Gong, and Lei Yang. Phased dmd: Few-step distribution matching distil- lation via score matching within subintervals,

  12. [12]

    One step diffusion via short- cut models, 2025

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via short- cut models, 2025. URLhttps://arxiv.org/ abs/2410.12557

    Pith/arXiv arXiv 2025

  13. [13]

    Scaling rectified flow transformers for high-resolution image synthe- sis, 2024

    Patrick Esser, SumithKulal, AndreasBlattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boe- sel, Dustin Podell, Tim Dockhorn, Zion En- glish, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthe- sis, 2024. URL https://arxiv.org/abs/...

    Pith/arXiv arXiv 2024

  14. [14]

    Zico Kolter, and Kaiming He

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J. Zico Kolter, and Kaiming He. Mean flows for one-step generative modeling, 2025. URL https://arxiv.org/abs/2505.13447

    Pith/arXiv arXiv 2025

  15. [15]

    URL https://arxiv.org/abs/2510. 27684

  16. [16]

    Geneval: An object-focused frame- work for evaluating text-to-image alignment,

    Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused frame- work for evaluating text-to-image alignment,

  17. [17]

    Zico Kolter

    Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J. Zico Kolter. Consistency mod- els made easy, 2024. URL https://arxiv. org/abs/2406.14548

    Pith/arXiv arXiv 2024

  18. [18]

    Splitmeanflow: Interval split- ting consistency in few-step generative modeling, 11 Parallel Decoding Distillation for Image and Video Generation

    Yi Guo, Wei Wang, Zhihang Yuan, Rong Cao, Kuan Chen, Zhengyang Chen, Yuanyuan Huo, Yang Zhang, Yuping Wang, Shouda Liu, and Yuxuan Wang. Splitmeanflow: Interval split- ting consistency in few-step generative modeling, 11 Parallel Decoding Distillation for Image and Video Generation

  19. [19]

    Zico Kolter, and Kaiming He

    Zhengyang Geng, Yiyang Lu, Zongze Wu, Eli Shechtman, J. Zico Kolter, and Kaiming He. Im- proved mean flows: On the challenges of fast- forward generative models, 2025. URLhttps: //arxiv.org/abs/2512.02012

    Pith/arXiv arXiv 2025

  20. [20]

    Classifier- free diffusion guidance, 2022

    Jonathan Ho and Tim Salimans. Classifier- free diffusion guidance, 2022. URL https: //arxiv.org/abs/2207.12598

    Pith/arXiv arXiv 2022

  21. [21]

    De- noising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. De- noising diffusion probabilistic models, 2020. URL https://arxiv.org/abs/2006.11239

    Pith/arXiv arXiv 2020

  22. [22]

    Anyflow: Any-step video diffusion model with on- policy flow map distillation, 2026

    Yuchao Gu, Guian Fang, Yuxin Jiang, Weijia Mao, Song Han, Han Cai, and Mike Zheng Shou. Anyflow: Any-step video diffusion model with on- policy flow map distillation, 2026. URLhttps: //arxiv.org/abs/2605.13724

    Pith/arXiv arXiv 2026

  23. [23]

    Cmt: Mid-training for ef- ficient learning of consistency, mean flow, and flow map models, 2026

    Zheyuan Hu, Chieh-Hsin Lai, Yuki Mitsufuji, and Stefano Ermon. Cmt: Mid-training for ef- ficient learning of consistency, mean flow, and flow map models, 2026. URLhttps://arxiv. org/abs/2509.24526

    arXiv 2026

  24. [24]

    URL https://arxiv.org/abs/2507. 16884

  25. [25]

    Ltx-2: Efficient joint audio-visual foundation model,

    YoavHaCohen, BennyBrazowski, NisanChiprut, Yaki Bitterman, Andrew Kvochko, Avishai Berkowitz, Daniel Shalem, Daphna Lifschitz, Dudu Moshe, Eitan Porat, Eitan Richardson, Guy Shiran, Itay Chachy, Jonathan Chetboun, Michael Finkelson, Michael Kupchick, Nir Zabari, Nitzan Guetta, Noa Kotler, Ofir Bibi, Ori Gor- don, Poriya Panet, Roi Benita, Shahar Armon, Vi...

  26. [26]

    URL https://arxiv.org/abs/2601. 03233

  27. [27]

    Decoupled meanflow: Turning flow models into flow maps for accelerated sampling, 2025

    Kyungmin Lee, Sihyun Yu, and Jinwoo Shin. Decoupled meanflow: Turning flow models into flow maps for accelerated sampling, 2025. URL https://arxiv.org/abs/2510.24474

    arXiv 2025

  28. [28]

    Diffusion adversar- ial post-training for one-step video generation,

    ShanchuanLin, XinXia, YuxiRen, CeyuanYang, Xuefeng Xiao, and Lu Jiang. Diffusion adversar- ial post-training for one-step video generation,

  29. [29]

    Ella: Equip diffusion modelswithllmforenhancedsemanticalignment,

    Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion modelswithllmforenhancedsemanticalignment,

  30. [30]

    Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. URL https://arxiv.org/abs/2209.03003

    Pith/arXiv arXiv 2022

  31. [31]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. URLhttps: //arxiv.org/abs/1711.05101

    Pith/arXiv arXiv 2019

  32. [32]

    Vbench: Comprehensive benchmark suite for video generative models, 2023

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chan- paisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench: Comprehensive benchmark suite for video generative models, 2023. URL https: //arxiv.org/abs/2311.17982

    Pith/arXiv arXiv 2023

  33. [33]

    Hunyuanvideo: A systematic framework for large video generative models, 2025

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, An- dong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Ji- awangBai, JianbingWu, JinbaoXue, JoeyWang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, ...

    Pith/arXiv arXiv 2025

  34. [34]

    FLUX.2: Frontier Visual In- telligence

    Black Forest Labs. FLUX.2: Frontier Visual In- telligence. https://bfl.ai/blog/flux-2, 2025

    2025

  35. [35]

    Soflow: Solution flow models for one-step gen- erative modeling, 2026

    Tianze Luo, Haotian Yuan, and Zhuang Liu. Soflow: Solution flow models for one-step gen- erative modeling, 2026. URLhttps://arxiv. org/abs/2512.15657

    arXiv 2026

  36. [36]

    Zico Kolter, and Guo jun Qi

    Weijian Luo, Zemin Huang, Zhengyang Geng, J. Zico Kolter, and Guo jun Qi. One-step diffu- sion distillation through score implicit matching, 12 Parallel Decoding Distillation for Image and Video Generation

  37. [37]

    URL https://arxiv.org/abs/2501. 08316

  38. [38]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben- Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling, 2023. URL https://arxiv.org/abs/2210.02747

    Pith/arXiv arXiv 2023

  39. [39]

    Transition matching distillation for fast video generation,

    Weili Nie, Julius Berner, Nanye Ma, Chao Liu, Saining Xie, and Arash Vahdat. Transition matching distillation for fast video generation,

  40. [40]

    Cosmos 3: Omnimodal world models for physical ai, 2026

    NVIDIA. Cosmos 3: Omnimodal world models for physical ai, 2026. URL https://arxiv. org/abs/2606.02800

    Pith/arXiv arXiv 2026

  41. [41]

    Simplifying, stabiliz- ing and scaling continuous-time consistency mod- els, 2025

    Cheng Lu and Yang Song. Simplifying, stabiliz- ing and scaling continuous-time consistency mod- els, 2025. URL https://arxiv.org/abs/ 2410.11081

    Pith/arXiv arXiv 2025

  42. [42]

    One-step latent-free image generation with pixel mean flows, 2026

    Yiyang Lu, Susie Lu, Qiao Sun, Hanhong Zhao, Zhicheng Jiang, Xianbang Wang, Tianhong Li, Zhengyang Geng, and Kaiming He. One-step latent-free image generation with pixel mean flows, 2026. URLhttps://arxiv.org/abs/ 2601.22158

    Pith/arXiv arXiv 2026

  43. [43]

    Latent consistency models: Synthesizing high-resolution images with few- step inference, 2023

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few- step inference, 2023. URL https://arxiv. org/abs/2310.04378

    Pith/arXiv arXiv 2023

  44. [44]

    Align your flow: Scaling continuous-time flow map distillation, 2025

    Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your flow: Scaling continuous-time flow map distillation, 2025. URL https:// arxiv.org/abs/2506.14603

    Pith/arXiv arXiv 2025

  45. [45]

    Progressive distillation for fast sampling of diffusion models,

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models,

  46. [46]

    URL https://arxiv.org/abs/2410. 16794

  47. [47]

    Im- proved training technique for shortcut models,

    Anh Nguyen, Viet Nguyen, Duc Vu, Trung Dao, Chi Tran, Toan Tran, and Anh Tran. Im- proved training technique for shortcut models,

  48. [48]

    URL https://arxiv.org/abs/2510. 21250

  49. [49]

    Nvidia fastgen: Fast generation from diffusion models, 2026

    Weili Nie, Julius Berner, Chao Liu, and Arash Vahdat. Nvidia fastgen: Fast generation from diffusion models, 2026. URLhttps://github. com/NVlabs/FastGen

    2026

  50. [50]

    Improved techniques for training consistency models,

    Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models,

  51. [51]

    URL https://arxiv.org/abs/2601. 09881

  52. [52]

    Consistency models, 2023

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models, 2023. URL https://arxiv.org/abs/2303.01469

    Pith/arXiv arXiv 2023

  53. [53]

    Eflow: Fast few-step video generator training from scratch via efficient solution flow,

    Dogyun Park, Yanyu Li, Sergey Tulyakov, and Anil Kag. Eflow: Fast few-step video generator training from scratch via efficient solution flow,

  54. [54]

    URL https://arxiv.org/abs/2603. 27086

  55. [55]

    High- resolution image synthesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models, 2022. URL https://arxiv.org/ abs/2112.10752

    Pith/arXiv arXiv 2022

  56. [56]

    ImageNet Large Scale Visual Recogni- tion Challenge.International Journal of Com- puter Vision (IJCV), 115(3):211–252, 2015

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhi- heng Huang, Andrej Karpathy, Aditya Khosla, MichaelBernstein, AlexanderC.Berg, andLiFei- Fei. ImageNet Large Scale Visual Recogni- tion Challenge.International Journal of Com- puter Vision (IJCV), 115(3):211–252, 2015. doi: 10.1007/s11263-015-0816-y

    doi:10.1007/s11263-015-0816-y 2015

  57. [57]

    Videomae v2: Scaling video masked autoencoders with dual masking, 2023

    Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking, 2023. URL https://arxiv.org/abs/2303.16727

    Pith/arXiv arXiv 2023

  58. [58]

    VidProM: A million- scale real prompt-gallery dataset for text-to- video diffusion models, 2024

    Wenhao Wang and Yi Yang. VidProM: A million- scale real prompt-gallery dataset for text-to- video diffusion models, 2024. URL https: //arxiv.org/abs/2403.06098

    Pith/arXiv arXiv 2024

  59. [59]

    Pro- lificdreamer: High-fidelity and diverse text-to- 3d generation with variational score distillation,

    Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Pro- lificdreamer: High-fidelity and diverse text-to- 3d generation with variational score distillation,

  60. [60]

    Adver- sarial diffusion distillation, 2023

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adver- sarial diffusion distillation, 2023. URL https://arxiv.org/abs/2311.17042

    Pith/arXiv arXiv 2023

  61. [61]

    Fast high-resolution image synthe- sis with latent adversarial diffusion distillation,

    Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthe- sis with latent adversarial diffusion distillation,

  62. [62]

    URL https://arxiv.org/abs/2403. 12015

  63. [63]

    Neta Shaul, Uriel Singer, Ricky T. Q. Chen, Matthew Le, Ali Thabet, Albert Pumarola, and Yaron Lipman. Bespoke non-stationary solvers for fast sampling of diffusion and flow models,

  64. [64]

    URL https://arxiv.org/abs/2403. 01329

  65. [65]

    Weiss, Niru Mah- eswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Mah- eswaranathan, and Surya Ganguli. Deep unsu- pervised learning using nonequilibrium thermo- dynamics, 2015. URL https://arxiv.org/ abs/1503.03585

    Pith/arXiv arXiv 2015

  66. [66]

    Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis,

  67. [67]

    URL https://arxiv.org/abs/2310. 14189

  68. [68]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative mod- eling through stochastic differential equations,

  69. [69]

    Alphaflow: Understanding and improving meanflow models,

    Huijie Zhang, Aliaksandr Siarohin, Willi Mena- pace, Michael Vasilkovsky, Sergey Tulyakov, Qing Qu, and Ivan Skorokhodov. Alphaflow: Understanding and improving meanflow models,

  70. [70]

    Large scale diffusion distillation via score-regularized continuous-time consistency,

    Kaiwen Zheng, Yuji Wang, Qianli Ma, Huayu Chen, Jintao Zhang, Yogesh Balaji, Jianfei Chen, Ming-Yu Liu, Jun Zhu, and Qinsheng Zhang. Large scale diffusion distillation via score-regularized continuous-time consistency,

  71. [71]

    Any-step genera- tion via n-th order recursive consistent veloc- ity field estimation

    Peng Sun and Tao Lin. Any-step genera- tion via n-th order recursive consistent veloc- ity field estimation. InThe Fourteenth In- ternational Conference on Learning Represen- tations, 2026. URL https://openreview. net/forum?id=GnawtLKGkP

    2026

  72. [72]

    Joshua Tian Jin Tee, Kang Zhang, Hee Suk Yoon, Dhananjaya Nagaraja Gowda, Chanwoo Kim, and Chang D. Yoo. Physics informed distillation for diffusion models, 2024. URL https://arxiv.org/abs/2411.08378

    Pith/arXiv arXiv 2024

  73. [73]

    Flow map distillation without data, 2025

    Shangyuan Tong, Nanye Ma, Saining Xie, and Tommi Jaakkola. Flow map distillation without data, 2025. URL https://arxiv.org/abs/ 2511.19428

    arXiv 2025

  74. [74]

    Wan: Open and advanced large-scale video generative models, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, JiayuWang, JingfengZhang, JingrenZhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, 13 Parallel Decoding Distillation for Image and Video Generation Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, ...

    Pith/arXiv arXiv 2025

  75. [78]

    URL https://arxiv.org/abs/2305. 16213

  76. [79]

    Transition models: Rethinking the gen- erative learning objective, 2025

    Zidong Wang, Yiyuan Zhang, Xiaoyu Yue, Xi- angyu Yue, Yangguang Li, Wanli Ouyang, and Lei Bai. Transition models: Rethinking the gen- erative learning objective, 2025. URL https: //arxiv.org/abs/2509.04394

    Pith/arXiv arXiv 2025

  77. [80]

    Qwen- image technical report, 2025

    Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingkun...

    Pith/arXiv arXiv 2025

  78. [81]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yix- iong Chen, Feng Zhu, Rui Zhao, and Hong- sheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023. URLhttps: //arxiv.org/abs/2306.09341

    Pith/arXiv arXiv 2023

  79. [82]

    Spectral progressive diffusion for efficient image and video generation, 2026

    Howard Xiao, Brian Chao, Lior Yariv, and Gor- don Wetzstein. Spectral progressive diffusion for efficient image and video generation, 2026. URL https://arxiv.org/abs/2605.18736

    Pith/arXiv arXiv 2026

  80. [83]

    One-step diffusion models with𝑓-divergence distribution matching, 2025

    YilunXu, WeiliNie, andArashVahdat. One-step diffusion models with𝑓-divergence distribution matching, 2025. URL https://arxiv.org/ abs/2502.15681

    Pith/arXiv arXiv 2025

Showing first 80 references.