Pith. sign in

REVIEW 4 major objections 6 minor 70 references

Alpha as an Efficiency Signal: Visibility-Routed RGBA Image-to-Video Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Joint RGBA video generation beats two-stage matting and runs 1.2x faster.

desk verdict A genuinely new visibility-routing idea for RGBA diffusion, honestly reported, but extractor-derived alpha and unmatched baselines leave the alpha-quality and FVD claims softer than the abstract suggests. read the letter →

arxiv 2608.09355 v1 pith:7OUKZ3ZU submitted 2026-08-10 cs.CV

Zhe Li , Honghao Qiao , Zhixin Xu , Qijie Wang , Bo Peng , Dawei Li This is my paper · ORCID
classification cs.CV
keywords RGBAvideogenerationalphamattingvisibilityroutingdiffusiontransformerflowmatchingtokenskippinggameassetanimationdataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that RGBA video can be generated in a single pass, jointly producing RGB frames and alpha mattes, with quality that beats the traditional generate-then-matte two-stage pipeline, and that the alpha channel itself can be used to reduce inference cost. It introduces GameAlpha-2.4K, a game-asset RGBA video dataset built from synthetic clips with alpha mattes recovered by off-the-shelf tools and checked by staged audits. It then adapts a pretrained image-to-video diffusion transformer through a merged RGB-alpha latent that adds no extra tokens, and adds a visibility router that predicts which tokens will end up fully transparent. Those tokens skip the final two denoising evaluations while being advanced along the original flow-matching schedule; the reported result is a 1.2x backbone speedup with a 35% reduction in late-token evaluations and lower Frechet Video Distance than the two-stage baselines at roughly unchanged identity, aesthetic, and motion quality.

What carries the argument

The load-bearing mechanism is the combination of a merged RGB-A latent and the visibility router with x0-lock. The merged latent keeps the pretrained image-to-video backbone's native 36-channel DiT input by concatenating a 16-channel RGB-alpha target latent with a 20-channel conditioning latent, so alpha costs no extra tokens or branches. The visibility router reads the clean estimate after the second evaluation through a logistic-regression probe, routes tokens predicted fully transparent, transports them along the unchanged flow-matching schedule via x0-lock, caches their K/V for attention context, and reactivates tokens that the evolving foreground boundary reaches. This lets alpha act as the efficiency signal: compute is spent where RGB can actually influence the composite.

What would settle it

Take a set of renderer-native RGBA clips with known ground-truth alpha, including semi-transparent hair, glow, and motion blur, run the paper's pipeline on them, and compare recovered alpha and single-stage outputs against ground truth: if the soft-opacity band shows large mean absolute difference or recompositing onto arbitrary backgrounds visibly fails, the extractor-supervision premise and the quality claim collapse.

Watch

Extended reading notes

Core claim

The central claim is that visibility, read from the model's own early clean estimate, is a valid and useful routing signal for RGBA diffusion inference. Because compositing satisfies $C = \alpha R + (1-\alpha)B$, RGB refinement cannot affect any composite wherever the final $\alpha$ is zero; the paper turns this analytic identity into an inference-time policy. A probe trained on dense trajectories predicts, after the second of four denoising evaluations, which tokens will be effectively transparent (maximum decoded $\alpha$ $\le 0.05$); a dilated spatiotemporal margin protects predicted foreground, uncertain clips fall back to dense inference, and routed tokens advance by x0-lock along the straight flow-matching interpolant toward their stored endpoints instead of passing through the DiT. The paper reports that this single-stage system obtains lower FVD than generate-then-matte baselines on its validation set and matches dense quality closely at 1.2x backbone speedup, with reactivation correcting most early routing errors.

Load-bearing premise

The load-bearing premise is that the extractor-derived alpha sequences are good enough to serve as supervision and evaluation ground truth; if those mattes are systematically wrong in semi-transparent regions, the dataset, the trained model, and the reported alpha metrics inherit that error.

Editorial extensions

If this is right

  • A single-stage RGBA generator can achieve lower FVD than generate-then-matte pipelines on game-style assets, so the decomposition into RGB synthesis followed by matting is not obligatory for quality.
  • The alpha channel gives a task-native token-importance criterion: tokens whose final alpha is zero can be skipped without directly affecting composited output, a signal generic saliency heuristics lack.
  • Skipping is safe when combined with a spatial-temporal margin, local reactivation, and a dense fallback: routing errors are contained rather than propagated.
  • Because the router changes neither the denoising schedule nor backbone parameters, it composes with step distillation and other acceleration techniques.
  • The approach transfers beyond the game-asset training domain: on a human-video matting benchmark the routed model still matches dense quality while dense fallback and reactivation absorb the domain shift.

Reading between the lines

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

  • If the compositing criterion is the right lens, the next natural step is continuous compute allocation by expected compositing error rather than a binary transparent/visible split, which could capture more speed on clips with glow, hair, and motion blur.
  • The 35% skip is well below the dataset's roughly 63% average fully-transparent token budget; foreground-dense clips and the first-frame protection cap the realized gain, so a budget-aware router might approach the oracle bound.
  • Since the merged RGB-A latent adds no token overhead, the cost of RGBA generation could approach that of ordinary RGB video; applying the router to multi-layer or multi-object RGBA generation is a direct testable extension.
  • The router's probe is fit to the model's own estimate rather than external saliency; this suggests the same early-readability result may hold in other flow-matching generators, where the clean estimate is already available.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a single-stage reference-conditioned RGBA image-to-video generator trained on a new GameAlpha-2.4K dataset of synthetic game-asset clips whose alpha mattes are recovered by extractors and selected by staged audits. To avoid extra DiT tokens, the method uses a merged RGB-A VAE, and it introduces a visibility router that predicts final-transparent tokens from the step-2 clean estimate, skips their later DiT updates, and advances them with an x0-lock transport under a spatiotemporal margin with local reactivation. The authors report lower FVD than frozen generate-then-matte baselines, a 1.2x backbone speedup at a 35% evaluation-2 routed fraction, and small quality deltas relative to dense inference. Extensive appendices document dataset construction, routing ablations, cross-domain evaluation on VideoMatte240K, and a blinded human evaluation.

Significance. The core idea, using alpha as a task-native signal for token-level computation routing, is attractive and the analytic criterion in Eq. (6) is a genuine simplification. The x0-lock construction in Eq. (7) is an elegant way to keep routed tokens on the global noise schedule, and the paper is commendably explicit about its limitations, including the extractor-derived alpha supervision and the full-system nature of the comparisons. The dataset is a carefully documented resource for a domain, game assets, that existing RGBA datasets under-serve. However, the headline claims are currently supported only under a specific supervision contract: the alpha used for training, for MAD, and for FVD references all come from the same extractors, and the two-stage baselines are not fine-tuned on the target domain. If the authors add a renderer-ground-truth alpha check and a same-backbone fine-tuned two-stage baseline, the paper would be substantially stronger and the claims of semi-transparent quality and FVD advantage would be on much firmer ground.

major comments (4)
  1. [Sec. 3.1, App. A.4, Sec. 5] The alpha sequences in GameAlpha-2.4K are produced by BiRefNet, SAM3, and MatAnyone2 from synthetic RGB videos and then selected by a multimodal evaluator; these same sequences supervise the RGB-A VAE, the DiT, the transparency-probe labels, and the MAD reference, and they define the reference composites for FVD. Appendix A.4 reports that only 2.2% of pixels fall in the soft band (2/255 < alpha < 253/255), which is exactly the regime where the paper claims two-stage matting fails (Sec. 1). The paper's own limitation statement (Sec. 5) concedes that extractor-selected alpha and the small soft-band coverage mean physical alpha accuracy and recompositability are not established. This is load-bearing because the motivation, the training signal, and the evaluation metric all share the same extractor bias, so the claimed semi-transparent advantage is not yet demonstrated. I would like to see a small renderer-ground-truth subset (or a compositing check with known alpha) used for evaluation and, ideally, for a brief fine-tuning probe.
  2. [Sec. 4.3, Table 1] The two-stage comparison in Table 1 is a full-system comparison: the single-stage model is fine-tuned on GameAlpha-2.4K training clips, while the Wan, CogVideoX, and HunyuanVideo generators and the matting models are used off-the-shelf and frozen. Because the FVD advantage could come from domain adaptation to the GameAlpha distribution rather than from joint RGB-alpha modeling, the claim that the model obtains lower FVD than traditional two-stage pipelines is not yet isolated. The paper acknowledges this in Sec. 4.3 and App. C.6, but the abstract states the comparison without this caveat. Please add a controlled baseline: fine-tune the same Wan2.1-I2V backbone on GameAlpha-2.4K RGB-only (or with dataset alpha as an auxiliary output) and then apply the same matting postprocessors; this would directly test the joint-modeling claim.
  3. [Sec. 4.3, App. B.4] MAD is computed against the extractor-selected alpha sequences, and FVD is computed on checkerboard composites whose reference composites are built from the same sequences. Because the model is trained to reproduce those sequences, both metrics can reward agreement with extractor artifacts rather than with true compositing quality. The description in App. B.4 correctly notes that MAD does not by itself establish physical opacity accuracy or arbitrary-background recompositability, but this caveat is absent from the main-text results. Please report at least one independent alpha metric on a subset where alpha is known by construction (for example, rendered semi-transparent effects or hand-composited overlays with known alpha), and state clearly in Sec. 4.3 that FVD is relative to extractor-selected references.
  4. [Sec. 4.5, Table 3] The operating point (k_route = 2, r_s = 2) and the dense-fallback thresholds are selected on the full validation split, and the headline FVD numbers (172.3 dense vs. 174.4 routed) are reported on that same split with no confidence intervals or significance tests. The 'negligible quality degradation' claim is therefore an in-sample comparison at a self-selected operating point. A held-out split for hyperparameter selection, or bootstrap confidence intervals on the FVD difference, would make the efficiency claim robust.
minor comments (6)
  1. [Sec. 3.1] Section 3.1 contains a near-duplicate passage beginning 'Arbitrary backgrounds make foreground and alpha recovery underconstrained...'; please remove the repeated sentences.
  2. [Abstract, App. C.4] The 35% figure in the abstract is the evaluation-2 initial routed fraction, not the realized skip over evaluations 3-4; Appendix C.4 reports 31.9% realized suffix skip at the requested 0.35 tier. Please make this distinction explicit in the abstract or keep the more precise wording in the body.
  3. [Table 3, Fig. 13] The dense reference FVD differs between the main ablation (172.3) and Appendix C.4 (172.9); please state that these are independent runs under the same harness so the numbers do not appear inconsistent.
  4. [App. C.8] The human evaluation uses an enriched pool of 20 cases rather than a random sample of the validation set; please label the study as exploratory in the main text or in the appendix caption.
  5. [Eqs. (6)-(7), Sec. 3.3] The paper does not justify why active queries should attend to cached K/V of routed tokens whose states are advanced by x0-lock, since those cached K/V come from evaluation 2 and therefore do not reflect the transported state; a brief note on this approximation would improve clarity.
  6. [References] Several references carry future-dated years (e.g., 2026); please check that all citations are complete and correctly dated in the final version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; efficiency claim rests on measured speedup and an analytic compositing criterion.

full rationale

The paper's derivation chain is self-contained. The visibility criterion in Eq. (6) follows from the compositing equation, and the x0-lock update in Eq. (7) is an exact algebraic transport along the flow-matching interpolant defined by the stored endpoint, not a fitted quantity. The central speedup claim is validated by comparing routed inference against dense inference of the same model under a matched protocol (Table 3), so the 1.2x backbone speedup is measured directly and does not depend on external labels. The transparency probe is fitted to the model's own dense trajectories and labeled by the model's own final alpha; this is a self-referential readout, but the paper discloses it explicitly ('The probe is a readout of the model's own evolving prediction, not an external saliency heuristic') and uses it only to select tokens for sparse computation, not to produce the reported quality metrics. The GameAlpha-2.4K alpha is extractor-selected rather than renderer ground truth, and the paper concedes this limitation: 'MAD measures agreement with that reference sequence and does not by itself establish physical opacity accuracy or arbitrary-background recompositability.' That is a data-quality and evaluation-bias limitation, not a constructional circularity: the FVD and MAD comparisons are empirical and could in principle go either way, and the cross-domain VideoMatte240K evaluation uses ground-truth alpha as an external check. No load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in by citation appear. The central efficiency claim is therefore independently grounded, and the disclosed limitations concern dataset supervision and external validity rather than circular reasoning.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a small number of hand-chosen thresholds and margins, on the assumption that extractor-derived alpha is adequate supervision, and on the self-referential probe that learns to predict the model's own final transparency. No new physical or conceptual entities are postulated.

free parameters (7)
  • transparency threshold eta = 0.05
    Decoded alpha at or below 0.05 marks a token effectively transparent; used to create probe self-labels and to define visible support.
  • routing time k_route = 2
    Router activates after DiT evaluation 2; selected from the risk-yield analysis in Sec 4.5, not independently validated across seeds.
  • spatial margin r_s = 2
    Nontransparent support dilated by 2 tokens; selected via risk-yield tradeoff, while Appendix C.3 shows mismatch tails beyond this margin.
  • temporal margin r_t = 1
    Adjacent-time guard for support changes; chosen by design without a dedicated ablation.
  • dense fallback thresholds = 10% candidates, 15% uncertain, delta=2
    Pre-routing gate that keeps some clips on the dense path; no sensitivity analysis is reported for these values.
  • probe feature patch = 2x2x16
    Logistic regression input size; arbitrary patch choice not justified by a sweep.
  • DoRA rank and training epochs = rank 32, 10 epochs
    Capacity and training-budget hyperparameters used for the Stage 2 DiT adaptation.
assumptions (5)
  • standard math Standard alpha compositing equation C = alpha R + (1-alpha) B
    Used in Eq. 6 to derive the visibility criterion; assumes straight alpha compositing without premultiplication or special blending modes.
  • domain assumption Extractor-selected alpha from matte-friendly synthetic videos is adequate supervision for RGBA generation
    GameAlpha-2.4K alpha is produced by BiRefNet, SAM3, or MatAnyone2 and selected by a multimodal evaluator; no renderer ground truth exists.
  • domain assumption The logistic probe trained on dense trajectories of the same model predicts final transparency at evaluation 2
    The probe labels come from the model's own dense final outputs; this self-consistency assumption is central to routing safety.
  • domain assumption Pretrained Wan-VAE encoder and decoder plus Wan2.1-I2V provide a valid base representation and motion prior
    The entire system inherits from frozen Wan components; their correctness and suitability are outside this paper.
  • standard math Flow matching formulation and the four-step LightX2V schedule are valid
    The x0-lock transport assumes the flow-matching interpolant and the fixed noise schedule; standard generative modeling assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Alpha as an Efficiency Signal: Visibility-Routed RGBA Image-to-Video Generation." pith.science (2026). https://pith.science/paper/7OUKZ3ZU

@misc{pith2026260809355,
  author       = {Pith},
  title        = {Pith review of: Alpha as an Efficiency Signal: Visibility-Routed RGBA Image-to-Video Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7OUKZ3ZU}},
  note         = {Machine review of arXiv:2608.09355}
}
read the original abstract

RGBA videos combine RGB appearance with an alpha channel, enabling animated assets to be applied across arbitrary backgrounds, which are heavily used in gaming industry. However, generating high-quality RGBA animations for games remains challenging for two reasons. First, most existing RGBA video datasets are dominated by photorealistic content, with limited coverage of game assets. Second, the traditional generate-then-matte pipelines estimate alpha only after RGB synthesis, so semi-transparent regions are often blurred by background, resulting in unstable matting outputs. More recently, many methods have begun to model RGB and alpha jointly, but existing approaches are mostly text-conditioned, and still have unresolved issues in efficiency and quality. To address these challenges, we introduce GameAlpha-2.4K, a 2.4K-clip game-style RGBA video dataset built with matte-friendly synthesis, multi-hypothesis alpha recovery, and compositing-based quality gates. Using this dataset, we train a reference-conditioned RGBA video generator that jointly produces RGB frames and alpha mattes in a single pass. To improve efficiency, we propose a visibility router that identifies transparent tokens in an early stage and bypasses their later DiT updates, while x_0-lock guides them along the original flow-matching schedule toward self-predicted endpoints. Our model obtains lower FVD than traditional two-stage pipelines, and the visibility router skips 35% of token evaluations in the final two DiT denoising steps, providing a 1.2x backbone speedup with negligible quality degradation compared to dense inference.

Figures

Figures reproduced from arXiv: 2608.09355 by the authors.

Figure 1
Figure 1. Visibility-routed RGBA inference. (a) Evaluations 1–2 process all tokens. After evaluation 2, the router uses xˆ [2] 0 to protect the predicted nontransparent support with a spatiotemporal margin and route the remaining tokens; uncertain samples retain dense inference. Evaluations 3–4 update active tokens, advance routed states with x0-lock, and reactivate tokens reached by the evolving support before merging both s… view at source ↗
Figure 2
Figure 2. Qualitative comparison with two-stage baselines. We pair three RGB generators with BiRefNet or MatAnyone2. Each group shows the RGBA reference (left) and four uniformly sampled frames from the two-stage baseline (top) and ours (bottom). Checkerboards denote transparency; prompts are provided in the appendix D.3. Metrics. FVD (Unterthiner et al. 2019; Skorokhodov, Tulyakov, and Elhoseiny 2021; Yu et al. 2022; Carreir… view at source ↗
Figure 3
Figure 3. plots P90 final-visible-token miss against median full-trajectory token-evaluation saving across routing times and spatial margins. At rs = 2, delaying routing from evalu￾ation 1 to 2 reduces P90 miss from 0.824% to 0.603%, while saving decreases from 22.99% to 15.68%. The 15.68% value is a per-sample median over all four evaluations, whereas the 35% in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: The GameAlpha-2.4K construction pipeline. Distribution-controlled prompt sampling (Stage 1) feeds matte-friendly reference synthesis with a repair loop (Stage 2). The approved RGB reference drives video synthesis and a raw-video gate (Stage 3), while a parallel branch …
Figure 5
Figure 5. Figure 5: Acceptance funnel of GameAlpha-2.4K. From 3,000 proposed tuples, 34 are removed at the reference stage (25 rejected by the reference gate after up to three prompt￾repair attempts, 9 synthesis-service failures), 13 at the video stage (synthesis-service and content scree…
Figure 7
Figure 7. Figure 7: Marginal distributions over equipment, style tier, and viewpoint tier. Closed-loop reweighting keeps every cat￾egory populated but not uniform: acceptance rates differ across categories, and sampling mixes a mainstream and an exploration mode, so observed shares deviat…
Figure 8
Figure 8. Figure 8: Transparency structure of GameAlpha-2.4K. (a) Per-pixel alpha over all GameAlpha-2.4K videos × 97 frames at the training geometry (624×640; log scale): 70.1% of pixels are fully transparent (α ≤ 2/255; exactly 0: 69.2%) and 27.8% effectively opaque (α ≥ 253/255; exactl…
Figure 9
Figure 9. Figure 9: GameAlpha-2.4K samples with their prompts. Each row shows the RGBA reference (left), the video prompt (below; viewpoint and action constraints highlighted), and uniformly sampled frames of the clip (right); checkerboards denote transparency. The prompts are the structu…
Figure 10
Figure 10. Figure 10: Early readability of token transparency (probe fitted on training trajectories; 48 held-out validation trajec￾tories; 1,872,000 tokens per evaluation; identical token set for both curves; the alpha-health gate fired on 0 of 48 clips). Blue (solid, circles): probe scor…
Figure 11
Figure 11. Figure 11: Evolution of decoded alpha estimates (deployed checkpoint; dense validation trajectories). Three manually selected alpha-healthy clips (rows) at evaluations 1–4 (numbered columns) and the final output (rightmost column), decoded directly from the stored clean-estimate…
Figure 12
Figure 12. Figure 12: Within-frame support-mismatch distance CDF (pilot). For alpha-healthy analysis trajectories at k=2: cumu￾lative fraction of raw per-frame support mismatches (tokens transparent in the evaluation-2 estimate but visible in the final output) within a given 2D Chebyshev d…
Figure 14
Figure 14. Figure 14: Deployment statistics of the safety mecha￾nisms (validation split, n=239, primary realization). Left: per-clip distribution of the evaluation-2 routed fraction over the 232 routed clips (dashed: the main paper’s 0.35 initial routed-fraction reference for evaluations 3…
Figure 15
Figure 15. Figure 15: Additional qualitative comparison with SAM3. Each example shows the prompt (left), common RGBA reference (center), and four frames at the same temporal indices from the indicated RGB generator followed by SAM3 (top) and from our single-stage model (bottom). Row pairs …
Figure 16
Figure 16. Figure 16: Additional qualitative comparison with UniVidX. The layout follows [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]
Figure 18
Figure 18. Figure 18: Qualitative routing ablation: static support and reactivation. The layout follows [PITH_FULL_IMAGE:figures/full_fig_p022_18.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 25 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

Show all 70 references
  1. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  2. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  3. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  4. [12]

    2025 , eprint=

    Video Generation with Stable Transparency via Shiftable RGB-A Distribution Learner , author=. 2025 , eprint=

    2025

  5. [13]

    2025 , eprint=

    SAM 3: Segment Anything with Concepts , author=. 2025 , eprint=

    2025

  6. [14]

    arXiv preprint arXiv:2408.00714 , url=

    SAM 2: Segment Anything in Images and Videos , author=. arXiv preprint arXiv:2408.00714 , url=

    arXiv

  7. [15]

    Yang, Peiqing and Zhou, Shangchen and Hao, Kai and Tao, Qingyi , booktitle =

  8. [16]

    Yang, Peiqing and Zhou, Shangchen and Zhao, Jixin and Tao, Qingyi and Loy, Chen Change , booktitle =

  9. [17]

    arXiv: 2306.05399 , year=

    Matting Anything , author=. arXiv: 2306.05399 , year=

    arXiv

  10. [18]

    arXiv preprint arXiv:2601.14255 , year=

    VideoMaMa: Mask-Guided Video Matting via Generative Prior , author=. arXiv preprint arXiv:2601.14255 , year=

  11. [19]

    arXiv preprint arXiv:2402.17113 , year=

    Transparent image layer diffusion using latent transparency , author=. arXiv preprint arXiv:2402.17113 , year=

    arXiv

  12. [20]

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

    Transpixeler: Advancing text-to-video generation with transparency , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  13. [21]

    arXiv preprint arXiv:2603.17944 , year=

    TransText: Alpha-as-RGB Representation for Transparent Text Animation , author=. arXiv preprint arXiv:2603.17944 , year=

  14. [22]

    ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    TransVDM: Motion-Constrained Video Diffusion Model for Transparent Video Synthesis , author=. ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2025 , organization=

    2025

  15. [23]

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

    Transanimate: Taming layer diffusion to generate rgba video , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  16. [24]

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

    Art: Anonymous region transformer for variable multi-layer transparent image generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  17. [25]

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

    Real-time high-resolution background matting , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  18. [26]

    GitHub repository , howpublished =

    LightX2V Contributors , title =. GitHub repository , howpublished =. 2025 , publisher =

    2025

  19. [27]

    arXiv preprint arXiv:2411.19108 , year=

    Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model , author=. arXiv preprint arXiv:2411.19108 , year=

    arXiv

  20. [28]

    The IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    DeepCache: Accelerating Diffusion Models for Free , author=. The IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  21. [29]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification , author=. Advances in Neural Information Processing Systems (NeurIPS) , year =

  22. [30]

    arXiv preprint arXiv:2207.01580 , year=

    Dynamic Spatial Sparsification for Efficient Vision Transformers and Convolutional Neural Networks , author=. arXiv preprint arXiv:2207.01580 , year=

    arXiv

  23. [31]

    2025 , eprint=

    Region-Adaptive Sampling for Diffusion Transformers , author=. 2025 , eprint=

    2025

  24. [32]

    Token Merging: Your

    Bolya, Daniel and Fu, Cheng-Yang and Dai, Xiaoliang and Zhang, Peizhao and Feichtenhofer, Christoph and Hoffman, Judy , booktitle=. Token Merging: Your

  25. [33]

    arXiv preprint arXiv:2410.05317 , year=

    Accelerating Diffusion Transformers with Token-wise Feature Caching , author=. arXiv preprint arXiv:2410.05317 , year=

    arXiv

  26. [34]

    CVPR Workshop on Efficient Deep Learning for Computer Vision , year=

    Token Merging for Fast Stable Diffusion , author=. CVPR Workshop on Efficient Deep Learning for Computer Vision , year=

  27. [35]

    arXiv preprint arXiv:2503.20314 , year=

    Wan: Open and Advanced Large-Scale Video Generative Models , author=. arXiv preprint arXiv:2503.20314 , year=

    arXiv

  28. [36]

    arXiv preprint arXiv:2408.06072 , year=

    CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer , author=. arXiv preprint arXiv:2408.06072 , year=

    arXiv

  29. [37]

    arXiv preprint arXiv:2205.15868 , year=

    CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers , author=. arXiv preprint arXiv:2205.15868 , year=

    arXiv

  30. [38]

    2023 , eprint=

    SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion Models , author=. 2023 , eprint=

    2023

  31. [39]

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

    VACE: All-in-One Video Creation and Editing , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  32. [40]

    ACM SIGGRAPH 2024 Conference Papers , pages=

    I2v-adapter: A general image-to-video adapter for diffusion models , author=. ACM SIGGRAPH 2024 Conference Papers , pages=

    2024

  33. [41]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  34. [42]

    Hu, Edward J and Shen, Yalu and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=

  35. [43]

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

    Adding conditional control to text-to-image diffusion models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  36. [44]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  37. [45]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  38. [46]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    arXiv

  39. [47]

    International Conference on Learning Representations , year=

    AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning , author=. International Conference on Learning Representations , year=

  40. [48]

    2026 , eprint=

    Seedance 2.0: Advancing Video Generation for World Complexity , author=. 2026 , eprint=

    2026

  41. [49]

    arXiv preprint arXiv:1711.05101 , volume=

    Fixing weight decay regularization in adam , author=. arXiv preprint arXiv:1711.05101 , volume=

    arXiv

  42. [50]

    arXiv preprint arXiv:2402.09353 , year=

    DoRA: Weight-Decomposed Low-Rank Adaptation , author=. arXiv preprint arXiv:2402.09353 , year=

    arXiv

  43. [51]

    CAAI Artificial Intelligence Research , volume =

    Bilateral Reference for High-Resolution Dichotomous Image Segmentation , author=. CAAI Artificial Intelligence Research , volume =

  44. [52]

    Transactions on Machine Learning Research Journal , year=

    Dinov2: Learning robust visual features without supervision , author=. Transactions on Machine Learning Research Journal , year=

  45. [53]

    arXiv:2309.16588 , year=

    Vision Transformers Need Registers , author=. arXiv:2309.16588 , year=

    arXiv

  46. [54]

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

    Dinov2 meets text: A unified framework for image-and pixel-level vision-language alignment , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  47. [55]

    FVD: A new metric for video generation , author=

  48. [56]

    arXiv preprint arXiv:2112.14683 , year=

    Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2 , author=. arXiv preprint arXiv:2112.14683 , year=

    arXiv

  49. [57]

    International Conference on Learning Representations , year=

    Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks , author=. International Conference on Learning Representations , year=

  50. [58]

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

  51. [59]

    , author=

    Perceptually Motivated Benchmark for Video Matting. , author=. BMVC , volume=

  52. [60]

    arXiv preprint arXiv:2412.03603 , year=

    Hunyuanvideo: A systematic framework for large video generative models , author=. arXiv preprint arXiv:2412.03603 , year=

    arXiv

  53. [61]

    ACM Transactions on Graphics , volume =

    UniVidX: A Unified Multimodal Framework for Versatile Video Generation via Diffusion Priors , author =. ACM Transactions on Graphics , volume =. 2026 , month = jul, doi =

    2026

  54. [62]

    2025 , journal=

    LayerFlow : A Unified Model for Layer-aware Video Generation , author=. 2025 , journal=

    2025

  55. [63]

    2021 , eprint=

    Robust High-Resolution Video Matting with Temporal Guidance , author=. 2021 , eprint=

    2021

  56. [64]

    arXiv preprint , year=

    VMFormer: End-to-End Video Matting with Transformer , author=. arXiv preprint , year=

  57. [65]

    2024 , eprint=

    Phased Consistency Models , author=. 2024 , eprint=

    2024

  58. [66]

    arxiv , year=

    FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality , author=. arxiv , year=

  59. [67]

    International Conference on Learning Representations , volume=

    Fastercache: Training-free video diffusion model acceleration with high quality , author=. International Conference on Learning Representations , volume=

  60. [68]

    2023 , eprint=

    UniMax: Fairer and more Effective Language Sampling for Large-Scale Multilingual Pretraining , author=. 2023 , eprint=

    2023

  61. [69]

    2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset , author=. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    2017

  62. [70]

    Scandinavian Conference on Image Analysis , year=

    Two-Frame Motion Estimation Based on Polynomial Expansion , author=. Scandinavian Conference on Image Analysis , year=

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.