Pith. sign in

REVIEW 3 major objections 4 minor 51 references

DAR — diffusion as renderer — claims that conditioning a pretrained video diffusion model on tracking, world position, and normals (a neural 4D G-buffer) alongside camera rays makes it render animated meshes with both camera and object moti

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

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

T0 review · deepseek-v4-flash

2026-08-04 00:34 UTC pith:SX4CYP5E

load-bearing objection Useful benchmark and a plausible conditioning idea, but the key ablation confounds representation with channel count, so the headline claim outruns the evidence. the 3 major comments →

arxiv 2608.00094 v1 pith:SX4CYP5E submitted 2026-07-30 cs.CV

Video Models as Native 4D Renderers: World-Grounded Conditioning from Animated Mesh

Junhao Chen , Mingjin Chen , Henghaofan Zhang , Minglin Chen , Liaoyuan Fan , Boran Zhang , Saining Zhang , Mingze Sun
show 4 more authors
Hao Zhao Ruqi Huang Zhihao Li Yufei Li
This is my paper
classification cs.CV
keywords 4D generative renderingvideo diffusion modelsworld-position conditioningneural G-bufferPlücker raystrackinganimated meshreference-guided rendering
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.

This paper is trying to establish that a pretrained video diffusion model can serve as a faithful renderer of an animated mesh if the conditioning signal exposes surface identity and scene-coordinate state rather than a camera-dependent depth map. The authors propose DAR, which injects a neural 4D G-buffer — tracking, world position, and normals, projected from the animated mesh — through a widened control adapter together with Plücker rays. The central design claim is that tracking plus world position is a 'visible 4D state code' that separates observer motion from object motion, while depth is a camera chart that mixes them. On a 68-case benchmark, LoRA-finetuned DAR reaches 23.22 PSNR (1.54 dB above an off-the-shelf depth-conditioned Wan2.2), and replacing world position by depth in matched ablations costs 1.26–1.55 dB PSNR at every checkpoint. If right, this points to a practical, camera-independent condition format for controllable generative rendering from animated 4D scenes.

Core claim

DAR's core discovery is that the conditioning representation, not the backbone, is the bottleneck in 4D generative rendering. By extending Wan2.2 camera control from 24 Plücker-ray channels to 33 channels that fuse rays with a mesh-projected tracking/world-position/normal buffer, a single pretrained diffusion model can follow both the target camera and the in-scene animation while re-appearing according to a first-frame reference. The paper argues this works because tracking tells the model which persistent surface element carries appearance and world position tells it where that element is in scene coordinates; depth, by contrast, is a camera-dependent chart in which camera translation, rot

What carries the argument

The load-bearing object is the neural 4D G-buffer B_t(u) = (q(s_t(u)), X̄_t(s_t(u)), n̄_t(s_t(u))): per-pixel tracking (persistent surface identity), scene-normalized world position (current 3D state), and normal (local shape). DAR concatenates these 9 channels with 24 channels of per-pixel Plücker rays into a single widened control adapter injected at the same early point as the original camera control, letting the transformer see both which ray is rendered and which animated surface state lies on it. The paper's theoretical contribution is the 'visible 4D state code' definition and the proposition that depth is a non-injective, camera-dependent chart while (tracking, normalized world posit

Load-bearing premise

The load-bearing premise is that the 1.26–1.55 dB advantage of world position over depth comes from the representation itself rather than from the higher channel count in the full DAR condition; the paper's matched ablation does not control channel count (33 vs 31).

What would settle it

Run the identical depth-swap ablation with the depth variant using 33 channels (e.g., depth plus two auxiliary identity/shape maps of matching capacity) under the same training budget; if the PSNR gap closes, the central representation claim is confounded by channel count. Separately, evaluate on noisy or reconstructed meshes: if DAR's advantage inverts or vanishes, the camera-invariance benefit is fragile to geometry error.

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

If this is right

  • If correct, the same pretrained video diffusion backbone can be reused as a controllable 4D renderer: the user supplies an animated mesh, a camera path, and one reference image, and the model renders the specified state instead of imagining a new one.
  • Depth-conditioned video models that already accept mesh state could swap depth for world position and gain roughly 1.3–1.5 dB PSNR without retraining the backbone, only the adapter.
  • The DAR-4D benchmark gives a reusable tuple — RGB, calibrated cameras, Plücker rays, and per-frame tracking/world-position/normal/depth buffers — so alternative conditioning representations can be compared under matched renderer inputs.
  • The appearance-control result suggests partial decoupling of geometry and appearance: with geometry and camera fixed, changing the reference image re-textures the scene while preserving pose and silhouette, which is useful for look development and previz.
  • Tracked persistent identity plus scene-coordinate state may be a general principle for other video-generation control tasks, since it factors observer motion from scene motion before the model sees the video.

Where Pith is reading between the lines

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

  • The reported advantage of world position over depth may be partly a channel-capacity effect: the full DAR condition uses 33 control channels, while the matching depth-swap uses 31, and the paper never runs an equal-capacity 33-channel depth variant. A direct test would be to pad depth to 33 channels.
  • I expect the world-position advantage to shrink on noisy or reconstructed meshes, since normalized world position depends on accurate scene-level geometry; depth is more forgiving. The paper itself notes real reconstructions will need confidence masks.
  • I would bet that any persistent surface parameterization (UV coordinates, DensePose-style maps, or NOCS-style coordinates) could serve as the identity channel; tracking is one implementation of the 'persistent identity' idea, not the only one.
  • If the throughput issue (1.5–2.5 minutes per clip at 480×832) is addressed, the same conditioning could be extended to long sequences; the paper mentions flicker beyond 200 frames, suggesting temporal smoothing of the G-buffer as a natural next step.

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 / 4 minor

Summary. The paper proposes DAR, a reference-guided 4D generative renderer that extends Wan2.2 video diffusion by injecting a mesh-derived neural 4D G-buffer—tracking, world position, and normal maps—alongside Plücker rays through a widened SimpleAdapter. The central claim is that tracking + world position is a better visible 4D state code than camera-dependent depth for simultaneous control of camera motion and object animation. The authors introduce a DAR-4D benchmark of 68 animated-mesh clips, report LoRA DAR at PSNR 23.22 and full fine-tuning at 25.36, and present a 'matched depth-swap' ablation in which world position is replaced by depth, yielding a +1.26–1.55 dB PSNR advantage for world position across checkpoints. The paper also includes a formulation of 4D generative rendering, an identifiability argument for tracking+world position over depth, and a dataset/benchmark contribution.

Significance. If the central representation claim were cleanly established, this would be a practically useful contribution: world position as an image-format condition is a simple, camera-independent way to expose animated-mesh state to a pretrained video diffusion model, and the paper ships a benchmark designed to compare conditioning representations under a shared training recipe. The geometric argument in Sec. 3.4/Appendix A.1 (depth is a camera chart, world position is not) is sound and correctly identifies an inductive-bias difference. The paper is also transparent about its limitations: OOD-34 is reported as a stress test, the user study is explicitly deferred, and compute/runtime constraints are stated. However, the empirical support for the headline representation claim is weakened by a channel-count confound in the matched depth-swap ablation and by an uncontrolled off-the-shelf baseline comparison, so the causal attribution to 'world position vs depth' is not identified in the current evidence.

major comments (3)
  1. [§5.1, Table 2, Table A.3, Eq. (5)] The 'matched depth-swap' ablation is not channel-matched. DAR uses 24+9=33 control channels; the tracking+normal+depth variant uses 24+7=31 channels (Table A.3). World position is a 3-channel vector, depth is a scalar, so replacing 'only world position by depth' necessarily changes the input dimensionality of the widened SimpleAdapter. The first-layer weights and subsequent feature statistics therefore differ, and the observed +1.26–1.55 dB PSNR advantage cannot be attributed solely to the semantic content of world coordinates. The same issue affects single-channel rows (World position only 27ch vs Depth only 25ch). To support the abstract and Sec. 6.1 claim, an equal-capacity depth condition is needed—e.g., depth stacked or expanded to 3 channels to match the 33-channel input—under the same training budget. The multi-channel rows in Table 2 even show that 'tracking + world position' (30
  2. [§6.1, Table 2, Abstract] The headline improvement over Wan2.2-Depth (+1.54 dB PSNR) compares a fine-tuned DAR model to an off-the-shelf, untuned baseline. Since DAR receives training on the benchmark and the baseline does not, this gain conflates training with representation. The paper acknowledges the baseline receives oracle depth, but the comparison is still not controlled. This is not the central controlled evidence, but the abstract presents the +1.54 dB as a headline improvement; it should be explicitly labeled as a non-matched reference point, with the controlled depth-swap pair carrying the representational claim.
  3. [§6.3, Table 8, Appendix A.4] The OOD-34 stress test appears to reverse the matched depth-swap conclusion: DAR achieves PSNR 19.04 while the tracking+normal+depth variant achieves 20.16, and Wan2.2-Depth achieves 22.25. The paper attributes this to LoRA capacity/data diversity, but it is an acknowledged counterexample to the general claim that world position beats depth. This is a load-bearing limitation for the paper's broader claim of a 'practical 4D rendering condition'; at minimum, the conclusion should be scoped to the 68-case in-distribution benchmark, and the OOD result should be explained more concretely rather than only as a capacity effect. A capacity-matched OOD ablation would help determine whether the reversal is due to representation or optimization.
minor comments (4)
  1. [Abstract and §3.4] The phrase 'matched ablations' is misleading given the channel-count confound; consider replacing with 'controlled ablations' or explicitly stating the channel counts of each condition.
  2. [Table 5] It is unclear whether all rows in Table 5 use the same checkpoint (10k) and the same protocol as Table 2, since the PSNR values vary substantially by trajectory. Please state the checkpoint and alignment protocol in the caption.
  3. [Appendix A.5] The paper honestly states that no user study was run and no synthetic preference numbers are inserted. This is commendable, but the absence of human evaluation means the perceptual-advantage claims rest on frame-aligned metrics alone; this should be stated in the main text, not only in the appendix.
  4. [References] Several technical contributions are cited via same-author preprints dated 2026 (e.g., OVOW, PairCoder++, NeoVerse). Please ensure these are publicly available and clearly distinguish established prior work from concurrent/unpublished work.

Circularity Check

0 steps flagged

No circular derivation: the geometric argument is a standard observation and the empirical target is external Blender ground truth; the unmatched-channel depth-swap is an experimental confound, not a circular step.

full rationale

Walking the claimed derivation chain, no load-bearing step reduces to its own inputs. Proposition 1 is a standard geometric fact: depth is non-injective and camera-dependent, while normalized world position is a scene-coordinate chart. The paper does not derive a numeric rendering-quality prediction from this definition; it makes a qualitative design argument and then tests it empirically. The empirical comparison is against Blender ground-truth frames and off-the-shelf baselines (Wan2.2-Camera, Wan2.2-Depth, DAS, VideoFrom3D), so the target is external to the training objective and not manufactured by the equations. Self-citations, chiefly the OVOW pipeline [Chen et al. 2026g] used to build DAR-4D, are for dataset provenance and are not the load-bearing justification for the representation claim; the benchmark itself is described in this paper and its RGB frames are rendered by Blender, not by the proposed model. The one substantive concern is the 'matched depth-swap' ablation: the paper says it 'replace[s] only world position by depth' (Sec. 3.4), but Table A.3 shows DAR uses 24+9=33 control channels while Tracking+normal+depth uses 24+7=31. Thus the comparison changes input dimensionality, not just the semantic content of the position channel, so the +1.26–1.55 dB advantage could partly reflect extra adapter channels rather than world-position representation. That is a real experimental confound affecting causal attribution, but it is not circularity under the stated rules: it is not a fitted parameter renamed as a prediction, nor an equation forcing the result by construction. The paper also candidly reports OOD-34 where DAR underperforms Wan2.2-Depth and a user-study protocol without synthetic numbers, which is consistent with an honest empirical presentation. Accordingly, the circularity score is 1: essentially self-contained derivation with a minor evaluation-design weakness that belongs in correctness risk, not circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The central claim rests on per-scene world-position normalization, uniqueness of tracking codes, injectivity of the normalization, and the assumption that a widened adapter preserves the Wan2.2 prior. These are domain assumptions, not fitted physical constants. No new physical or latent entity is introduced; the 'neural 4D G-buffer' is a repackaging of existing image-format maps into a control tensor.

free parameters (1)
  • scene world-position normalization bounds = per-scene min/max bounds (exact rule not specified)
    Normalized world position X̄ in Eq. (7) requires per-scene normalization; Appendix A.1 assumes injectivity inside scene bounds, a condition the paper does not verify numerically.
axioms (5)
  • domain assumption The animated mesh surface state X_t(s), normal n_t(s), and tracking identity q(s) are user-specified and accurate, with no noise or camera error.
    Sec. 3.1 defines the rendering problem with exact mesh state; Sec. 7 admits real reconstructions would require confidence masks. The benchmark uses synthetic rendered inputs, so this holds only for the controlled benchmark.
  • domain assumption Tracking color q(s) is unique at the selected surface/part granularity.
    Appendix A.1 states 'Assume the tracking code q is unique at the selected surface/part granularity' — load-bearing for the identifiability corollary.
  • domain assumption The scene normalization for X̄ is injective inside the scene bounds.
    Appendix A.1: 'assume ... the scene normalization for X̄ is injective inside the scene bounds.' If two distinct points map to the same normalized coordinate, the claimed identifiability fails.
  • domain assumption Widening the Wan2.2 camera adapter from 24 to 33 channels preserves the pretrained image-to-video prior while adding geometry conditioning.
    Sec. 3.2 replaces the 24-channel adapter with a 33-channel adapter; the paper assumes the residual injection point keeps the pretrained prior. This is an architectural assumption not formally proven.
  • domain assumption Frame-aligned PSNR/SSIM/LPIPS against Blender ground truth is a valid measure of renderer fidelity.
    Sec. 5.2 defines metrics and leaves tracking IoU, ATE/RTE, FVD to future work. Whether pixel metrics capture camera/object-motion adherence is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 18193 in / 14792 out tokens · 137754 ms · 2026-08-04T00:34:58.613482+00:00 · methodology

0 comments
read the original abstract

Pretrained video diffusion models can act as renderers when the desired scene state is already specified by an animated mesh, a camera trajectory, and a reference image. This 4D generative rendering setting raises a representation question: what image-format condition lets a video backbone obey both camera motion and scene-internal animation? We propose DAR, a reference-guided renderer that extends Wan2.2 camera control from Pl\"ucker rays alone to a joint camera-plus-geometry interface. DAR projects a neural 4D G-buffer (tracking, world position, and normal) from the animated mesh and injects it through a widened control adapter while preserving the pretrained image-to-video prior. The central design choice is the pair of tracking and world position. Tracking identifies the persistent surface element that should carry appearance; world position gives its current scene-coordinate state; normal supplies local shape. Depth plus calibrated rays can recover 3D in principle, but depth is a camera-dependent chart in which camera and object motion are mixed. On the 68-case DAR-4D benchmark, LoRA DAR reaches PSNR 23.22, SSIM 0.895, and LPIPS 0.134, improving over off-the-shelf Wan2.2-Depth by 1.54 dB PSNR; a full fine-tune reaches PSNR 25.36 and SSIM 0.917. Matched ablations show that replacing world position by depth reduces PSNR by 1.26--1.55 dB at every checkpoint, supporting tracking+world-position correspondence as a practical 4D rendering condition.

Figures

Figures reproduced from arXiv: 2608.00094 by Boran Zhang, Hao Zhao, Henghaofan Zhang, Junhao Chen, Liaoyuan Fan, Mingjin Chen, Minglin Chen, Mingze Sun, Ruqi Huang, Saining Zhang, Yufei Li, Zhihao Li.

Figure 1
Figure 1. Figure 1: DAR is a reference-guided 4D renderer: with the animated mesh and camera fixed, changing only the first-frame reference image re-renders the scene’s appearance. The same animated car mesh, camera trajectory, and 4D geometry are held fixed; each column uses a different reference style and each row is a normalized animation time. DAR alters paint style, color, and vehicle appearance while keeping the camera,… view at source ↗
Figure 2
Figure 2. Figure 2: DAR pipeline. An animated 4D mesh, target camera trajectory, reference image, and text prompt are converted into aligned video-diffusion conditions. Plücker rays encode the observer; a mesh projection renderer produces tracking, world-position, and normal maps; a widened SimpleAdapter injects the combined camera-plus-geometry control into the pretrained Wan2.2 video DiT. 3.3 Neural 4D G-buffer For each vis… view at source ↗
Figure 3
Figure 3. Figure 3: 4D rotation gallery. DAR outputs under a horizontal orbit (±60◦ ) around six Blender meshes. Each row shows one scene; columns show the first-frame reference and DAR outputs at 𝑡=0, 𝑇 3 , 2𝑇 3 , 𝑇 −1. DAR follows the target camera trajectory while preserving the animated mesh state and reference appearance across humans, animals, vehicles, mechanical objects, and indoor scenes. 7 Conclusion We introduced 4… view at source ↗
Figure 4
Figure 4. Figure 4: DAR controls both camera trajectory and in-scene 4D mesh animation for generative rendering. Each row shows one OVOW Blender asset under a ±60◦ horizontal orbit. Except for the reference column, all panels show the final frame 𝑡=𝑇 , where the target view is farthest from the first-frame reference. Columns show reference, mesh-derived 9-channel geometry, Wan2.2-Camera, Wan2.2-Depth, Blender GT, and DAR. Wan… view at source ↗
Figure 5
Figure 5. Figure 5: Executed baseline suite across animation time. Three OVOW assets are shown at five normalized time points (𝑡=0/25%/50%/75%/𝑇 ) across Reference, Blender GT, Wan2.2-Camera, Wan2.2-Depth, DAS, VideoFrom3D, and DAR. All videos are sampled by normalized animation time. DAR more consistently preserves both target camera trajectory and animated mesh state. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

51 extracted references · 13 linked inside Pith

  1. [1]

    Greff, Klaus and Belletti, Francois and Beyer, Lucas and Doersch, Carl and Du, Yilun and Duckworth, Daniel and Fleet, David J and Gnanapragasam, Dan and Golemo, Florian and Herrmann, Charles and Kipf, Thomas and Kundu, Abhijit and others , booktitle=

  2. [2]

    Xu, Zhongcong and Zhang, Jianfeng and Liew, Jun Hao and Yan, Hanshu and Liu, Jia-Wei and Zhang, Chenxu and Feng, Jiashi and Shou, Mike Zheng , booktitle=

  3. [3]

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

    Li, Yang and Takehara, Hikari and Taketomi, Takafumi and Zheng, Bo and Nie. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

  4. [4]

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

    ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  5. [5]

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

    Texture Generation on 3D Meshes with Point-UV Diffusion , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

  6. [6]

    arXiv preprint arXiv:2503.20314 , year=

    Wan: Open and advanced large-scale video generative models , author=. arXiv preprint arXiv:2503.20314 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2604.02329 , year=

    Generative World Renderer , author=. arXiv preprint arXiv:2604.02329 , year=

    arXiv

  8. [8]

    arXiv preprint arXiv:2607.20889 , year=

    Engine-Native Editable 3D World Reconstruction with Objects and Lighting , author=. arXiv preprint arXiv:2607.20889 , year=

    Pith/arXiv arXiv

  9. [9]

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

    Generative Rendering: Controllable 4D-Guided Video Generation with 2D Diffusion Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  10. [10]

    arXiv preprint arXiv:2606.31388 , year=

    One Video, One World: Turning Monocular Video into Physical 4D Scenes , author=. arXiv preprint arXiv:2606.31388 , year=

    Pith/arXiv arXiv

  11. [11]

    The Fourteenth International Conference on Learning Representations (ICLR) , year=

    DanceTogether: Generating Interactive Multi-Person Video without Identity Drifting , author=. The Fourteenth International Conference on Learning Representations (ICLR) , year=

  12. [12]

    arXiv preprint arXiv:2503.15147 , year=

    PBR-Inspired Controllable Diffusion for Image Generation , author=. arXiv preprint arXiv:2503.15147 , year=

    arXiv

  13. [13]

    arXiv preprint arXiv:2509.07978 , year=

    One View, Many Worlds: Single-Image to 3D Object Meets Generative Domain Randomization for One-Shot 6D Pose Estimation , author=. arXiv preprint arXiv:2509.07978 , year=

    Pith/arXiv arXiv

  14. [14]

    Zeng, Xianfang and Chen, Xin and Qi, Zhongqi and Liu, Wen and Zhao, Zibo and Wang, Zhibin and Fu, Bin and Liu, Yong and Yu, Gang , booktitle=

  15. [15]

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

    Animator-Centric Skeleton Generation on Objects with Fine-Grained Details , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  16. [16]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=

  17. [17]

    arXiv preprint arXiv:2605.27351 , year=

    Feedforward 3D Editing Learns from Semantic-Part Transformation , author=. arXiv preprint arXiv:2605.27351 , year=

    Pith/arXiv arXiv

  18. [18]

    Deitke, Matt and Liu, Ruoshi and Wallingford, Matthew and Ngo, Huong and Michel, Oscar and Kusupati, Aditya and Fan, Alan and Laforte, Christian and Voleti, Vikram and Gadre, Samir Yitzhak and others , booktitle=

  19. [19]

    ACM Transactions on Graphics (TOG) , volume=

    Functional maps: a flexible representation of maps between shapes , author=. ACM Transactions on Graphics (TOG) , volume=

  20. [20]

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

    HVG-3D: Bridging Real and Simulation Domains for 3D-Conditional Hand-Object Interaction Video Synthesis , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  21. [21]

    arXiv preprint arXiv:2607.01883 , year=

    PairCoder++: Pair Programming as a Universal Paradigm for Verified Code-Driven Multimodal and Structured-Artifact Generation , author=. arXiv preprint arXiv:2607.01883 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2506.01103 , year=

    DeepVerse: 4D Autoregressive Video Generation as a World Model , author=. arXiv preprint arXiv:2506.01103 , year=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2604.09415 , year=

    Physinone: Visual physics learning and reasoning in one suite , author=. arXiv preprint arXiv:2604.09415 , year=

    Pith/arXiv arXiv

  24. [24]

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

    GenesisTex: Adapting Image Denoising Diffusion to Texture Space , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  25. [25]

    SIGGRAPH Asia 2025 Conference Papers , year=

    VideoFrom3D: 3D Scene Video Generation via Complementary Image and Video Diffusion Models , author=. SIGGRAPH Asia 2025 Conference Papers , year=

    2025

  26. [26]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  27. [27]

    arXiv preprint arXiv:2503.09733 , year=

    I2V3D: Controllable image-to-video generation with 3D guidance , author=. arXiv preprint arXiv:2503.09733 , year=

    Pith/arXiv arXiv

  28. [28]

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

    SceneTex: High-Quality Texture Synthesis for Indoor Scenes via Diffusion Priors , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  29. [29]

    arXiv preprint arXiv:2601.00393 , year=

    NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos , author=. arXiv preprint arXiv:2601.00393 , year=

    arXiv

  30. [30]

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

    CAT4D: Create Anything in 4D with Multi-View Video Diffusion Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  31. [31]

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

    Normalized Object Coordinate Space for Category-Level 6D Object Pose and Size Estimation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  32. [32]

    Weng, Fangsheng and Chen, Junhao and Li, Xiang and Qin, Jie and Guo, Hanzhong and Shao, Chunhao and Han, Xiaoguang , booktitle=

  33. [33]

    arXiv preprint arXiv:2602.01661 , year=

    From Frames to Sequences: Temporally Consistent Human-Centric Dense Prediction , author=. arXiv preprint arXiv:2602.01661 , year=

    arXiv

  34. [34]

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

    Recammaster: Camera-controlled generative rendering from a single video , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

  35. [35]

    Wen, Kairun and Huang, Yuzhi and Chen, Runyu and Zheng, Hui and Lin, Yunlong and Pan, Panwang and Li, Chenxin and Cong, Wenyan and Zhang, Jian and Lu, Junbin and others , journal=

  36. [36]

    arXiv preprint arXiv:2412.15050 , year=

    Uni-Renderer: Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion , author=. arXiv preprint arXiv:2412.15050 , year=

    Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2503.14492 , year=

    Cosmos-transfer1: Conditional world generation with adaptive multimodal control , author=. arXiv preprint arXiv:2503.14492 , year=

    Pith/arXiv arXiv

  38. [38]

    Ling, Lu and Sheng, Yichen and Tu, Zhi and Zhao, Wentian and Xin, Cheng and Wan, Kun and Yu, Lantao and Guo, Qianyu and Yu, Zixun and Lu, Yawen and others , booktitle=

  39. [39]

    arXiv preprint arXiv:2501.18590 , year=

    DiffusionRenderer: Neural Inverse and Forward Rendering with Video Diffusion Models , author=. arXiv preprint arXiv:2501.18590 , year=

    Pith/arXiv arXiv

  40. [40]

    Machine Vision and Applications , volume=

    Ultraman: ultra-fast and high-resolution texture generation for 3D human reconstruction from a single image , author=. Machine Vision and Applications , volume=

  41. [41]

    Seedance 1.0: Exploring the Boundaries of Video Generation Models , author=

  42. [42]

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

    DRiVE: Diffusion-based Rigging Empowers Generation of Versatile and Expressive Characters , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  43. [43]

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

    RoomPainter: View-Integrated Diffusion for Consistent Indoor Scene Texturing , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  44. [44]

    Proceedings of the 31st International Conference on Computational Linguistics (COLING) , pages=

    Idea23d: Collaborative lmm agents enable 3d model generation from interleaved multimodal inputs , author=. Proceedings of the 31st International Conference on Computational Linguistics (COLING) , pages=

  45. [45]

    arXiv preprint arXiv:2501.03847 , year=

    Diffusion as Shader: 3D-aware Video Diffusion for Versatile Video Generation Control , author=. arXiv preprint arXiv:2501.03847 , year=

    Pith/arXiv arXiv

  46. [46]

    Guler, Riza Alp and Neverova, Natalia and Kokkinos, Iasonas , booktitle=

  47. [47]

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

    Vace: All-in-one video creation and editing , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

  48. [48]

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

    LottieGPT: Tokenizing Vector Animation for Autoregressive Generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  49. [49]

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

    Text2Tex: Text-driven Texture Synthesis via Diffusion Models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

  50. [50]

    Karaev, Nikita and Rocco, Ignacio and Graham, Benjamin and Neverova, Natalia and Vedaldi, Andrea and Rupprecht, Christian , booktitle=

  51. [51]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages=

    Joint 3D Face Reconstruction and Dense Alignment with Position Map Regression Network , author=. Proceedings of the European Conference on Computer Vision (ECCV) , pages=