Pith

open record

sign in

arxiv: 2607.06173 · v1 · pith:2F4ETHR7 · submitted 2026-07-07 · cs.CV

MobileWan: Closing the Quality Gap for Mobile Video Diffusion

Mohsen Ghafoorian , Denis Korzhenkov , Adil Karjauv , Ioannis Lelekas , Noor Fathima , Spyridon Stasis , Hanno Ackermann , Boris van Breugel
show 4 more authors
Markus Nagel Fatih Porikli Animesh Karnewar Amirhossein Habibian
This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-08 14:18 UTCglm-5.2pith:2F4ETHR7record.jsonopen to challenge →

classification cs.CV
keywords video diffusionmobile deploymentrecurrent attentionmodel pruningstep distillationlinear attentionon-device generation
0
0 comments X

The pith

5B video diffusion runs on mobile via recurrence and pruning

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

MobileWan challenges the assumption that mobile video generation requires small models. The authors take a server-scale 5B-parameter video diffusion transformer (Wan2.2-5B) and make it deployable on commercial mobile hardware through four complementary optimizations: learnable attention head pruning, recurrence distillation that converts the transformer into a chunk-wise RNN with constant-memory attention, sampling-step distillation reducing inference to 3 denoising steps, and a memory-optimized VAE decoder with extended temporal context. The resulting system generates 5-second 480x832 videos at 16 FPS in 20 seconds end-to-end on a Snapdragon 8 Gen. 5 NPU, achieving a VBench score of 83.79—higher than previous mobile models that used 0.4–1.8B parameters. The central mechanism is the recurrence distillation framework: by replacing full softmax attention with a hybrid of local softmax (within a chunk) and causal linear attention (across chunks), the transformer operates as an RNN at inference time, maintaining a fixed-size state (s_t, z_t) that summarizes all past key-value information. This reduces memory from growing with video length to constant, which is the key enabler for on-device deployment of a model this large.

Core claim

The paper demonstrates that a 5B-parameter video diffusion transformer can be reformulated to run within the memory and latency constraints of a mobile NPU without reducing the model to a small architecture. The load-bearing technique is converting all 30 transformer blocks to recurrent hybrid attention, where intra-chunk interactions use softmax attention and all cross-chunk context is compressed into fixed-size running accumulators. Combined with 23% head pruning (selected via learnable binary gates trained with a high-noise-biased schedule), 3-step distillation, and an efficient decoder, the system achieves VBench 83.79—competitive with the unoptimized server model's 83.12 and surpassing

What carries the argument

recurrence distillation (chunk-wise RNN reformulation with causal linear attention); learnable per-head binary gates with noise-biased sparsity for head pruning; distribution-matching step distillation; extended temporal receptive field VAE decoder

If this is right

  • The recurrence-as-RNN reformulation means video length is no longer bounded by quadratic attention memory—generation duration becomes a linear cost trade-off rather than a hard memory wall, which could extend to arbitrarily long video generation on fixed hardware.
  • The finding that high-noise-biased training improves pruning decisions suggests that attention head importance is noise-level-dependent: heads critical for global structure (high noise) differ from those for fine detail (low noise), and pruning decisions should be made at the structure-forming stage.
  • If large pretrained models can be ported to mobile via recurrence and pruning rather than training small models from scratch, the standard approach of building compact architectures for edge deployment may be suboptimal compared to compressing existing strong models.
  • The quantization-aware training pipeline (8-bit weights, mixed-precision activations) that recovers quality lost from post-training quantization suggests that aggressive quantization of large diffusion models is feasible when combined with block-wise distillation.

Where Pith is reading between the lines

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

  • The recurrence distillation approach is architecture-agnostic in principle: any transformer-based video diffusion model with softmax attention could be converted to a chunk-wise RNN using the same hybrid attention + causal linear attention recipe, potentially extending the mobile deployment strategy beyond the Wan family.
  • The temporal discontinuities the authors acknowledge suggest that the fixed-size RNN state (s_t, z_t) is a lossy summary of past context—a fundamental information bottleneck. The severity of this bottleneck likely scales with video complexity (more objects, more motion = more information to compress), which could explain why certain scene types (e.g., human faces at middle distance) suffer more.
  • The observation that step distillation trades motion dynamism for sampling speed (lower dynamic degree with fewer steps) implies a fundamental tension between temporal expressivity and inference brevity that may require dedicated motion-preserving distillation objectives rather than generic distribution matching.
  • The 80% human preference over the previous mobile SOTA, combined with competitive scores against server models, suggests the quality gap between mobile and server video generation is now primarily a latency gap rather than a fidelity gap.

Load-bearing premise

The recurrence distillation assumes that causal linear attention with learnable polynomial feature maps can approximate full softmax attention well enough to preserve temporal coherence across chunks, and that the fixed-size RNN state carries sufficient information from past chunks. The authors acknowledge this does not always hold: the full-RNN model exhibits temporal discontinuities and higher temporal jitter, partially mitigated but not fully resolved by subsequent step蒸馏.

What would settle it

If the temporal discontinuities from the RNN reformulation persist across diverse video content types and cannot be resolved by larger state sizes or better feature maps, the constant-memory property that enables mobile deployment would come at an unacceptable quality cost for general-purpose video generation.

Figures

Figures reproduced from arXiv: 2607.06173 by Adil Karjauv, Amirhossein Habibian, Animesh Karnewar, Boris van Breugel, Denis Korzhenkov, Fatih Porikli, Hanno Ackermann, Ioannis Lelekas, Markus Nagel, Mohsen Ghafoorian, Noor Fathima, Spyridon Stasis.

Figure 1
Figure 1. Figure 1: We significantly advance the quality of Mobile Video Generation compared to the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MobileWan: We apply a sequence of complementary optimizations, targeting both the diffusion transformer and the decoder, to the Base Wan 2.2 5B model which requires a typical Server GPU like 80GB-A100 and transform it into MobileWan running natively on a mobile device powered by Snapdragon® 8 Gen. 5 NPU. 1 Introduction Video diffusion models have rapidly become a central paradigm for generative visual mode… view at source ↗
Figure 3
Figure 3. Figure 3: VBench comparison of head pruning meth￾ods. Learnable pruning with high-noise-biased train￾ing performs best, especially at aggressive pruning ratios [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the VAE architecture. We retain the Wan2.2 encoder and modify the MemBlock from LightX2V [35] by extending the causal look-back window at each layer, enabling the prediction of frame 𝐹𝑡 to leverage information from up to 𝐹𝑡−4 [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Evaluation of MobileWan model. We visualize uniformly selected frames from the 81 × 480 × 832 videos generated on Qualcomm Snapdragon® 8 Gen. 5 NPU. tokens, resulting in a drastic reduction in attention weights. This enables a denoising step over the entire video to be completed in 6.6 seconds. Human Evaluation We extend our quantitative evaluation with two internal user studies. In the first s… view at source ↗
Figure 6
Figure 6. Figure 6: Head Pruning FLOPs vs Pruning Ratio. Number of FLOPs and the corresponding reduction at different pruning ratios with respect to the baseline [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Radar plot comparing a subset of our hybrid models with the original Wan2.2 5B model on the full [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Our optimized pipeline matches exactly the base Wan 2.2 model’s instantiation of the latent video [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompt: A bear and a zebra [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Prompt: A dog running happily [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Prompt: A happy fuzzy panda playing guitar nearby a campfire, snow mountain in the background [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Prompt: A person eating a burger [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Prompt: A person is pushing up [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Prompt: A teddy bear is swimming in the ocean [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Full set of VBench dimensions for optimizations steps used by [PITH_FULL_IMAGE:figures/full_fig_p028_15.png] view at source ↗
read the original abstract

Recent advances in video diffusion have been driven by scaling transformer-based architectures to billions of parameters, substantially improving visual fidelity and motion coherence. In contrast, existing mobile video diffusion models remain limited to relatively small parameter budgets, typically 0.4-1.8B, restricting generation quality. In this work, we show that high-quality mobile video generation does not require small models. Instead, we demonstrate that a server-scale 5B-parameter video diffusion transformer can be deployed efficiently on memory-constrained mobile hardware through recurrent reformulation and structured compression. Starting from Wan2.2-5B, we rely on a recurrence distillation framework that converts video generation into a chunk-wise autoregressive process with constant-memory attention computation. Combined with causal linear attention, the model operates as an RNN at inference time while preserving temporal coherence across chunks. We further propose a learnable attention head pruning method based on binary per-head gates optimized end-to-end using a noise-biased sparsity objective and distillation-based finetuning. Together with sampling-step distillation and memory-optimized VAE decoding, MobileWan becomes the first 5B-scale video diffusion model deployable on a commercial mobile device. Our system generates 5-second 480x832 videos at 16 FPS in 20 seconds end-to-end latency, achieving a VBench score of 83.79 and establishing a new state of the art in mobile video generation. Project page: https://qualcomm-ai-research.github.io/mobilewan

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

4 major / 9 minor

Summary. The paper presents MobileWan, a system that deploys a 5B-parameter video diffusion transformer (Wan2.2-5B) on commercial mobile hardware (Snapdragon 8 Gen. 5 NPU). The approach combines four optimizations: (1) learnable per-head attention pruning with a noise-biased sparsity objective, (2) recurrence distillation that converts the transformer into a chunk-wise autoregressive RNN with constant-memory attention via causal linear attention, (3) step distillation (DMD/DMD2) to reduce NFE to 2-3, and (4) an extended-memory 2D VAE decoder. The final system generates 5-second 480x832 videos at 16 FPS in 20 seconds end-to-end, achieving a VBench score of 83.79. The paper provides ablation tables tracing each component's impact, on-device latency/memory measurements, and a user study comparing against Wan2.2-5B and Neodragon.

Significance. Deploying a 5B-parameter video DiT on a mobile NPU is a genuine engineering achievement. The on-device measurements (Table 11) are concrete and include RAM, initialization time, run-time, and spill/fill data — this level of deployment detail is rare in the literature. The learnable head-pruning method with high-noise-biased gate training (Table 1, Figure 3) is a clean, falsifiable contribution that shows consistent improvement over heuristic pruning. The recurrence distillation framework (§3.2, Eqs. 5-10) provides a clear RNN reformulation with well-defined state updates. The commitment to release models and training recipes adds value. The paper builds directly on two prior methods by the same group (ReHyAt [16], Attention Surgery [15]), which is disclosed transparently; the novel contributions are the head-pruning method, the full-RNN conversion of all 30 blocks, the decoder modifications, and the end-to-end mobile integration.

major comments (4)
  1. §3.3 and Table 6: The headline VBench score of 83.79 (Table 5, Table 7) comes from the DMD2 3-step configuration (VBench 84.03 before decoder optimization, per Table 3). However, the user study (Table 6) uses the D-DMD 2-step configuration, which achieves VBench 82.25 (Table 3) — below the Wan2.2 baseline's 83.12. The paper states in §3.3: 'we used D-DMD model with 2 sampling steps for our user study.' This means the 'closing the quality gap' claim combines the best VBench from one variant with the user study from a different, weaker variant. The 83.79-VBench model was never human-evaluated, and the human-evaluated model scores below the baseline on VBench. The paper should either (a) run the user study on the actual deployed DMD2 3-step configuration, or (b) clearly state which configuration is the final deployed system and report all metrics (VBench, user study, latency) for that same,
  2. §3.3, Table 3: The DMD2 3-step model achieves VBench 84.03 before decoder optimization (Table 5), but the semantic score drops from 80.90 (DMD 3-step) to 78.89 (DMD2 3-step) — a 2-point decline. The paper notes that DMD2 produces 'oversaturated' outputs and reduced dynamic degree (45.83 vs 68.33 for the multi-step model). Appendix A.6 acknowledges that 'VBench toolkit gives higher scores to slightly oversaturated colors.' This raises a correctness risk: the 83.79 headline VBench may partly reward oversaturation artifacts that humans dislike. The paper should explicitly discuss whether the VBench improvement from DMD2 reflects genuine quality gains or metric artifacts, and ideally report a color-fidelity or saturation-aware metric for both DMD and DMD2 variants.
  3. Table 6: The user study shows MobileWan loses to Wan2.2 5B in 53% of comparisons (with 25% 'no preference' and 22% preferring MobileWan). The paper frames this as 'in half of the cases is not performing worse,' but 53% loss rate means the baseline is preferred more than twice as often as MobileWan (53% vs 22%). This directly tempers the 'closing the quality gap' claim. The paper should present this result more transparently and discuss what quality dimensions drive the baseline preference.
  4. §3.2, Table 2: Converting all 30 blocks to recurrence drops VBench Total from 83.12 to 83.09, and Appendix A.3.1 attributes the higher dynamic degree of the full-RNN model to 'higher temporal jitters.' The Conclusion also acknowledges that 'in certain cases our RNN reformulation leads to temporal discontinuities.' Since the final system uses 30/30 recurrent blocks, this is a load-bearing quality concern. The paper should quantify the temporal discontinuity rate (e.g., what fraction of generated videos exhibit visible temporal artifacts) and report whether step distillation and decoder optimization fully resolve these artifacts or merely mask them.
minor comments (9)
  1. Figure 2 caption: 'Snapdragon® 8 Gen. 5 NPU' — the period after 'Gen' is unusual; verify this is the correct product name formatting.
  2. Table 1: The 'Total' column for 'High Noise Biased' is 82.19, but the text states the pruning ratio is 'approximately 33%.' Figure 3 shows the learnable method at ~33% achieving VBench ~82.77. These numbers should be reconciled or the difference explained (different pruning ratios, training stages, etc.).
  3. §3.1: The temperature annealing schedule mentions η going from 1.0 to 0.1, but the initial logit value of 5.0 with η=1.0 gives sigmoid(5.0)≈0.993, not exactly 1.0. Minor, but the text says 'close to 1' which is fine — just noting for precision.
  4. Table 4: The 'Mobile Time (ms)' column shows 501 for 'Ours' but 72.60 for LightX2V and Tiny VAE. The text in §4 states decoder latency of 0.50s, which matches 501ms, but the 7x increase over LightX2V deserves discussion in the main text (currently only in the table).
  5. Table 7: The 'Wan2.2 5B*' entry has an asterisk but no footnote explaining what it signifies (presumably the finetuned-for-target-resolution version from Table 10).
  6. §3.2, Eq. (3): The stabilizing constant c_t is described as 'typically the maximum exponent' but its exact computation is not specified. A brief clarification would help reproducibility.
  7. Appendix A.3.1: 'Recurrence latent chunk size T_s is 4' — the main text uses T_c for chunk size. Notation should be consistent.
  8. Table 11: The decoder row for 'Our Decoder' shows 501ms, but Figure 2 states 'Dec. Latency: 0.50s' — consistent, but the text in §4 says '0.50s' while Table 11 says 501ms. Minor rounding inconsistency.
  9. References [15] and [16] are by the same author group and are core to the method. This is disclosed, but the paper could more explicitly state which components are novel vs. inherited from these prior works (e.g., the polynomial feature maps and two-stage training recipe from [15]).

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the thorough and constructive review. The referee correctly identifies a genuine inconsistency in our reporting: the headline VBench score (83.79) comes from the DMD2 3-step configuration, while the user study was conducted on the D-DMD 2-step configuration (VBench 82.25). We agree this must be fixed. We also accept the referee's points about VBench oversaturation artifacts, the framing of the user study results, and the need to quantify temporal discontinuities from the full-RNN conversion. We will revise the manuscript to address all four major comments.

read point-by-point responses

  1. Referee: §3.3 and Table 6: The headline VBench score of 83.79 comes from the DMD2 3-step configuration, but the user study uses the D-DMD 2-step configuration (VBench 82.25, below the Wan2.2 baseline's 83.12). The 'closing the quality gap' claim combines the best VBench from one variant with the user study from a different, weaker variant. The paper should either (a) run the user study on the actual deployed DMD2 3-step configuration, or (b) clearly state which configuration is the final deployed system and report all metrics for that same configuration.

    Authors: The referee is correct, and we acknowledge this inconsistency without reservation. The manuscript as written conflates two different step-distilled variants: the DMD2 3-step model (VBench 83.79, used for the headline number and all quantitative tables) and the D-DMD 2-step model (VBench 82.25, used for the user study). This was an oversight in our reporting, not an intentional cherry-pick, but the effect is the same: the reader cannot form a coherent picture of the deployed system's quality. We will fix this in the revision. Concretely, we will (a) run a new user study on the DMD2 3-step configuration that we report as the final deployed system, and (b) restructure Tables 3, 5, 6, and 7 so that VBench, user study, and latency are all reported for the same configuration. If the DMD2 user study results differ materially from the D-DMD results, we will report both transparently and adjust the 'closing the quality gap' framing accordingly. We will also add an explicit statement in §3.3 identifying the DMD2 3-step model as the final deployed configuration and explaining why the D-DMD variant was initially used for the user study (its more favorable color palette at the time the study was conducted), while making clear that this does not justify reporting metrics across configurations. revision: yes

  2. Referee: §3.3, Table 3: DMD2 3-step achieves VBench 84.03 before decoder optimization, but semantic score drops from 80.90 (DMD 3-step) to 78.89 (DMD2 3-step). Appendix A.6 acknowledges VBench gives higher scores to oversaturated colors. The 83.79 headline VBench may partly reward oversaturation artifacts that humans dislike. The paper should explicitly discuss whether the VBench improvement reflects genuine quality gains or metric artifacts, and ideally report a color-fidelity or saturation-aware metric for both DMD and DMD2 variants.

    Authors: This is a fair and important concern. We already acknowledge in Appendix A.6 that 'VBench toolkit gives higher scores to slightly oversaturated colors, and this does not always align with the human preference,' but this acknowledgment is buried and the main text does not confront the implication for our headline score. We agree that the VBench improvement from DMD2 (84.03 vs. 83.20 for DMD) may be partly inflated by oversaturation, particularly given the semantic score drop (78.89 vs. 80.90) and the reduced dynamic degree (45.83 vs. 68.33 for the multi-step model). In the revision, we will: (1) add an explicit discussion in §3.3 about the risk that DMD2's higher VBench partly reflects metric artifacts rather than genuine quality gains; (2) report the per-dimension VBench breakdown (including the Color dimension) for both DMD and DMD2 variants in the main text rather than only in the appendix; (3) report a saturation-aware metric — we will compute color histogram statistics (mean saturation, saturation variance) for both variants on the VBench prompt set, and if a standard color-fidelity metric is available for the VBench evaluation pipeline, we will include it. We will also temper the headline claims to make clear that the VBench advantage of DMD2 over DMD is not unambiguous. revision: yes

  3. Referee: Table 6: The user study shows MobileWan loses to Wan2.2 5B in 53% of comparisons (22% preferring MobileWan, 25% no preference). The paper frames this as 'in half of the cases is not performing worse,' but 53% loss rate means the baseline is preferred more than twice as often as MobileWan. The paper should present this result more transparently and discuss what quality dimensions drive the baseline preference.

    Authors: We agree that the current framing is misleading. Stating that MobileWan 'in half of the cases is not performing worse' obscures the fact that the baseline is preferred more than twice as often (53% vs. 22%). This is a legitimate criticism of our presentation. In the revision, we will: (1) restate the user study results in neutral terms, reporting the raw preference percentages without spin; (2) acknowledge that the baseline is preferred more often and that this tempers the 'closing the quality gap' claim; (3) add a brief analysis of which quality dimensions drive baseline preference — based on informal feedback from study participants and our own inspection, we expect the main factors to be temporal coherence and motion naturalness (where the full-RNN formulation and step distillation introduce artifacts), but we will verify this with a targeted follow-up survey if feasible. We will also note that the user study was conducted on the D-DMD 2-step variant (per the first comment), so the revised user study on the DMD2 3-step model may yield different results. revision: yes

  4. Referee: §3.2, Table 2: Converting all 30 blocks to recurrence drops VBench Total from 83.12 to 83.09, and Appendix A.3.1 attributes the higher dynamic degree of the full-RNN model to 'higher temporal jitters.' The Conclusion also acknowledges temporal discontinuities. Since the final system uses 30/30 recurrent blocks, this is a load-bearing quality concern. The paper should quantify the temporal discontinuity rate and report whether step distillation and decoder optimization fully resolve these artifacts or merely mask them.

    Authors: We accept this comment. The temporal jitter and discontinuity issue is acknowledged in scattered locations (Appendix A.3.1, Conclusion) but is never quantified, and the claim that 'step distillation resolves the artifacts' (Appendix A.3.1) is made without evidence. This is a gap in our analysis, especially given that the final system uses all 30 recurrent blocks. In the revision, we will: (1) quantify the temporal discontinuity rate by having human annotators rate a random sample of generated videos (e.g., 100 prompts from the VBench set) for visible temporal artifacts, reporting the fraction of videos with noticeable discontinuities; (2) report this rate at multiple pipeline stages — after recurrence distillation (30 blocks), after step distillation, and after decoder optimization — to show whether each stage reduces or masks the artifacts; (3) report the VBench Motion Smoothness and Temporal Flickering sub-scores at each stage as quantitative proxies; (4) revise the Appendix A.3.1 claim to accurately reflect whether step distillation and decoder optimization fully resolve, partially mitigate, or merely mask the temporal artifacts. If the artifacts are not fully resolved, we will state this plainly rather than implying they are resolved. revision: yes

Circularity Check

0 steps flagged

Self-citations [15] and [16] are load-bearing for the recurrence distillation component, but are transparently adopted prior methods, not unverified claims or uniqueness theorems; the central mobile-deployment claim has substantial independent content.

full rationale

The paper builds its recurrence distillation framework (§3.2) on two prior works by overlapping authors: ReHyAt [16] for the hybrid attention formulation (Eq. 3) and Attention Surgery [15] for the learnable polynomial feature maps and training recipe (Appendix A.3.1: 'We follow the efficient training recipe from [15]'). These are self-citations that are load-bearing for one component of the system. However, they do not constitute circularity: (1) the paper transparently cites them as prior published methods (CVPR 2026), not as derived results or uniqueness theorems; (2) the RNN reformulation (Eqs. 5–10) is a standard mathematical equivalence of causal linear attention, not a fit renamed as prediction; (3) no 'prediction' or 'first-principles result' reduces to fitted inputs by construction — all VBench scores, latency measurements, and user study results are empirically measured; (4) the central claim (5B model deployable on mobile) has substantial independent content through the paper's own contributions: learned head pruning with binary gates and noise-biased training (§3.1), decoder optimization with extended temporal look-back (§3.4), step distillation configuration analysis (§3.3), quantization pipeline, and full system integration. The skeptic's concern about different model variants used for VBench vs. user study is a validity concern, not a circularity concern. Score 2 reflects the presence of load-bearing self-citations for one component without reduction of the central claim.

Axiom & Free-Parameter Ledger

10 free parameters · 5 axioms · 2 invented entities

The axiom ledger reveals 10 free parameters (several with unreported values like λ and CFG scale c), 5 axioms (3 domain assumptions, 1 ad-hoc-to-paper, 1 domain assumption with partial counter-evidence), and 2 invented entities both with independent falsifiable evidence. The free parameter count is moderate for a systems paper but several key values are missing from the main text. The most fragile axiom is the causal linear attention sufficiency assumption, which the paper's own limitations section partially contradicts.

free parameters (10)
  • Gate logits α_{b,i} = initialized to 5.0
    Learnable per-head gate logits optimized end-to-end; control which attention heads are pruned. Fitted to data via gradient signals.
  • Sparsity regularizer λ = not explicitly stated
    Controls pruning strength in L_pruning = L_FM + λ Σ g_{b,i}. Value not reported in main text.
  • Temperature η = annealed 1.0 → 0.1
    Controls gate binarization schedule. Annealing schedule is specified but the exact values are design choices.
  • Pruning ratio = 23%
    Selected pruning ratio for the final model. Chosen based on the quality-efficiency trade-off shown in Figure 3.
  • Number of recurrent blocks = 30 (all blocks)
    Selected based on Table 2 comparison of 15/20/25/30 blocks. All 30 blocks chosen for maximum memory efficiency.
  • Chunk size T_c = 4
    Temporal slices per chunk in recurrence distillation. Stated in Appendix A.3.1.
  • Chunk overlap T_o = 2
    Overlapping temporal slices for softmax attention. Stated in Appendix A.3.1.
  • Polynomial degree for φ_q, φ_k = 3
    Degree of learnable polynomial feature maps for linear attention. Stated in Appendix A.3.1.
  • CFG scale c = not explicitly stated
    Classifier-free guidance scale used in DMD step distillation. Referenced in §3.3 but exact value not reported.
  • Noise distribution parameters (LogitNormal) = LogitNormal(1.5,1) for high-noise biased; LogitNormal(0,1) for standard
    Parameters of the noise distribution during gate learning. Stated in §3.1.
axioms (5)
  • domain assumption Linear attention with learnable polynomial feature maps can approximate softmax attention sufficiently for video generation quality.
    Invoked in §3.2, Eq. 2-3. The paper follows [16] for this assumption. The quality degradation (Table 2: 83.12 → 83.09 for 30 blocks) provides partial evidence but also shows the approximation is lossy.
  • domain assumption VBench is a reliable proxy for video generation quality.
    Used throughout as the primary evaluation metric. The paper itself notes in Appendix A.6 that VBench 'gives higher scores to slightly oversaturated colors' that do not align with human preference, undermining this axiom.
  • ad hoc to paper A causal linear attention state (s_t, z_t) of fixed size D'×D carries sufficient temporal context for coherent video generation across chunks.
    Invoked in §3.2, Eqs. 5-10. The paper acknowledges in the Conclusion that 'our RNN reformulation leads to temporal discontinuities,' indicating this assumption does not always hold.
  • domain assumption High-noise timesteps are more informative for pruning decisions than low-noise timesteps.
    Invoked in §3.1 to justify noise-biased gate learning. Supported by Table 1 showing high-noise-biased training outperforms alternatives, but the intuition is stated rather than derived from first principles.
  • domain assumption Post-training quantization with QAT recovery can approximate full-precision model quality on mobile NPUs.
    Invoked in §4 for the quantization pipeline. The paper states PTQ introduces 'noticeable degradation' and QAT 'substantially improves' quality, but no quantized-vs-full-precision VBench comparison is reported.
invented entities (2)
  • Noise-biased sparsity objective independent evidence
    purpose: Guides learnable head pruning by biasing gate optimization toward high-noise timesteps where global structure matters most.
    Falsifiable: Table 1 compares high-noise, standard, and low-noise biased training, showing measurable VBench differences (82.19 vs 80.91 vs 82.19). The method makes a testable prediction about which timesteps are informative for pruning.
  • Extended-memory 2D MemBlock decoder independent evidence
    purpose: Replaces causal Conv3D with extended look-back 2D convolutions for efficient mobile VAE decoding.
    Falsifiable: Table 4 compares the decoder against LightX2V and Tiny VAE baselines on PSNR and VBench, showing measurable improvements (PSNR 28.85 vs 25.01).

pith-pipeline@v1.1.0-glm · 26822 in / 4049 out tokens · 619407 ms · 2026-07-08T14:18:04.840983+00:00 · methodology

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

86 extracted references · 86 canonical work pages · 11 internal anchors

  1. [1]

    InICLR, 2026

    Animesh Karnewar, Denis Korzhenkov, Ioannis Lelekas, Noor Fathima, Adil Karjauv, Mohsen Ghafoorian,andAmirhosseinHabibian.Neodragon: Mobilevideogenerationusingdiffusiontransformer. InICLR, 2026

    work page 2026

  2. [2]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. https://openai.com/index/ video-generation-models-as-world-simulators/, 2024. OpenAI technical report, accessed 2026-05-05

    work page 2024

  3. [3]

    HunyuanVideo: A Systematic Framework For Large Video Generative Models

    Hunyuan Foundation Model Team. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Zhang Yuxuan, Weihan Wang, Yean Cheng, Bin Xu, Xiaotao Gu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. InICLR, 2025

    work page 2025

  5. [5]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Team Wan et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    Efficient Video Diffusion Models: Advancements and Challenges

    Shitong Shao, Lichen Bai, Pengfei Wan, James Kwok, and Zeke Xie. Efficient video diffusion models: Advancements and challenges.arXiv preprint arXiv:2604.15911, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  7. [7]

    Moalign: Motion-centric representation alignment for video diffusion models.arXiv preprint arXiv:2510.19022, 2025

    Aritra Bhowmik, Denis Korzhenkov, Cees GM Snoek, Amirhossein Habibian, and Mohsen Ghafoo- rian. Moalign: Motion-centric representation alignment for video diffusion models.arXiv preprint arXiv:2510.19022, 2025

    work page arXiv 2025

  8. [8]

    A Systematic Post-Train Framework for Video Generation

    ZeyueXue,SimingFu,JieHuang,ShuaiLu,HaoranLi,YijunLiu,YumingLi,XiaoxuanHe,Mengzhao Chen, Haoyang Huang, et al. A systematic post-train framework for video generation.arXiv preprint arXiv:2604.25427, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  9. [9]

    Mobile-oriented video diffusion: Enabling text-to-video generation on mobile devices without retraining, compression, or pruning

    Bosung Kim, Kyuhwan Lee, Isu Jeong, Jungmin Cheon, Yeojin Lee, and Seulki Lee. Mobile-oriented video diffusion: Enabling text-to-video generation on mobile devices without retraining, compression, or pruning. InWACV, 2026

    work page 2026

  10. [10]

    Tamingdiffusiontransformer for real-time mobile video generation.arXiv preprint arXiv:2507.13343, 2025

    Yushu Wu, Yanyu Li, Anil Kag, Ivan Skorokhodov, Willi Menapace, Ke Ma, Arpit Sahni, Ju Hu, AliaksandrSiarohin,DhritimanSagar,YanzhiWang,andSergeyTulyakov. Tamingdiffusiontransformer for real-time mobile video generation.arXiv preprint arXiv:2507.13343, 2025

    work page arXiv 2025

  11. [11]

    S2dit: Sandwichdiffusiontransformerformobilestreaming video generation.arXiv preprint arXiv:2601.12719, 2026

    Lin Zhao, Yushu Wu, Aleksei Lebedev, Dishani Lahiri, Meng Dong, Arpit Sahni, Michael Vasilkovsky, HaoChen,JuHu,AliaksandrSiarohin,etal. S2dit: Sandwichdiffusiontransformerformobilestreaming video generation.arXiv preprint arXiv:2601.12719, 2026

    work page arXiv 2026

  12. [12]

    Mobile video diffusion

    Haitam Ben Yahia, Denis Korzhenkov, Ioannis Lelekas, Amir Ghodrati, and Amirhossein Habibian. Mobile video diffusion. InICCV, 2025

    work page 2025

  13. [13]

    Snapgen-v: Generating a five-second video within five seconds on a mobile device

    YushuWu, ZhixingZhang, YanyuLi, YanwuXu, AnilKag, YangSui, HuseyinCoskun, KeMa, Aleksei Lebedev, Ju Hu, et al. Snapgen-v: Generating a five-second video within five seconds on a mobile device. InCVPR, 2025

    work page 2025

  14. [14]

    Sana-video: Efficient video generation with block linear diffusion transformer

    Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, et al. Sana-video: Efficient video generation with block linear diffusion transformer. InProceedings of the International Conference on Learning Representations (ICLR), 2026

    work page 2026

  15. [15]

    Attention surgery: An efficient recipe to linearize your video diffusion transformer

    Mohsen Ghafoorian, Denis Korzhenkov, and Amirhossein Habibian. Attention surgery: An efficient recipe to linearize your video diffusion transformer. InCVPR, 2026

    work page 2026

  16. [16]

    Rehyat: Recurrent hybrid attention for video diffusion transformers

    Mohsen Ghafoorian and Amirhossein Habibian. Rehyat: Recurrent hybrid attention for video diffusion transformers. InCVPR, 2026

    work page 2026

  17. [17]

    MoViE: Mobile Diffusion for Video Editing

    Adil Karjauv, Noor Fathima, Ioannis Lelekas, Fatih Porikli, Amir Ghodrati, and Amirhossein Habibian. Movie: Mobile diffusion for video editing.arXiv preprint arXiv:2412.06578, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    Clockwork diffusion: Efficient generation with model-step distillation

    Amirhossein Habibian, Amir Ghodrati, Noor Fathima, Guillaume Sautiere, Risheek Garrepalli, Fatih Porikli, and Jens Petersen. Clockwork diffusion: Efficient generation with model-step distillation. In CVPR, 2024

    work page 2024

  19. [19]

    Amd-hummingbird: Towards an efficient text-to-video model.arXiv preprint arXiv:2503.18559, 2025

    Takashi Isobe, He Cui, Dong Zhou, Mengmeng Ge, Dong Li, and Emad Barsoum. Amd-hummingbird: Towards an efficient text-to-video model.arXiv preprint arXiv:2503.18559, 2025

    work page arXiv 2025

  20. [20]

    Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023

    work page 2023

  21. [21]

    On-device Sora: Enabling Training-Free Diffusion-based Text-to-Video Generation for Mobile Devices

    Bosung Kim, Kyuhwan Lee, Isu Jeong, Jungmin Cheon, Yeojin Lee, and Seulki Lee. On-device sora: Enabling training-free diffusion-based text-to-video generation for mobile devices.arXiv preprint arXiv:2502.04363, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    SANA 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer

    Enze Xie, Junsong Chen, Yuyang Zhao, Jincheng YU, Ligeng Zhu, Yujun Lin, Zhekai Zhang, Muyang Li, Junyu Chen, Han Cai, Bingchen Liu, Daquan Zhou, and Song Han. SANA 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer. InICML, 2025

    work page 2025

  23. [23]

    FastLightGen: Fast and light video generation with fewer steps and parameters

    Shitong Shao, Yufei Gu, and Zeke Xie. FastLightGen: Fast and light video generation with fewer steps and parameters. InCVPR, 2026

    work page 2026

  24. [24]

    Hadzic, Manling Li, AgrimGupta,StefanoMassaroli,AzaliaMirhoseini,JuanCarlosNiebles,StefanoErmon,andLiFei-Fei

    Keshigeyan Chandrasegaran, Michael Poli, Daniel Y Fu, Dongjun Kim, Lea M. Hadzic, Manling Li, AgrimGupta,StefanoMassaroli,AzaliaMirhoseini,JuanCarlosNiebles,StefanoErmon,andLiFei-Fei. Exploring diffusion transformer designs via grafting. InNeurIPS, 2026

    work page 2026

  25. [25]

    Freeman, and Taesung Park

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand, William T. Freeman, and Taesung Park. One-step Diffusion with Distribution Matching Distillation. InCVPR, 2024

    work page 2024

  26. [26]

    Classifier-Free Diffusion Guidance

    Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021

    work page 2021

  27. [27]

    Decoupled DMD: CFG augmentation as the spear, distribution matching as the shield

    Dongyang Liu, Peng Gao, David Liu, Ruoyi Du, Zhen Li, Qilong Wu, Xin Jin, Sihan Cao, Shifeng Zhang, Steven HOI, and Hongsheng Li. Decoupled DMD: CFG augmentation as the spear, distribution matching as the shield. InICLR, 2026

    work page 2026

  28. [28]

    Barron, and Ben Mildenhall

    Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. InICLR, 2023

    work page 2023

  29. [29]

    Jacobs, Alexei A

    David McAllister, Songwei Ge, Jia-Bin Huang, David W. Jacobs, Alexei A. Efros, Aleksander Holynski, and Angjoo Kanazawa. Rethinking Score Distillation as a Bridge Between Image Distributions. In NeurIPS, 2024

    work page 2024

  30. [30]

    Compositional visual generation with energy based models

    Yilun Du, Shuang Li, and Igor Mordatch. Compositional visual generation with energy based models. InNeurIPS, 2020

    work page 2020

  31. [31]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. InICML, 2024

    work page 2024

  32. [32]

    Faster video diffusion with trainable sparse attention

    Peiyuan Zhang, Yongqi Chen, Haofeng Huang, Will Lin, Zhengzhong Liu, Ion Stoica, Eric P Xing, and Hao Zhang. Faster video diffusion with trainable sparse attention. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    work page 2025

  33. [33]

    Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved Distribution Matching Distillation for Fast Image Synthesis. InNeurIPS, 2024

    work page 2024

  34. [34]

    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

    work page 2026

  35. [35]

    Lightx2v: Light video generation inference framework.https://github

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

    work page 2025

  36. [36]

    https://github.com/madebyollin/ taehv, 2025

    OllinBoerBohan.Taehv: Tinyautoencoderforhunyuanvideo. https://github.com/madebyollin/ taehv, 2025

    work page 2025

  37. [37]

    Abenchmark dataset and evaluation methodology for video object segmentation

    F.Perazzi,J.Pont-Tuset,B.McWilliams,L.VanGool,M.Gross,andA.Sorkine-Hornung. Abenchmark dataset and evaluation methodology for video object segmentation. InCVPR, 2016

    work page 2016

  38. [38]

    The 2018 DAVIS Challenge on Video Object Segmentation

    Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation.arXiv:1803.00557, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  39. [39]

    VBench: Comprehensive benchmark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu,QingyangJin,NattapolChanpaisit,YaohuiWang,XinyuanChen,LiminWang,DahuaLin,YuQiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. InCVPR, 2024

    work page 2024

  40. [40]

    Fastforward: Neural network quantization for research and prototyping

    Qualcomm AI Research. Fastforward: Neural network quantization for research and prototyping. https://github.com/Qualcomm-AI-research/fastforward, 202X

    work page

  41. [41]

    Open-Sora Plan: Open-Source Large Video Generation Model

    Bin Lin, Yunyang Ge, Xinhua Cheng, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  42. [42]

    Open-Sora: Democratizing Efficient Video Production for All

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  43. [43]

    LTX-Video: Realtime Video Latent Diffusion

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  44. [44]

    Pyramidal flow matching for efficient video generative modeling

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong MU, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. InThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  45. [45]

    M4V: Multi-Modal Mamba for Text-to-Video Generation

    Jiancheng Huang, Gengwei Zhang, Zequn Jie, Siyu Jiao, Yinlong Qian, Ling Chen, Yunchao Wei, and Lin Ma. M4v: Multi-modal mamba for text-to-video generation.arXiv preprint arXiv:2506.10915, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  46. [46]

    Fast video generation with sliding tile attention

    Peiyuan Zhang, Yongqi Chen, Runlong Su, Hangliang Ding, Ion Stoica, Zhengzhong Liu, and Hao Zhang. Fast video generation with sliding tile attention. InICML, 2025

    work page 2025

  47. [47]

    Tinyfusion: Diffusion transformers learned shallow

    Gongfan Fang, Kunjun Li, Xinyin Ma, and Xinchao Wang. Tinyfusion: Diffusion transformers learned shallow. InCVPR, 2025

    work page 2025

  48. [48]

    Learnable sparsity for vision generative models

    Yang Zhang, Er Jin, Wenzhong Liang, Yanfei Dong, Ashkan Khakzar, Philip Torr, Johannes Stegmaier, and Kenji Kawaguchi. Learnable sparsity for vision generative models. InICLR, 2026

    work page 2026

  49. [49]

    OBS-diff: Accurate pruning for diffusion models in one-shot

    Junhan Zhu, Hesong Wang, Mingluo Su, Zefang Wang, and Huan Wang. OBS-diff: Accurate pruning for diffusion models in one-shot. InICLR, 2026

    work page 2026

  50. [50]

    Generating long sequences with sparse transformers, 2019

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

    work page 2019

  51. [51]

    Peters, and Arman Cohan

    Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The long-document transformer, 2020

    work page 2020

  52. [52]

    Bigbird: Transformers for longer sequences

    Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon,PhilipPham,AnirudhRavula,QifanWang,LiYang,andAmrAhmed. Bigbird: Transformers for longer sequences. InNeurIPS, 2020

    work page 2020

  53. [53]

    Efficient content-based sparse attention with routing transformers.Transactions of the Association for Computational Linguistics, 2021

    Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. Efficient content-based sparse attention with routing transformers.Transactions of the Association for Computational Linguistics, 2021

    work page 2021

  54. [54]

    Object-centric diffusion for efficient video editing

    Kumara Kahatapitiya, Adil Karjauv, Davide Abati, Fatih Porikli, Yuki M Asano, and Amirhossein Habibian. Object-centric diffusion for efficient video editing. InECCV, 2024

    work page 2024

  55. [55]

    Li, Madian Khabsa, Han Fang, and Hao Ma

    Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity, 2020

    work page 2020

  56. [56]

    Reformer: The efficient transformer

    Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. InICLR, 2020

    work page 2020

  57. [57]

    Transformers are rnns: fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: fast autoregressive transformers with linear attention. InICML, 2020

    work page 2020

  58. [58]

    Rethinking attention with performers

    Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J Colwell, and Adrian Weller. Rethinking attention with performers. InICLR, 2021

    work page 2021

  59. [59]

    Nyströmformer: A nyström-based algorithm for approximating self-attention

    Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nyströmformer: A nyström-based algorithm for approximating self-attention. InAAAI, 2021

    work page 2021

  60. [60]

    Flashattention: Fast and memory-efficient exact attention with IO-awareness

    Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Re. Flashattention: Fast and memory-efficient exact attention with IO-awareness. InNeurIPS, 2022

    work page 2022

  61. [61]

    Flashattention-2: Faster attention with better parallelism and work partitioning

    Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. InICLR, 2024

    work page 2024

  62. [62]

    HLA:HadamardLinear Attention, 2026

    HannoAckermann,HongCai,MohsenGhafoorian,andAmirhosseinHabibian. HLA:HadamardLinear Attention, 2026

    work page 2026

  63. [63]

    Gonzalez, Jianfei Chen, and Jun Zhu

    Jintao Zhang, Haoxu Wang, Kai Jiang, Shuo Yang, Kaiwen Zheng, Haocheng Xi, Ziteng Wang, Hongzhou Zhu, Min Zhao, Ion Stoica, Joseph E. Gonzalez, Jianfei Chen, and Jun Zhu. SLA: Beyond sparsity in diffusion transformers via fine-tunable sparse–linear attention. InICLR, 2026

    work page 2026

  64. [64]

    Liteattention: A temporal sparse attention for diffusion transformers, 2025

    Dor Shmilovich, Tony Wu, Aviad Dahan, and Yuval Domb. Liteattention: A temporal sparse attention for diffusion transformers, 2025

    work page 2025

  65. [65]

    Adaptor: Adaptive token reduction for video diffusion transformers

    Elia Peruzzo, Adil Karjauv, Nicu Sebe, Amir Ghodrati, and Amir Habibian. Adaptor: Adaptive token reduction for video diffusion transformers. InCVPR, 2025

    work page 2025

  66. [66]

    Astraea: A token-wise acceleration framework for video diffusion transformers

    Haosong Liu, Yuge Cheng, Wenxuan Miao, Zihan Liu, Aiyue Chen, Jing Lin, Yiwu Yao, Chen Chen, Jingwen Leng, Minyi Guo, and Yu Feng. Astraea: A token-wise acceleration framework for video diffusion transformers. InICLR, 2026

    work page 2026

  67. [67]

    TPDiff: Temporal pyramid video diffusion model

    Lingmin Ran and Mike Zheng Shou. TPDiff: Temporal pyramid video diffusion model. InICLR, 2026

    work page 2026

  68. [68]

    PyramidalWan: On making pretrained video model pyramidal for efficient inference

    Denis Korzhenkov, Adil Karjauv, Animesh Karnewar, Mohsen Ghafoorian, and Amirhossein Habibian. PyramidalWan: On making pretrained video model pyramidal for efficient inference. InCVPR, 2026

    work page 2026

  69. [69]

    Fast high-resolution image synthesis with latent adversarial diffusion distillation

    Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. InSIGGRAPH Asia 2024 Conference Papers, 2024

    work page 2024

  70. [70]

    Align your flow: Scaling continuous-time flow map distillation

    Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your flow: Scaling continuous-time flow map distillation. InNeurIPS, 2026

    work page 2026

  71. [71]

    Sana-sprint: One-step diffusion with continuous-time consistency distillation

    Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Song Han, and Enze Xie. Sana-sprint: One-step diffusion with continuous-time consistency distillation. InICCV, 2025

    work page 2025

  72. [72]

    Twinflow: Realizing one-step generation on large models with self-adversarial flows

    Zhenglin Cheng, Peng Sun, Jianguo Li, and Tao Lin. Twinflow: Realizing one-step generation on large models with self-adversarial flows. InICLR, 2026

    work page 2026

  73. [73]

    Metaxas, Sergey Tulyakov, and Jian Ren

    Zhixing Zhang, Yanyu Li, Yushu Wu, yanwu xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris N. Metaxas, Sergey Tulyakov, and Jian Ren. SF-v: Single forward video generation model. InNeurIPS, 2024

    work page 2024

  74. [74]

    Diffusion adversarial post-training for one-step video generation

    Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation. InICML, 2025

    work page 2025

  75. [75]

    Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models

    Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models. InNeurIPS, 2023

    work page 2023

  76. [76]

    Phaseddmd: Few-stepdistributionmatchingdistillationviascorematching within subintervals, 2026

    Xiangyu Fan, Zesong Qiu, Zhuguanyu Wu, Fanzhou Wang, Zhiqian Lin, Tianxiang Ren, Dahua Lin, RuihaoGong,andLeiYang. Phaseddmd: Few-stepdistributionmatchingdistillationviascorematching within subintervals, 2026

    work page 2026

  77. [77]

    MagicDistillation: Weak-to-strong video distillation for large-scale few-step synthesis, 2025

    Shitong Shao, Hongwei Yi, Hanzhong Guo, Tian Ye, Daquan Zhou, Michael Lingelbach, Zhiqiang Xu, and Zeke Xie. MagicDistillation: Weak-to-strong video distillation for large-scale few-step synthesis, 2025

    work page 2025

  78. [78]

    Transition matching distillation for fast video generation, 2026

    Weili Nie, Julius Berner, Nanye Ma, Chao Liu, Saining Xie, and Arash Vahdat. Transition matching distillation for fast video generation, 2026

    work page 2026

  79. [79]

    HiStream: Efficient high-resolution video generation via redundancy-eliminated streaming, 2025

    Haonan Qiu, Shikun Liu, Zijian Zhou, Zhaochong An, Weiming Ren, Zhiheng Liu, Jonas Schult, Sen He, Shoufa Chen, Yuren Cong, Tao Xiang, Ziwei Liu, and Juan-Manuel Perez-Rua. HiStream: Efficient high-resolution video generation via redundancy-eliminated streaming, 2025

    work page 2025

  80. [80]

    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. InICLR, 2026

    work page 2026

Showing first 80 references.