Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

A pointwise multi-view reprojection reward can align video diffusion models toward geometric consistency without labeled depth or camera data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A VGGT-based pointwise reprojection reward with geometry-aware sampling improves video geometric consistency via SFT/DPO and causal test-time search.

T0 review reviewed 2026-07-13 challenge →

load-bearing objection Useful practical geometry reward plus real causal TTS search; VGGT circularity softens the 3D claim but does not erase the relative gains. the 3 major comments →

arxiv 2603.16271 v3 pith:27XHHWAK submitted 2026-03-17 cs.CV

VIGOR: VIdeo Geometry-Oriented Reward for Temporal Generative Alignment

Tengjiao Yin , Jinglei Shi , Heng Guo , Xi Wang This is my paper
classification cs.CV
keywords video diffusiongeometry-based rewardmulti-view consistencypointwise reprojectiontest-time scalingcausal video generationpreference alignmentgeometry-aware sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Video diffusion models still produce object deformation, spatial drift, and depth violations because training never supplies explicit geometry. This paper argues that a pretrained geometric foundation model can act as a reward: keep only points in attention-salient, high-structure regions, back-project them with estimated depth and cameras, reproject into other frames, and treat the tracker-versus-reprojection mismatch as a consistency score. That score is used two ways—building preference pairs for supervised fine-tuning and direct preference optimization on bidirectional generators, and guiding seed-and-path search at inference time on causal streaming generators. Experiments report that the pointwise form is more robust than epipolar distance or pixel-warping rewards, and that modest test-time budgets already lift 3D reconstruction and multi-view metrics. The practical claim is geometric cleanup of open video models without massive retraining or paired 3D labels.

Core claim

Multi-view geometric consistency of generated video can be measured, and therefore optimized, by the mean pointwise L2 error between tracker correspondences and foundation-model reprojections of the same points, after geometry-aware sampling has discarded low-texture and non-semantic regions. This reward is more physically grounded than pixel-intensity warping or pure epipolar scores, and it is sufficient both to curate preference data for SFT and DPO and to serve as a path verifier for test-time search on causal generators.

What carries the argument

Geometry-based reward (Reproj-Pts): geometry-aware sampling of patch centers from shallow global attention of a geometric foundation model, followed by tracker correspondences and mean L2 cross-frame reprojection error using feed-forward depth and cameras. That scalar ranks videos for preference learning and verifies generation paths at test time.

Load-bearing premise

The method assumes a pretrained geometric foundation model still gives trustworthy depth and camera estimates on synthetic, artifact-heavy video, so that tracker-versus-reprojection disagreement truly ranks geometric quality rather than shared foundation-model failures.

What would settle it

Inject controlled geometric faults (warp, depth flip, or drift) into otherwise clean clips and check whether the reward still ranks clean above corrupted versions when the foundation model’s depth and cameras are replaced by an independent reconstruction stack; if rankings reverse or collapse, the reward is not a faithful geometry signal.

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

If this is right

  • Bidirectional video models fine-tuned with SFT or DPO on this reward improve 3D reconstruction and multi-view consistency metrics without geometric ground truth.
  • Causal streaming generators can be improved at inference by Search-on-Start, Search-on-Path, or Beam Search that treat the reward as a path verifier, with gains that scale with budget.
  • Open-source video models can gain geometric realism through inference-time scaling alone when retraining budgets are unavailable.
  • Preference pairs ranked by this reward form a reusable geometry-alignment resource for other generators.

Where Pith is reading between the lines

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

  • If the same foundation model is both reward and primary 3D evaluator, reported gains may partly reflect agreement with that model rather than independent geometric truth; an external reconstruction pipeline would be a sharper test.
  • The causal path-search strategies suggest any sequential generator could use intermediate geometric verifiers to prune rollouts, a pattern that may transfer to long-horizon 3D or world models.
  • Geometry-aware attention sampling may also serve as a cheap saliency prior for other dense geometric losses beyond reward learning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes VIGOR, a geometry-based reward for video diffusion models that scores multi-view consistency via pointwise cross-frame reprojection error using VGGT-estimated depth and cameras, combined with Geometry-Aware Sampling (GAS) that selects patches from shallow VGGT global attention. The reward is used in two pathways: (i) offline preference construction (GB3DV-25k) for SFT and Flow-DPO on a bidirectional DiT, and (ii) test-time scaling on bidirectional Best-of-N and, more notably, on causal/streaming generators via Search on Start, Search on Path, and Beam Search with the reward as a path verifier. Experiments report gains over baseline, epipolar, and pixel-reprojection rewards on VGGT re-render metrics, multi-view scores, and VBench, with qualitative reductions in deformation, drift, and depth violations.

Significance. If the geometric signal is reliable, the work is a useful practical contribution for open-source video generation: it avoids paired geometric ground truth, packages a reusable preference dataset and reward, and is among the first systematic treatments of reward-guided test-time search for causal/streaming video models (SoS/SoP/Beam Search). The dual post-training and inference-time pathways, and the explicit comparison of pointwise vs pixel-space reprojection, are valuable for the community. Strengths include a clear algorithmic framework, budget-scaling curves for causal TTS, and consistent relative gains versus Epipolar and Reproj-Pix on several reported metrics. The main significance is engineering and empirical rather than a new geometric theory; the value hinges on whether the reward measures true multi-view consistency rather than agreement with a single foundation model.

major comments (3)
  1. Circularity between reward and primary 3D evaluation. The reward (§4, Eqs. 1–4) is defined from VGGT depth/extrinsics plus tracker–reprojection disagreement; the main “3D reconstruction quality” protocol (§6.1) again runs VGGT and reports PSNR/SSIM/LPIPS of reprojected clouds, and Tables 1–2 also report RPT (the paper’s own pointwise metric). Relative gains over Reproj-Pix partially control for shared VGGT use, and EPI/VBench are more independent, but the central claim of a “more physically grounded” geometric reward is not isolated from matching VGGT’s own geometry on generated (often OOD) video. At minimum, re-evaluate primary 3D metrics with an independent estimator (e.g., DUSt3R/MASt3R or COLMAP-style SfM where feasible) and report whether Reproj-Pts still leads; ideally also ablate the reward oracle itself.
  2. Table 2: post-hoc alignment collapses Dynamic Degree. Baseline DD is 57.74; SFT falls to 15.62 and DPO (Reproj-Pts) to 25.78, while geometric metrics improve. This is a load-bearing trade-off for the claim that geometric preference alignment “enhances perceptual quality alongside geometric fidelity” (§6.4). The motion penalty (Eq. 8) does not prevent the collapse, suggesting the preference signal may favor low-motion clips that are easier to make multi-view consistent. The paper needs quantitative analysis (motion magnitude vs reward, win-rate stratified by dynamics), stronger motion-preserving constraints, and/or human preference on dynamic scenes before the post-training pathway can be considered successful.
  3. Reliability of VGGT as oracle on generated video is assumed, not tested (§4; weakest premise for Eq. 4). Tracker–reprojection error can reward shared failure modes (bad depth/pose on deformed or textureless regions) rather than true consistency; GAS (§4.1) further concentrates samples on VGGT high-attention patches, amplifying dependence. A short stress test—corruptions, known geometric failure cases, or correlation of reward with human geometric rankings on a held-out set—would make the “superior robustness” claim falsifiable rather than circular with the evaluation stack.
minor comments (6)
  1. Causal TTS budget study uses a 16-clip subset (§6.3); state selection criteria and report variance/error bars so scaling trends in Fig. 4 are interpretable.
  2. Computational cost of the reward (VGGT + tracking per candidate/path) is not reported; for TTS this is central to the “efficient inference-time scaling” claim in the abstract.
  3. Preference construction (best/worst of 10 seeds per prompt, §5.1) should be justified: how often is the geometric gap large, and is Bradley–Terry appropriate when both clips can be geometrically poor?
  4. Clarify free hyperparameters (GAS top-k fraction, DPO β/λ, SoP window, beam B/K) and sensitivity; several appear only implicitly in experiments.
  5. Related work: position more carefully against concurrent VideoGPA [11] (also VGGT-based) and state what is unique beyond pointwise error and causal TTS.
  6. Presentation: many equations and symbols in the submitted text are hard to parse (garbled subscripts); ensure camera model (Eqs. 1–2) and Flow-DPO (Eq. 7) are clean in the camera-ready version. Fig. 3 would benefit from a clearer legend for attention heatmaps vs sampled centers.

Circularity Check

2 steps flagged

Partial evaluation circularity: the pointwise reprojection reward (Eq. 4) and primary 3D metrics both rest on the same VGGT oracle, so Best-of-N / preference gains on RPT and VGGT re-render scores are partly by construction; EPI and VBench remain independent.

specific steps
  1. self definitional [§4.2 Eq. 4 (reward) vs §6.1 multi-view metrics / Tables 1–2 (RPT)]
    "The final reprojection error metric is the mean L2 distance over all valid point-frame pairs: ������� = (1/|V|) ∑_{(i,j,k)∈V} ||û^{(i)}_j − u^{(i)}_j||_2 (4) where V denotes the set of valid point-frame pairs. [...] For multi-view consistency, we compute three scores on the evaluated videos: epipolar consistency (EPI), reprojection-pixelwise (RPX), and reprojection-pointwise (RPT)."

    RPT is the same pointwise reprojection construction used as the reward. Under Best-of-N TTS, candidates are selected by maximizing this score; reporting superior RPT for Reproj-Pts is then true by construction, not an independent test of geometric robustness. Preference ranking for SFT/DPO likewise optimizes the same quantity that later appears as an evaluation column.

  2. other [§4 (reward via VGGT depth/extrinsics + GAS) and §6.1 (3D reconstruction quality via VGGT)]
    "we propose a geometry-based reward model that leverages pretrained geometric foundation models to evaluate multi-view consistency through cross-frame reprojection error. [...] For 3D reconstruction quality, we apply VGGT [42] to uniformly sampled frames to obtain depth maps and camera poses, reproject the recovered 3D point cloud into each target frame, and measure reprojection fidelity via PSNR, SSIM, and LPIPS."

    The reward’s geometry (depth, extrinsics, and attention-based sampling of “geometrically meaningful” patches) is supplied by VGGT; the headline 3D metrics re-apply the same VGGT stack and score re-render fidelity of its point clouds. Gains on PSNR/SSIM/LPIPS therefore partly measure agreement with the same foundation-model failure modes that shaped the preference signal, rather than geometry external to that oracle. This is evaluation contamination, not a fitted-parameter prediction, but it still couples training signal to reported geometric success.

full rationale

This is an engineering alignment paper, not a first-principles derivation, so classical self-definitional or uniqueness-import circularity is absent. The load-bearing methodological coupling is evaluation contamination: the reward is defined as VGGT-backed pointwise reprojection error (Eqs. 1–4, with GAS from VGGT attention), while §6.1’s primary “3D reconstruction quality” again runs VGGT and reports PSNR/SSIM/LPIPS of reprojected clouds plus RPT—the paper’s own pointwise metric. Under Best-of-N TTS (Table 1), selecting by Reproj-Pts and then ranking methods on RPT (and highly correlated VGGT re-render scores) makes those particular wins forced rather than independent evidence of geometric robustness. Preference SFT/DPO (Table 2) is weaker but still trains on pairs ranked by the same oracle and re-evaluates with it. Mitigations are real and keep the score moderate: EPI is a more independent multi-view check, VBench is external/perceptual, qualitative examples are not VGGT-scored, and comparisons to Epipolar / Reproj-Pix on shared metrics still have content. No self-citation uniqueness chain, no fitted parameter renamed as a physical prediction, and no ansatz smuggled via author-only theorems. Score 4 reflects partial circularity confined to the shared-oracle geometric metrics, not collapse of the whole claim.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 4 invented entities

The central claim rests on treating a frozen geometric foundation model and a tracker as an external preference oracle, plus standard flow-matching and preference-optimization machinery. Free parameters are mostly engineering knobs (sampling fraction, LoRA, DPO weights, search budget). Invented pieces are the reward formulation, sampling module, preference dataset, and causal search recipes—not new physical entities. The load-bearing domain assumption is VGGT reliability on generated video.

free parameters (5)
  • GAS top-k patch fraction
    Fraction of attention-ranked patches retained for point sampling; chosen by design and not derived; controls which regions enter the reward.
  • LoRA rank r=32 and α=64
    PEFT capacity for SFT/DPO on Wan2.1; hand-chosen adaptation budget that affects how much the policy can move.
  • DPO temperature/scale β and motion penalty weight λ
    Trade off preference strength vs reference model and static-collapse penalty in Flow-DPO; fitted/tuned rather than predicted.
  • TTS budget N / beam (B,K) and SoP context window
    Search width and temporal context size determine reported scaling curves; free inference hyperparameters.
  • Preference construction: 10 seeds per prompt, keep best/worst
    Discrete sampling and extreme-pair selection define the GB3DV preference distribution; ad hoc design choice that magnifies reward differences.
axioms (5)
  • domain assumption VGGT (and similar geometric foundation models) produce sufficiently accurate dense depth and camera parameters on generated video frames for reprojection error to rank geometric quality.
    Invoked throughout §3–4 and evaluation §6.1; without it the reward and several metrics lose meaning.
  • domain assumption A pretrained point tracker yields reliable cross-frame correspondences on the GAS-selected points, so tracker–reprojection disagreement measures multi-view inconsistency.
    §4.2 tracking step before Eq. 4; failures would inject noise into the reward.
  • standard math Rectified-flow / flow-matching video generators and Flow-DPO preference optimization are valid alignment vehicles for the reward.
    §3 and §5.2; standard in the cited literature, not re-derived here.
  • ad hoc to paper Best-vs-worst clips among fixed seeds for the same prompt form valid Bradley–Terry geometric preferences for SFT/DPO.
    §5.1 dataset construction; assumes seed-induced variation is primarily geometric and preference-relevant.
  • ad hoc to paper Shallow global-attention maps of VGGT preferentially highlight geometrically meaningful regions suitable for sampling.
    §4.1 and Fig. 3; empirical design choice supporting GAS.
invented entities (4)
  • Reproj-Pts geometry reward (pointwise cross-frame reprojection error) no independent evidence
    purpose: Scalar multi-view consistency score used as preference/TTS verifier.
    Core proposed reward; defined via Eq. 4 from VGGT geometry + tracker.
  • Geometry-Aware Sampling (GAS) no independent evidence
    purpose: Filter low-texture/non-semantic regions using VGGT attention before error aggregation.
    Claimed robustness mechanism in §4.1; evidence is qualitative heatmaps and ablations only within this paper.
  • GB3DV-25k preference dataset no independent evidence
    purpose: Offline pairs for SFT/DPO ranked by the geometry reward.
    Constructed artifact (§6.1); not an external benchmark with independent labels.
  • Causal TTS path verifiers (Search on Start, Search on Path, Beam Search) no independent evidence
    purpose: Inference-time search over seeds/time using the reward without weight updates.
    Algorithmic recipes in §5.3; empirical scaling in Fig. 4 is paper-internal.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of VIGOR: VIdeo Geometry-Oriented Reward for Temporal Generative Alignment." pith.science (2026). https://pith.science/paper/27XHHWAK

@misc{pith2026260316271,
  author       = {Pith},
  title        = {Pith review of: VIGOR: VIdeo Geometry-Oriented Reward for Temporal Generative Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/27XHHWAK}},
  note         = {Machine review of arXiv:2603.16271}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Video diffusion models lack explicit geometric supervision during training, leading to inconsistency artifacts such as object deformation, spatial drift, and depth violations in generated videos. To address this limitation, we propose a geometry-based reward model that leverages pretrained geometric foundation models to evaluate multi-view consistency through cross-frame reprojection error. Unlike previous geometric metrics that measure inconsistency in pixel space, where pixel intensity may introduce additional noise, our approach conducts error computation in a pointwise fashion, yielding a more physically grounded and robust error metric. Furthermore, we introduce a geometry-aware sampling strategy that filters out low-texture and non-semantic regions, focusing evaluation on geometrically meaningful areas with reliable correspondences to improve robustness. We apply this reward model to align video diffusion models through two complementary pathways: post-training of a bidirectional model via SFT or Reinforcement Learning and inference-time optimization of a Causal Video Model (e.g., Streaming video generator) via test-time scaling with our reward as a path verifier. Experimental results validate the effectiveness of our design, demonstrating that our geometry-based reward provides superior robustness compared to other variants. By enabling efficient inference-time scaling, our method offers a practical solution for enhancing open-source video models without requiring extensive computational resources for retraining.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

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

  1. Geo-Align: Video Generation Alignment via Metric Geometry Reward

    cs.CV 2026-05 unverdicted novelty 7.0

    Geo-Align applies RL with a perceptual reward derived from 3D camera trajectory estimation to improve controllability and fidelity in video generation without paired training data.

  2. GEOPHYS: The Geometry of Physical Plausibility

    cs.CV 2026-06 unverdicted novelty 6.0

    GEOPHYS defines five geometric properties of per-frame embeddings from image encoders that detect physical implausibility in videos with SOTA accuracy and serve as an efficient verifier.

Reference graph

Works this paper leans on

53 extracted references · 43 linked inside Pith · cited by 2 Pith papers

  1. [1]

    In: IEEE Conf

    Asim, M., Wewer, C., Wimmer, T., Schiele, B., Lenssen, J.E.: MEt3R: Measuring multi-view consistency in generated images. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2025), arXiv:2501.06336 3

    arXiv 2025

  2. [2]

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., et al.: Qwen3-vl technical report. Tech. rep. (2025), arXiv:2511.21631 11

    Pith/arXiv arXiv 2025

  3. [3]

    Bratulić, J., Mittal, S., Brox, T., Rupprecht, C.: On geometric understanding and learned data priors in vggt (2025), arXiv:2512.11508 6

    arXiv 2025

  4. [4]

    Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., et al.: Video generation models as world simulators. Tech. rep., OpenAI (2024) 1, 2, 4

    2024

  5. [5]

    Cao,Y.,Wu,F.,Chen,D.Z.,Zhong,Y.,Hong,L.,Xu,D.:Vggt-det:Miningvggtin- ternal priors for sensor-geometry-free multi-view indoor 3d object detection (2026), arXiv:2603.00912 6

    arXiv 2026

  6. [6]

    Chen, B., Monso, D.M., Du, Y., Simchowitz, M., Tedrake, R., Sitzmann, V.: Diffu- sion forcing: Next-token prediction meets full-sequence diffusion. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2024), arXiv:2407.01392 4

    Pith/arXiv arXiv 2024

  7. [7]

    Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. Adv. Neural Inform. Process. Syst. (NeurIPS)36, 9353– 9387 (2023), arXiv:2305.10855 2

    Pith/arXiv arXiv 2023

  8. [8]

    Christiano, P., Leike, J., Brown, T.B., Martic, M., Legg, S., Amodei, D.: Deep reinforcement learning from human preferences. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2023), arXiv:1706.03741 4, 8

    Pith/arXiv arXiv 2023

  9. [9]

    Clark, K., Vicol, P., Swersky, K., Fleet, D.J.: Directly fine-tuning diffusion mod- els on differentiable rewards. In: Int. Conf. Learn. Represent. (ICLR) (2024), arXiv:2309.17400 4

    Pith/arXiv arXiv 2024

  10. [10]

    Dai, Y., Jiang, F., Wang, C., Xu, M., Qi, Y.: Fantasyworld: Geometry-consistent world modeling via unified video and 3d prediction (2025), arXiv:2509.21657 2

    arXiv 2025

  11. [11]

    Du, H., Ye, J., Cong, X., Li, R., Ni, J., et al.: Videogpa: Distilling geometry priors for 3d-consistent video generation (2026), arXiv:2601.23286 3, 4, 8, 11, 12

    Pith/arXiv arXiv 2026

  12. [12]

    Gao, Y., Guo, H., Hoang, T., Huang, W., Jiang, L., et al.: Seedance 1.0: Exploring the boundaries of video generation models (2025), arXiv:2506.09113 3, 4

    Pith/arXiv arXiv 2025

  13. [13]

    Ghosh, D., Hajishirzi, H., Schmidt, L.: Geneval: An object-focused framework for evaluating text-to-image alignment. Adv. Neural Inform. Process. Syst. (NeurIPS) 36, 52132–52152 (2023), arXiv:2310.11513 2

    Pith/arXiv arXiv 2023

  14. [14]

    Google DeepMind: Veo 3 technical report. Tech. rep., Google DeepMind (2025) 1, 2, 4

    2025

  15. [15]

    Gupta, A., Yu, L., Sohn, K., Gu, X., Hahn, M., Li, F.F., Essa, I., Jiang, L., Lezama, J.: Photorealistic video generation with diffusion models. In: Eur. Conf. Comput. Vis. (ECCV) (2023), arXiv:2312.06662 4

    Pith/arXiv arXiv 2023

  16. [16]

    Han, J., Hong, S., Jung, J., Jang, W., An, H., et al.: Emergent outlier view rejection in visual geometry grounded transformers (2025), arXiv:2512.04012 6

    arXiv 2025

  17. [17]

    He, H., Xu, Y., Guo, Y., Wetzstein, G., Dai, B., et al.: CameraCtrl: Enabling camera control for text-to-video generation (2025), arXiv:2404.02101 2

    Pith/arXiv arXiv 2025

  18. [18]

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., Fleet, D.J.: Video diffusion models. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2022), arXiv:2204.03458 4

    Pith/arXiv arXiv 2022

  19. [19]

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: Int. Conf. Learn. Represent. (ICLR) (2021), arXiv:2106.09685 13 VIdeo Geometry-Oriented Reward for Temporal Generative Alignment 17

    Pith/arXiv arXiv 2021

  20. [20]

    Hu, Y., Cheng, C., Yu, S., Guo, X., Wang, H.: VGGT4D: Mining motion cues in visual geometry transformers for 4d scene reconstruction (2025), arXiv:2511.19971 6

    arXiv 2025

  21. [21]

    Huang, T., Zheng, W., Wang, T., Liu, Y., Wang, Z., Wu, J., et al.: Voyager: Long-range and world-consistent video diffusion for explorable 3d scene generation (2025), arXiv:2506.04225 2

    Pith/arXiv arXiv 2025

  22. [22]

    Huang, X., Li, Z., He, G., Zhou, M., Shechtman, E.: Self forcing: Bridging the train-test gap in autoregressive video diffusion (2025), arXiv:2506.08009 4

    Pith/arXiv arXiv 2025

  23. [23]

    In: IEEE Conf

    Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., et al.: VBench: Comprehensive bench- mark suite for video generative models. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024), arXiv:2311.17982 11

    Pith/arXiv arXiv 2024

  24. [24]

    Kong, W., Tian, Q., Zhang, Z., Min, R., Dai, Z., et al.: Hunyuanvideo: A systematic framework for large video generative models (2025), arXiv:2412.03603 1, 4

    Pith/arXiv arXiv 2025

  25. [25]

    Kupyn,O.,Manhardt,F.,Tombari,F.,Rupprecht,C.:Epipolargeometryimproves video generation models (2025), arXiv:2510.21615 3, 4, 8, 11, 12

    Pith/arXiv arXiv 2025

  26. [26]

    Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback (2023), arXiv:2302.12192 4, 6, 8

    Pith/arXiv arXiv 2023

  27. [27]

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with MASt3R. In: Eur. Conf. Comput. Vis. (ECCV) (2024), arXiv:2406.09756 5

    Pith/arXiv arXiv 2024

  28. [28]

    Lin, H., Chen, S., Liew, J., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views (2025), arXiv:2511.10647 5

    Pith/arXiv arXiv 2025

  29. [29]

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow match- ing for generative modeling. In: Int. Conf. Learn. Represent. (ICLR) (2023), arXiv:2210.02747 4

    Pith/arXiv arXiv 2023

  30. [30]

    Liu, J., Liu, G., Liang, J., Yuan, Z., Liu, X., et al.: Improving video generation with human feedback (2025), arXiv:2501.13918 2, 3, 4, 6, 9

    Pith/arXiv arXiv 2025

  31. [31]

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: Int. Conf. Learn. Represent. (ICLR) (2022), arXiv:2209.03003 4

    Pith/arXiv arXiv 2022

  32. [32]

    Lu, Y., Zeng, Y., Li, H., Ouyang, H., Wang, Q., et al.: Reward forcing: Effi- cient streaming video generation with rewarded distribution matching distillation (2025), arXiv:2512.04678 2, 3

    Pith/arXiv arXiv 2025

  33. [33]

    Ma,N.,Tong,S.,Jia,H.,Hu,H.,Su,Y.C.,etal.:Inference-timescalingfordiffusion models beyond scaling denoising steps (2025), arXiv:2501.09732 2, 4, 10

    Pith/arXiv arXiv 2025

  34. [34]

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., et al.: Training lan- guage models to follow instructions with human feedback. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2022), arXiv:2203.02155 4, 8

    Pith/arXiv arXiv 2022

  35. [35]

    In: IEEE Int

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: IEEE Int. Conf. Comput. Vis. (ICCV) (2023), arXiv:2212.09748 4

    Pith/arXiv arXiv 2023

  36. [36]

    Polyak, A., Zohar, A., Brown, A., Tjandra, A., Sinha, A., et al.: Movie gen: A cast of media foundation models (2025), arXiv:2410.13720 1, 2, 4

    Pith/arXiv arXiv 2025

  37. [37]

    Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. In: Adv. Neural Inform. Process. Syst. (NeurIPS) (2024), arXiv:2305.18290 4, 8

    Pith/arXiv arXiv 2024

  38. [38]

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., et al.: DeepSeekMath: Pushing the limitsofmathematicalreasoninginopenlanguagemodels(2024),arXiv:2402.03300 4, 8

    Pith/arXiv arXiv 2024

  39. [39]

    Team, K., Chen, J., Ci, Y., Du, X., Feng, Z., et al.: Kling-omni technical report. Tech. rep. (2025), arXiv:2512.16776 4 18 T. Yin et al

    Pith/arXiv arXiv 2025

  40. [40]

    Verdun, C.M., Oesterling, A., Lakkaraju, H., Calmon, F.P.: Soft best-of-n sampling for model alignment (2025), arXiv:2505.03156 4

    Pith/arXiv arXiv 2025

  41. [41]

    Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., et al.: Wan: Open and advanced large-scale video generative models (2025), arXiv:2503.20314 1, 4, 11, 13

    Pith/arXiv arXiv 2025

  42. [42]

    In: IEEE Conf

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Vi- sual geometry grounded transformer. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2025), arXiv:2503.11651 3, 4, 5, 6, 11

    Pith/arXiv arXiv 2025

  43. [43]

    In: IEEE Conf

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: DUSt3R: Geometric 3d vision made easy. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2024), arXiv:2312.14132 5

    Pith/arXiv arXiv 2024

  44. [44]

    Wang, X., Courant, R., Christie, M., Kalogeiton, V.: Akira: Augmentation kit on rays for optical video generation (2025), arXiv:2412.14158 2

    Pith/arXiv arXiv 2025

  45. [45]

    In: ACM SIGGRAPH (SIG- GRAPH) (2024), arXiv:2312.03641 2

    Wang, Z., Yuan, Z., Wang, X., Chen, T., Xia, M., et al.: MotionCtrl: A unified and flexible motion controller for video generation. In: ACM SIGGRAPH (SIG- GRAPH) (2024), arXiv:2312.03641 2

    Pith/arXiv arXiv 2024

  46. [46]

    In: IEEE Conf

    Weyand, T., Araujo, A., Cao, B., Sim, J.: Google Landmarks Dataset v2 – a large-scale benchmark for instance-level recognition and retrieval. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2020), arXiv:2004.01804 11

    Pith/arXiv arXiv 2020

  47. [47]

    In: IEEE Int

    Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: IEEE Int. Conf. Comput. Vis. (ICCV). pp. 2096–2105 (2023), arXiv:2303.14420 2

    Pith/arXiv arXiv 2096

  48. [48]

    Xing, J., Xia, M., Liu, Y., Zhang, Y., Zhang, Y., He, Y., et al.: Make-your- video: Customized video generation using textual and structural guidance (2023), arXiv:2306.00943 2

    Pith/arXiv arXiv 2023

  49. [49]

    Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., et al.: CogVideoX: Text-to- video diffusion models with an expert transformer. In: Int. Conf. Learn. Represent. (ICLR) (2025), arXiv:2408.06072 1, 4

    Pith/arXiv arXiv 2025

  50. [50]

    In: IEEE Conf

    Yin, T., Zhang, Q., Zhang, R., Freeman, W.T., Durand, F., et al.: From slow bidirectional to fast autoregressive video diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (2025), arXiv:2412.07772 4, 11, 12

    arXiv 2025

  51. [51]

    Zhang, X., Lin, H., Ye, H., Zou, J., Ma, J., et al.: Inference-time scaling of diffusion models through classical search (2025), arXiv:2505.23614 2, 4, 10

    arXiv 2025

  52. [52]

    Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnifica- tion: Learning view synthesis using multiplane images37(4), 65:1–65:12 (2018), arXiv:1805.09817 11

    Pith/arXiv arXiv 2018

  53. [53]

    Zhu, H., Zhao, M., He, G., Su, H., Li, C., Zhu, J.: Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video genera- tion (2026), arXiv:2602.02214 4, 6, 10, 11, 14

    Pith/arXiv arXiv 2026

This paper was first reviewed by grok-4.5 on July 13, 2026.