Pith. sign in

REVIEW 3 major objections 6 minor 61 references

BlenderFusion: 3D-Grounded Visual Editing and Generative Compositing

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read BlenderFusion claims that the most reliable way to get precise, 3D-aware control in visual editing is to perform the edit in a graphics engine (Blender) and let a diffusion model handle photorealistic synthesis; the paper shows this…

desk verdict A strong integration paper that earns real gains from 3D-grounding, but the unqualified 'precise object control' claim is bounded by the 2.5D lift failing on large rotations. read the letter →

arxiv 2506.17450 v2 pith:SYHMSCMX submitted 2025-06-20 cs.CV cs.GR

classification cs.CVcs.GR
keywords visualcompositing3D-groundededitingdiffusionmodelsBlenderobjectliftingsourcemaskingsimulatedjitteringmulti-objectscene
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

BlenderFusion argues that the cleanest way to get precise, controllable visual compositing is to stop asking a generative model to infer 3D edits from text, and instead perform edits explicitly in a 3D graphics engine (Blender) while a diffusion model handles the synthesis. The paper proposes a layering-editing-compositing pipeline: objects are segmented and lifted into editable 3D triangle meshes, edited in Blender to produce a target render, and a dual-stream diffusion compositor trained on video frames fuses that render with the original context into a coherent, photorealistic image. With two training strategies—source masking and simulated object jittering—the model learns to follow object edits while the camera stays fixed, and to ignore regions of the source that have been removed or replaced. If the claim holds, BlenderFusion generalizes from simple training videos to fine-grained multi-object edits, novel asset insertion, and background replacement beyond what previous text- or token-driven 3D editing methods achieve.

What carries the argument

The load-bearing component is the dual-stream diffusion compositor operating on a source stream (original image, its Blender render, camera parameters, object poses) and a target stream (the edited Blender render, camera, poses), connected through shared weights and self-attention, with extra input channels for the render, an instance mask, and Plücker camera embeddings. The two training strategies make it work: source masking randomly blanks objects in the source stream so the model learns to ignore removed or replaced regions, and simulated object jittering reconstructs target frames from a fixed camera while jittering object poses, teaching disentangled object control. These strategies only function because the compositor sees both source and target Blender renders as 3D-grounded control signals, not just text tokens.

What would settle it

Run BlenderFusion on Objectron images with only the 2.5D lift and rotate an object by 150–180 degrees; if the produced composite loses the object's identity or geometry (measured by object-level DINO similarity or human judging the object is wrong), the central claim of precise 3D-grounded control fails in exactly the regime the paper flags as unreliable. A controlled comparison of large-angle rotation with and without image-to-3D meshes would isolate whether the compositor or the mesh lift is the bottleneck.

Watch

Extended reading notes

Core claim

The central claim is that decoupling control from generation—editing in Blender and generating in a diffusion model—yields 3D-grounded control over objects, camera, and background at a granularity that prior single-stream, text-conditioned or token-conditioned methods cannot match. Trained only on video frames of simple object and camera motion, the dual-stream compositor learns to transform a noisy Blender render of the edited scene into a photorealistic image, and in doing so inherits Blender's full editing vocabulary: attribute changes, deformation, duplication, and insertion of objects not seen during training. The paper demonstrates this with quantitative gains over Object 3DIT and Neural Assets on MOVi-E, Objectron, and Waymo, and with human preference results showing the gap widens as tasks become more compositional.

Load-bearing premise

The whole pipeline assumes the Blender render of the edited scene is close enough to the target that the diffusion model only needs to polish it; if the 2.5D mesh drifts badly—as it does for large rotations—the edit signal is wrong and the output fails.

Editorial extensions

If this is right

  • Users can rearrange, recolor, deform, duplicate, or replace multiple objects in a photo, with the diffusion model filling in realistic material, shading, and occlusion details.
  • Simple video data with entangled object and camera motion is enough to learn disentangled control, so the method does not require expensive paired 3D editing datasets.
  • The framework carries over to in-the-wild images from SUN-RGBD, ARKitScenes, and Hypersim without retraining, suggesting it is a general compositing tool rather than a dataset-specific fit.
  • Because edits are performed in a graphics engine, the method inherits any Blender operation, enabling progressive, step-by-step editing that a pure generative model cannot express.
  • Standard video-frame metrics (PSNR, SSIM, LPIPS, FID) understate the gains; the real advantage appears in fine-grained compositional tasks where human evaluation shows BlenderFusion winning the large majority of comparisons.

Reading between the lines

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

  • A natural next step the paper leaves implicit is to extend the compositor to video by adding temporal consistency layers, since the training signal comes from video frames and the framework is already per-frame.
  • The source masking strategy could transfer to other conditional diffusion pipelines as a general way to tell the model which parts of the conditioning context are untrustworthy or obsolete.
  • Replacing the 2.5D lift with a full multi-view reconstruction, as the paper does optionally for complex edits, should close the large-rotation failure mode; a testable prediction is that identity preservation on large-angle rotations would then match small-angle performance.
  • The design choice of serialized 3D bounding boxes as text tokens hints that explicit geometric conditioning may reduce reliance on natural-language prompts for 3D-aware editing, which could generalize to other world-model editing tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. BlenderFusion proposes a three-stage visual compositing pipeline: (i) segmenting input images and lifting foreground objects into editable 3D entities using SAM2 and Depth Pro, (ii) editing these entities in Blender with 3D-grounded control over object pose, geometry, appearance, camera, and background, and (iii) compositing the coarse Blender renderings into a photorealistic output with a dual-stream diffusion model. The diffusion compositor, built on Stable Diffusion v2.1, processes source and target render streams in parallel and is trained on video frames with two additional strategies: source masking and simulated object jittering. Experiments on MOVi-E, Objectron, and Waymo Open Dataset compare BlenderFusion against re-implemented versions of Object 3DIT and Neural Assets, reporting consistent quantitative improvements on standard video-frame prediction and qualitative gains on disentangled object control and fine-grained multi-object editing, plus a small human evaluation.

Significance. If the claims hold, BlenderFusion is a valuable and practical framework that couples the precise, interpretable control of a graphics engine with the synthesis power of a diffusion model. The paper has concrete strengths: consistent quantitative gains over two baselines on three datasets, carefully re-implemented baselines with matched training and inference settings, clear ablations of the architectural and training choices, and an honest failure-case appendix. The central idea of decoupling control from generation is compelling and likely to influence follow-up work. However, the breadth of the headline claim, especially 'precise object control' and generalization to fine-grained multi-object editing, is currently bounded by the default 2.5D reconstruction's known unreliability under large rotations and by the lack of quantitative evaluation on the disentangled-control tasks that motivate the method.

major comments (3)
  1. [§3.1, §4.1, Appendix B.2] The central claim that BlenderFusion provides 'precise object control' under large rotations is not established for the default pipeline. Section 3.1 describes the default layering as back-projecting a single Depth Pro depth map into 2.5D triangle meshes, and Section 4.1 states that test-time layering uses these 2.5D reconstructions without image-to-3D meshes except for Figure 7 (Bottom). Appendix B.2 concedes that 'when the object reconstruction is 2.5D, the renders can be unreliable when the object is rotated significantly' and Figure 9 shows failures at 150 and 180 degrees. Yet Figures 5 and 11 demonstrate rotation control at exactly those angles. The paper should either quantify performance as a function of rotation angle for the default pipeline, restrict the rotation claims to the reliable range, or adopt complete meshes for large-rotation experiments.
  2. [§4.2, Table 2] The quantitative evaluation does not directly measure the fine-grained disentangled control that motivates the method. In the standard video-frame setup, the Blender render R_tgt is produced by applying the annotated object and camera poses to the lifted meshes, so the target render already encodes the exact transformation that defines the ground-truth frame; Table 2 therefore evaluates the compositor's ability to repair pose-conditioned renders, not the user's ability to specify arbitrary edits. Disentangled rotation, translation, scaling, and multi-object composition are supported mainly by qualitative figures and a 54-example user study (Section 4.4, Table 3), which lacks confidence intervals or inter-annotator agreement. Adding quantitative metrics on synthetic edit benchmarks, such as object pose error or identity preservation under known transformations, would materially strengthen the main claim.
  3. [Table 4, §4.4] The load-bearing training contribution for disentangled control is simulated object jittering, but the quantitative ablation shows that it slightly hurts standard video-frame performance relative to source masking alone (object-level PSNR 16.18 vs. 16.06, SSIM 0.393 vs. 0.389), and the paper explains this as expected because jittering is an image-reconstruction setup. No quantitative measure of disentanglement is provided, so the reader cannot determine whether jittering improves object control or merely trades away standard performance. A quantitative disentanglement metric, for example fixed-camera object translation and rotation accuracy versus camera-only novel-view fidelity, is needed to support the core claim that jittering enables disentangled control.
minor comments (6)
  1. [Appendix A.2] There is a typo: 'Simialr' should be 'Similar'.
  2. [Figure 8] The label 'Dual-streamBlender' in the figure appears to be missing a space; it should read 'Dual-stream + Blender'.
  3. [§3.2/§4.1] The source-masking probability (0.5), the training mix ratios, and the classifier-free guidance scale are reported but not ablated; a sentence noting their sensitivity would be helpful.
  4. [Appendix B.2] The terminology is inconsistent: the method section refers to Hunyuan3D v2 meshes, while the appendix calls them '3D-Gen meshes'; unify the naming.
  5. [Table 1] The column header 'Novel-Obj' is ambiguous; consider spelling it out as 'Novel Object Insertion'.
  6. [§4.1] The WOD results rely on initialization from MOVi-E, as stated in the text, but the main narrative could emphasize this dependency more clearly when interpreting the WOD numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BlenderFusion's compositing output is generated by a learned model; R_tgt is a noisy conditioning signal, not the supervision, and the only self-citation is positioning, not load-bearing.

full rationale

BlenderFusion is an empirical system paper, not a formal derivation, and its central claim is supported by generated outputs from a fine-tuned diffusion compositor rather than by an equation that reduces to its own inputs. The target render R_tgt is constructed by transforming a lifted 2.5D mesh using video annotations, but the training and evaluation supervision is the actual video frame I_tgt; the paper explicitly frames R_tgt as noisy and in need of correction: 'Since the reconstructed 3D scene S_src is derived from 2D images, its transformation often introduces noise, leading to artifacts in the target render R_tgt. The compositor corrects these artifacts with learned 3D shape priors.' Thus the output is not R_tgt renamed as a prediction, and the quantitative protocol measures fidelity to held-out target frames, not to the conditioning render. The training strategies (source masking and simulated object jittering) are ablations over conditioning and supervision design, and the ablation table separately reports that the jittering strategy slightly lowers standard video metrics, confirming it is not a fitted quantity masquerading as a result. The only author self-citation is Image Sculpting [57], which shares co-author Saining Xie and is cited in Related Work and Table 1 to position prior single-object Blender-based editing; this citation is not load-bearing for BlenderFusion's central claims. Appendix B.2 concedes that 2.5D reconstructions produce unreliable renders for large rotations and that 3D-Gen meshes resolve most such cases; this is an honest limitation of the default layering choice, not a circular step, because the claimed control is explicitly bounded by render faithfulness and the optional complete-mesh path is disclosed. No fitted parameter is later called a prediction, no uniqueness theorem is imported from the authors' prior work, and no known empirical pattern is merely renamed. The finding is therefore no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's empirical claims rest on off-the-shelf perception models, dataset annotations, and Blender renders as a faithful conditioning signal. There is no derivation, so free parameters are limited to hand-set training and inference choices. No invented physical entities are introduced; the contributions are an architecture and two training strategies.

free parameters (4)
  • source masking probability = 0.5
    Each source object is randomly masked with probability 0.5 during training (Section 3.2); this choice controls how strongly the model relies on source context.
  • training mix ratios = 0.35 / 0.30 / 0.30 / 0.05
    Ratios of vanilla video training, source masking, source masking plus jittering, and unconditional training (Section 4.1); hand-set and central to the disentanglement claims.
  • classifier-free guidance scale = 2.0
    Inference CFG scale used for sampling (Section 4.1); hand-selected.
  • learning rates = 5e-5 for diffusion model, 1e-4 for MLP
    Optimization hyperparameters reported in Section 4.1; chosen by hand.
assumptions (4)
  • domain assumption SAM2, Depth Pro, and Grounding DINO provide sufficiently accurate masks, metric depth, and detections for the lifted meshes.
    Introduced in Section 3.1 Object-centric 3D Layering; the whole pipeline depends on these off-the-shelf perception models.
  • domain assumption Object poses and camera parameters in MOVi-E, Objectron, and Waymo are accurate enough to simulate test-time edits in Blender.
    Used in Section 3.2 and Section 4.1 to construct source-target pairs and render R_tgt from S_src; noisy annotations would make the render conditioning wrong.
  • domain assumption Blender renderings R_src and R_tgt are a reliable 3D control signal, and the compositor can repair reconstruction artifacts while following the render geometry.
    Central to Section 3.1 Generative Compositing; Appendix B.2 admits this fails for 2.5D meshes under large rotation.
  • ad hoc to paper A diffusion model fine-tuned only on simple rigid frame-to-frame transformations generalizes to non-rigid deformation, attribute change, and novel-object insertion at test time.
    Figure 7 demonstrates examples, but there is no quantitative evaluation for these advanced controls and they are outside the training distribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BlenderFusion: 3D-Grounded Visual Editing and Generative Compositing." pith.science (2026). https://pith.science/paper/SYHMSCMX

@misc{pith2026250617450,
  author       = {Pith},
  title        = {Pith review of: BlenderFusion: 3D-Grounded Visual Editing and Generative Compositing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SYHMSCMX}},
  note         = {Machine review of arXiv:2506.17450}
}
read the original abstract

We present BlenderFusion, a generative visual compositing framework that synthesizes new scenes by recomposing objects, camera, and background. It follows a layering-editing-compositing pipeline: (i) segmenting and converting visual inputs into editable 3D entities (layering), (ii) editing them in Blender with 3D-grounded control (editing), and (iii) fusing them into a coherent scene using a generative compositor (compositing). Our generative compositor extends a pre-trained diffusion model to process both the original (source) and edited (target) scenes in parallel. It is fine-tuned on video frames with two key training strategies: (i) source masking, enabling flexible modifications like background replacement; (ii) simulated object jittering, facilitating disentangled control over objects and camera. BlenderFusion significantly outperforms prior methods in complex compositional scene editing tasks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 39 canonical work pages

  1. [1]

    Ahmadyan, L

    A. Ahmadyan, L. Zhang, A. Ablavatski, J. Wei, and M. Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2021

  2. [2]

    Magic Fixup: Streamlining Photo Editing by Watching Dynamic Videos

    H. Alzayer, Z. Xia, X. Zhang, E. Shechtman, J.-B. Huang, and M. Gharbi. Magic fixup: Stream- lining photo editing by watching dynamic videos.arXiv preprint arXiv:2403.13044, 2024

  3. [3]

    Baldridge, J

    J. Baldridge, J. Bauer, M. Bhutani, N. Brichtova, A. Bunner, K. Chan, Y. Chen, S. Dieleman, Y. Du, Z. Eaton-Rosen, et al. Imagen 3.arXiv preprint arXiv:2408.07009, 2024

  4. [4]

    Baruch, Z

    G. Baruch, Z. Chen, A. Dehghan, T. Dimry, Y. Feigin, P. Fu, T. Gebauer, B. Joffe, D. Kurz, A. Schwartz, et al. Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data.arXiv preprint arXiv:2111.08897, 2021

  5. [5]

    Bochkovskii, A

    A. Bochkovskii, A. Delaunoy, H. Germain, M. Santos, Y. Zhou, S. R. Richter, and V. Koltun. Depth pro: Sharp monocular metric depth in less than a second.arXiv preprint arXiv:2410.02073, 2024

  6. [6]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision (ICCV), 2021

  7. [7]

    W. Chen, H. Hu, C. Saharia, and W. W. Cohen. Re-imagen: Retrieval-augmented text-to-image generator.arXiv preprint arXiv:2209.14491, 2022

  8. [8]

    W. Chen, H. Hu, Y. Li, N. Ruiz, X. Jia, M.-W. Chang, and W. W. Cohen. Subject-driven text-to- image generation via apprenticeship learning.Advances in Neural Information Processing Systems (NeurIPS), 2023

Show all 61 references
  1. [9]

    X. Chen, L. Huang, Y. Liu, Y. Shen, D. Zhao, and H. Zhao. Anydoor: Zero-shot object-level image customization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2024

  2. [10]

    Cheng, K

    Y.-C. Cheng, K. K. Singh, J. S. Yoon, A. Schwing, L. Gui, M. Gadelha, P. Guerrero, and N. Zhao. 3d-fixup: Advancing photo editing with 3d priors.arXiv preprint arXiv:2505.10566, 2025. 15 BlenderFusion: 3D-Grounded Visual Editing and Generative Compositing

  3. [11]

    R. Gao, A. Holynski, P. Henzler, A. Brussee, R. Martin-Brualla, P. Srinivasan, J. T. Barron, and B. Poole. Cat3d: Create anything in 3d with multi-view diffusion models.arXiv preprint arXiv:2405.10314, 2024

  4. [12]

    Garibi, S

    D. Garibi, S. Yadin, R. Paiss, O. Tov, S. Zada, A. Ephrat, T. Michaeli, I. Mosseri, and T. Dekel. Tokenverse: Versatile multi-concept personalization in token modulation space.arXiv preprint arXiv:2501.12224, 2025

  5. [13]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets.Advances in neural information processing systems (NeurIPS), 2014

  6. [14]

    Greff, F

    K. Greff, F. Belletti, L. Beyer, C. Doersch, Y. Du, D. Duckworth, D. J. Fleet, D. Gnanapragasam, F. Golemo, C. Herrmann, et al. Kubric: A scalable dataset generator. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2022

  7. [15]

    Y. Gu, I. Huang, J. Je, G. Yang, and L. Guibas. Blendergym: Benchmarking foundational model systems for graphics editing.arXiv preprint arXiv:2504.01786, 2025

  8. [16]

    K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask r-cnn. InProceedings of the IEEE international conference on computer vision (ICCV), 2017

  9. [17]

    Ho and T

    J. Ho and T. Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

  10. [18]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems (NeurIPS), 2020

  11. [19]

    H. Hu, K. C. Chan, Y.-C. Su, W. Chen, Y. Li, K. Sohn, Y. Zhao, X. Ben, B. Gong, W. Cohen, et al. Instruct-imagen: Image generation with multi-modal instruction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2024

  12. [20]

    Z. Hu, A. Iscen, A. Jain, T. Kipf, Y. Yue, D. A. Ross, C. Schmid, and A. Fathi. Scenecraft: An llm agent for synthesizing 3d scenes as blender code. InForty-first International Conference on Machine Learning (ICML), 2024

  13. [21]

    Huang, G

    I. Huang, G. Yang, and L. Guibas. Blenderalchemy: Editing 3d graphics with vision-language models. InEuropean Conference on Computer Vision (ECCV), 2024

  14. [22]

    Huang, Y

    I. Huang, Y. Bao, K. Truong, H. Zhou, C. Schmid, L. Guibas, and A. Fathi. Fireplace: Geometric refinements of llm common sense reasoning for 3d object placement. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025

  15. [23]

    Karras, S

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila. Analyzing and improving the image quality of stylegan. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2020

  16. [24]

    Leroy, Y

    V. Leroy, Y. Cabon, and J. Revaud. Grounding image matching in 3d with mast3r. InEuropean Conference on Computer Vision, 2024

  17. [25]

    R. Liu, R. Wu, B. Van Hoorick, P. Tokmakov, S. Zakharov, and C. Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. InProceedings of the IEEE/CVF international conference on computer vision, 2023. 16 BlenderFusion: 3D-Grounded Visual Editing and Generative Compositing

  18. [26]

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. InEuropean Conference on Computer Vision (ECCV). Springer, 2024

  19. [27]

    S. Lu, Y. Liu, and A. W.-K. Kong. Tf-icon: Diffusion-based training-free cross-domain image composition. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

  20. [28]

    Michel, A

    O. Michel, A. Bhattad, E. VanderBilt, R. Krishna, A. Kembhavi, and T. Gupta. Object 3dit: Language-guided 3d-aware image editing.Advances in Neural Information Processing Systems (NeurIPS), 2024

  21. [29]

    Pandey, P

    K. Pandey, P. Guerrero, M. Gadelha, Y. Hold-Geoffroy, K. Singh, and N. J. Mitra. Diffusion handles enabling 3d edits for diffusion models by lifting activations to 3d. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

  22. [30]

    Radford, L

    A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolu- tional generative adversarial networks.arXiv preprint arXiv:1511.06434, 2015

  23. [31]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), 2021

  24. [32]

    Ramesh, P

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen. Hierarchical text-conditional image generation with clip latents.arXiv preprint arXiv:2204.06125, 2022

  25. [33]

    N. Ravi, V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

  26. [34]

    Roberts, J

    M. Roberts, J. Ramapuram, A. Ranjan, A. Kumar, M. A. Bautista, N. Paczan, R. Webb, and J. M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. InProceedings of the IEEE/CVF international conference on computer vision (ICCV), 2021

  27. [35]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2022

  28. [36]

    N. Ruiz, Y. Li, V. Jampani, Y. Pritch, M. Rubinstein, and K. Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2023

  29. [37]

    Saharia, W

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems (NeurIPS), 2022

  30. [38]

    Sajnani, J

    R. Sajnani, J. Vanbaar, J. Min, K. Katyal, and S. Sridhar. Geodiffuser: Geometry-based image editing with diffusion models. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025

  31. [39]

    Salimans and J

    T. Salimans and J. Ho. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512, 2022

  32. [40]

    Y. Shi, P. Wang, J. Ye, M. Long, K. Li, and X. Yang. Mvdream: Multi-view diffusion for 3d generation.arXiv preprint arXiv:2308.16512, 2023. 17 BlenderFusion: 3D-Grounded Visual Editing and Generative Compositing

  33. [41]

    Sitzmann, S

    V. Sitzmann, S. Rezchikov, B. Freeman, J. Tenenbaum, and F. Durand. Light field networks: Neural scene representations with single-evaluation rendering.Advances in Neural Information Processing Systems (NeurIPS), 2021

  34. [42]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational conference on machine learning (ICML), 2015

  35. [43]

    K. Sohn, N. Ruiz, K. Lee, D. C. Chin, I. Blok, H. Chang, J. Barber, L. Jiang, G. Entis, Y. Li, et al. Styledrop: Text-to-image generation in any style.arXiv preprint arXiv:2306.00983, 2023

  36. [44]

    S. Song, S. P. Lichtenberg, and J. Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. InProceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2015

  37. [45]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

  38. [46]

    Y. Song, Z. Zhang, Z. Lin, S. Cohen, B. Price, J. Zhang, S. Y. Kim, and D. Aliaga. Objectstitch: Generative object compositing.arXiv preprint arXiv:2212.00932, 2022

  39. [47]

    Y. Song, Z. Zhang, Z. Lin, S. Cohen, B. Price, J. Zhang, S. Y. Kim, H. Zhang, W. Xiong, and D.Aliaga. Imprint: Generativeobjectcompositingbylearningidentity-preservingrepresentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

  40. [48]

    P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V. Patnaik, P. Tsui, J. Guo, Y. Zhou, Y. Chai, B. Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2020

  41. [49]

    S. Tang, J. Chen, D. Wang, C. Tang, F. Zhang, Y. Fan, V. Chandra, Y. Furukawa, and R. Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. InEuropean Conference on Computer Vision (ECCV), 2024

  42. [50]

    G. C. Tarrés, Z. Lin, Z. Zhang, H. Zhang, A. Gilbert, J. Collomosse, and S. Y. Kim. Multitwine: Multi-object compositing with text and layout control.arXiv preprint arXiv:2502.05165, 2025

  43. [51]

    Attention is all you need.Advances in neural information processing systems (NeurIPS), 2017

    A.Vaswani, N.Shazeer, N.Parmar, J.Uszkoreit, L.Jones, A.N.Gomez, Ł.Kaiser, andI.Polosukhin. Attention is all you need.Advances in neural information processing systems (NeurIPS), 2017

  44. [52]

    von Platen, S

    P. von Platen, S. Patil, A. Lozhkov, P. Cuenca, N. Lambert, K. Rasul, M. Davaadorj, D. Nair, S. Paul, W. Berman, Y. Xu, S. Liu, and T. Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022

  45. [53]

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny. Vggt: Visual geometry grounded transformer.arXiv preprint arXiv:2503.11651, 2025

  46. [54]

    S. Wang, V. Leroy, Y. Cabon, B. Chidlovskii, and J. Revaud. Dust3r: Geometric 3d vision made easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

  47. [55]

    Z. Wu, Y. Rubanova, R. Kabra, D. Hudson, I. Gilitschenski, Y. Aytar, S. van Steenkiste, K. Allen, and T. Kipf. Neural assets: 3d-aware multi-object scene synthesis with image diffusion models. Advances in Neural Information Processing Systems (NeurIPS), 2024. 18 BlenderFusion:...

  48. [56]

    Xiang, Z

    J. Xiang, Z. Lv, S. Xu, Y. Deng, R. Wang, B. Zhang, D. Chen, X. Tong, and J. Yang. Structured 3d latents for scalable and versatile 3d generation.arXiv preprint arXiv:2412.01506, 2024

  49. [57]

    Yenphraphai, X

    J. Yenphraphai, X. Pan, S. Liu, D. Panozzo, and S. Xie. Image sculpting: Precise object editing with 3d geometry control. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

  50. [58]

    Customnet: Zero-shotobjectcustomization with variable-viewpoints in text-to-image diffusion models.arXiv preprint arXiv:2310.19784, 2023

    Z.Yuan,M.Cao,X.Wang,Z.Qi,C.Yuan,andY.Shan. Customnet: Zero-shotobjectcustomization with variable-viewpoints in text-to-image diffusion models.arXiv preprint arXiv:2310.19784, 2023

  51. [59]

    Zhang, Y

    B. Zhang, Y. Duan, J. Lan, Y. Hong, H. Zhu, W. Wang, and L. Niu. Controlcom: Controllable image composition using diffusion model.arXiv preprint arXiv:2308.10040, 2023

  52. [60]

    Zhang, A

    L. Zhang, A. Rao, and M. Agrawala. Adding conditional control to text-to-image diffusion models. InProceedings of the IEEE/CVF international conference on computer vision (ICCV), 2023

  53. [61]

    source” image and source render are different – because the “source

    Z. Zhao, Z. Lai, Q. Lin, Y. Zhao, H. Liu, S. Yang, Y. Feng, M. Yang, S. Zhang, X. Yang, et al. Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation. arXiv preprint arXiv:2501.12202, 2025. 19 BlenderFusion: 3D-Grounded Visual Editing and Gen...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.