Pith. sign in

REVIEW 3 major objections 5 minor 58 references

Towards Affordance-Aware Articulation Synthesis for Rigged Objects

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A3Syn synthesizes affordance-aware articulation for arbitrary rigged objects from a scene and a text prompt, with no training data and no rig-topology assumptions.

desk verdict A genuinely new task and a cleanly built pipeline, but the evaluation is too thin to fully support the 'plausible affordance' claim. read the letter →

arxiv 2501.12393 v1 pith:QBWY3L3C submitted 2025-01-21 cs.CV

Yu-Chu Yu , Chieh Hubert Lin , Hsin-Ying Lee , Chaoyang Wang , Yu-Chiang Frank Wang , Ming-Hsuan Yang This is my paper · ORCID
classification cs.CV
keywords riggedobjectsaffordance-awarearticulationlinearblendskinningdiffusioninpaintingsemanticcorrespondencedifferentiablerenderingscenecompositionzero-shotposesynthesis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces A3Syn, a system that takes a rigged 3D object, a scene mesh, an approximate placement point, and a text prompt such as “the dog is climbing the stairs,” and produces the per-bone rotations and global transform that pose the object in that action. The authors aim to show that no task-specific training data or assumptions about rig topology are needed: the affordance signal is drawn from a pretrained 2D inpainting diffusion model, and the pose is optimized directly through differentiable rendering and semantic correspondence. If the method works as claimed, an artist could obtain a physically and semantically plausible first-pass pose in minutes rather than spending hours manipulating tens or hundreds of bones by hand.

What carries the argument

The load-bearing mechanism is the bone-correspondence loss: for each visible vertex, the bone with the largest skinning weight is chosen, per-bone mean pixel positions are computed in the rendered image and in the inpainted reference, and the L2 distance between those means is minimized. This turns a 2D semantic correspondence problem into explicit 3D deformation through differentiable rasterization and linear blend skinning. The multi-view stage adds partial denoising, which starts from a noisy latent at an intermediate step instead of regenerating the image from scratch, and a grid prior that tiles four views into a 2x2 grid so the inpainted references stay cross-view consistent; a thresholded multi-view loss discards outlier views whose correspondence loss is abnormally large.

What would settle it

One concrete test: pick a rig and a scene where the inpainting model is known to generate a reference image that contradicts the prompt (the paper itself reports reverse head-to-tail postures in multi-view inpainting) and check whether A3Syn's final multi-view render reproduces the contradiction rather than the prompt. If the final articulation follows the wrong image rather than the prompt, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that affordance-aware articulation for arbitrary, open-domain rigged objects can be synthesized without retraining, using a single pretrained 2D inpainting model plus geometric optimization. A3Syn first renders the rest-pose object from a good viewpoint, dilates its silhouette, and lets a prompt-conditioned inpainting model generate a reference image of the object already interacting with the scene. A bone-correspondence loss then aligns the average projected position of each visible bone in the current render to its semantically corresponding position in the inpainted reference, using features from a semantic-correspondence network; this loss, combined with a hierarchical bone rotation penalty and an SDF collision-and-contact loss, is minimized end-to-end by gradient descent on the bone rotations and global transform. A second multi-view stage repeats the alignment from several cameras with partial-denoising grid-prior inpainting to resolve depth ambiguity. The paper reports convergence within minutes and better CLIP, non-collision, and contact scores than an SDS baseline.

Load-bearing premise

The whole pipeline assumes the pretrained inpainting model produces a reference image whose pose and text alignment are good enough that the 2D bone-correspondence loss can drive the 3D rig to a plausible pose; if the inpainted pose is wrong or misaligned with the prompt, the optimization faithfully reproduces that wrong pose.

Editorial extensions

If this is right

  • Artists can start from a rest-pose rig and receive a plausible, context-aware first-pass pose in minutes, then make small manual edits instead of posing every bone from scratch.
  • Because the method makes no topological assumptions, the same optimization applies to rigs with extra bones, different joint hierarchies, or unusual artist-created control structures.
  • The two-stage design---single-view coarse placement followed by multi-view fine alignment---resolves single-view depth ambiguity and keeps convergence steady, avoiding the noise and scaling spikes observed with SDS-based pose optimization.
  • The released benchmark and rig-to-PyTorch conversion pipeline make it possible for other rigged objects from the Internet to be tested under the same protocol.
  • The framework needs no task-specific training data, so new scenes and prompts are handled at inference time by the pretrained inpainting model.

Reading between the lines

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

  • Editorial inference: the method's ceiling is set by the inpainting model's text-image alignment; if unconditional or video diffusion models improve, swapping the affordance source could extend A3Syn from static poses to motion, which the authors name as future work.
  • Editorial inference: since the affordance signal flows through a single inpainting model, the method inherits that model's biases; prompts involving rare animals or unusual actions may degrade before the geometric optimization can compensate.
  • Editorial inference: a natural stress test is texture-less or symmetric rigs, where semantic correspondence features have less signal; success there would broaden the method beyond furry four-legged assets.
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 / 5 minor

Summary. The paper proposes A3Syn, a training-free framework that synthesizes articulation parameters for arbitrary rigged objects placed into a scene mesh, given a text prompt describing the desired posture. The method uses a 2D inpainting diffusion model (Flux-Controlnet-Inpainting in the first stage, Flux-Inpainting with partial denoising in the second) to generate affordance reference images, then optimizes per-bone SO(3) rotations and global transforms through differentiable rendering, semantic correspondence (a bone-correspondence loss), an SDF contact/penetration loss, and a hierarchical bone rotation penalty. The pipeline operates in two stages: single-view coarse placement with a VLM-based verification of the inpainted reference, and multi-view fine-grained alignment using a grid prior for cross-view consistency. Experiments on six rigged objects and 20 object-scene pairs compare against an SDS baseline, with ablations on the bone rotation penalty, multi-view alignment, and grid prior.

Significance. If the central claim holds, the paper addresses a practically valuable and previously unstudied task: automatically posing open-domain rigged assets in context without training data or topology assumptions. The method is clearly specified, is topology-agnostic, and the ablations demonstrate that each proposed component contributes to the final result. The authors also contribute a benchmark and a pipeline for converting internet rigs into a programmable format, which is a useful community resource. The main weakness is that the quantitative support is thin and the central 'plausible affordance' claim rests on a diffusion model that the authors themselves describe as having questionable text-image alignment.

major comments (3)
  1. [Sec. 3.2.1, Sec. 3.3, Limitations] The central affordance signal in A3Syn comes entirely from 2D inpainting diffusion models, yet the paper's own Limitations section states that 'inpainting diffusion models have a worse text-image alignment' than unconditional models and questions whether the gap is 'a fundamentally unsolvable issue.' Because the bone-correspondence loss (Eq. 5 and Eq. 8) optimizes the rig so that rendered pixels move toward the inpainted reference, any semantic or physical error in that reference is inherited by the final 3D pose. The VLM verification (Sec. A.2) is applied only in the single-view stage, its rejection rate is not reported, and the multi-view stage uses Flux-Inpainting with partial denoising, filtered only by the loss threshold eps_t (Eq. 8), which cannot detect a semantically wrong but internally consistent reference. The paper needs additional experiments that validate the reference pose quality (for example, human evaluation of the inpainted images, or a comparison of final poses when the affordance signal comes from an unconditional model) and that analyze how reference errors propagate to the final articulation.
  2. [Sec. 4.2, Table 1] The quantitative evaluation is too thin to support the paper's claim that A3Syn synthesizes 'plausible affordance.' The dataset contains only 20 object-scene pairs, no error bars or significance tests are reported, and the CLIP score improvement over the rest-pose baseline is only 0.006 (0.291 to 0.297), which is within typical CLIP noise and may not be perceptually meaningful. In addition, the non-collision and contact scores for the initial rest pose are left blank, so it is not possible to judge whether the physical plausibility gains over simply placing the object are substantial. The authors should report per-pair perforance statistics, error bars, and ideally a human perceptual study or a more sensitive semantic metric.
  3. [Sec. 3.3.2, Sec. 3.3.4, Suppl. A.3] The multi-view loss threshold eps_t in Eq. (8) is used to exclude views with anomalously high bone-correspondence loss, and when all views exceed the threshold the paper says in the supplementary material that 'we simply ignore this round, and re-inpaint a set of new multi-view images.' This introduces a potential selection bias: the reported results may be dominated by cases where the inpainting model happened to produce references that pass the heuristic. The paper should report how often rounds or individual views are rejected, how many re-inpainting attempts are made, and whether the final results change when the threshold is removed or varied.
minor comments (5)
  1. [Sec. 3.3.2, Eq. (8)] The paragraph below Eq. (8) refers to 'Nvalid is the number of LmBC lower than ℓt', but the threshold is defined as eps_t in the equation and in the supplementary material; the symbol ℓt appears to be a typo and should be eps_t.
  2. [Sec. 3.2.2, Eqs. (3)-(5)] The notation for the number of foreground vertices is inconsistent: Eq. (3) and Eq. (4) use N_b_fg while Eq. (5) uses N_b_f; the two should be unified.
  3. [Sec. 3.2.2] The outlier filtering step says points are removed when the feature cosine similarity is below a threshold τ or 'outside a standard deviation', but the threshold τ and the standard deviation are not specified anywhere; this reduces the reproducibility of the correspondence step.
  4. [Sec. 4.2, Table 1] Leaving the non-collision and contact scores blank for the 'Initial' rest-pose placement weakens the comparison, as those metrics are well-defined for any placement; the authors could report the values and note that the rest pose is not placed with physical constraints.
  5. [Sec. 4.1, Fig. 4 caption] The sentence 'Fig. 4 show the qualitative comparison' contains a subject-verb agreement error and should be 'Fig. 4 shows'.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: A3Syn optimizes against an external diffusion prior and is evaluated with independent metrics.

full rationale

The derivation chain is self-contained with respect to its external priors. A3Syn takes scene mesh, prompt, approximate location, and rigged object; it renders the rest pose, calls pretrained inpainting diffusion models (Flux-CN-Inp. in stage one, Flux-Inpainting with partial denoising in stage two) to propose a 2D affordance reference, and optimizes articulation A and global transform T so that differentiable renders of the skinned mesh move toward those reference pixels via the bone-correspondence losses (Eq. 5 and Eq. 8), regularized by SDF (Eq. 7) and rotation penalty (Eq. 6). No equation defines the objective in terms of the optimized output; the target pixels come from an external pretrained model, not from A, T, or the final pose. The evaluation metrics (CLIP, non-collision, contact) are also external to the optimization, so the reported gains are not guaranteed by construction. The self-citations present—semantic correspondence [50], Virtual Pets [4], and text-guided 3D scene composition [51]—are either component tools or motivating observations from the same research group; none is invoked as a uniqueness theorem or as the sole justification of the central claim, so they do not constitute load-bearing circularity. The paper's own Limitations section states that 'inpainting diffusion models have a worse text-image alignment, compared to the unconditional diffusion models,' which is an acknowledged quality risk in the external prior rather than a circularity: an unreliable reference can degrade the pose, but the optimization is still not equivalent to its input. Likewise, hand-tuned hyperparameters and the unreported GPT-4o rejection rate are reproducibility and correctness concerns, not fitted quantities renamed as predictions. Overall, no load-bearing step reduces by definition or by self-citation to its own output.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The framework rests on a stack of pretrained models (inpainting, semantic correspondence, VLM) and a handful of hand-set hyperparameters; no new physical entities are introduced. The main risk is the unverified reliability of each pretrained component, especially the inpainting model that supplies the affordance reference.

free parameters (6)
  • α (bone rotation penalty base) = 1.2
    Eq. (6); chosen empirically to balance naturalness and diversity of articulation; hand-tuned.
  • λRP (bone rotation penalty weight) = 100
    Sec. 3.2.4; set to normalize the scale of this loss relative to the bone correspondence loss.
  • λSDF = λpen = λno-cont = 1000
    Sec. 3.2.4 and Appendix A.4; the authors note the SDF loss is typically 1000x smaller than the correspondence loss, so a large weight is applied.
  • ϵt (multi-view loss threshold) = 1000
    Eq. (8) and Appendix A.3; hand-set threshold used to discard anomalous view losses caused by reverse head-to-tail inpaintings.
  • γ (partial denoising ratio) = 0.8 to 0.6 over 3 rounds
    Sec. 3.3.4 and Appendix A.3; descending schedule tuned to reduce the chance of completely different inpainted references across iterations.
  • Scale learning rate = 1e-5
    Appendix A.4; set much smaller than the 1e-2 used for articulation and translation to prevent unnatural object scaling.
assumptions (5)
  • domain assumption Linear blend skinning is a sufficient deformation model for the rigs
    Sec. 3.1; the pipeline converts artist rigs to LBS in PyTorch and verifies correctness, so rigs that do not fit LBS (or fail conversion verification) are excluded.
  • domain assumption Flux-CN-Inp inpainting model produces affordance-correct images for open-domain objects and scenes
    Sec. 3.2.1; the paper's own Limitations question the text-image alignment of inpainting models, making this premise load-bearing and uncertain.
  • domain assumption Semantic correspondence model [50] gives meaningful dense matches between rendered and inpainted arbitrary categories
    Eq. (1) in Sec. 3.2.2; no fine-tuning is applied per object, so the pretrained features must generalize to cartoon and realistic animals.
  • domain assumption Grid prior with partial denoising keeps multi-view inpainted references 3D-consistent
    Sec. 3.3.1 and Fig. 7; consistency is only approximated in 2D, and the paper still needs a loss threshold to discard contradictory views (Fig. 12).
  • domain assumption GPT-4o VLM accurately judges image-prompt alignment during selection
    Appendix A.2; the VLM output is treated as ground truth for accepting or rejecting inpainted candidates, and no accuracy rate or rejection rate is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Affordance-Aware Articulation Synthesis for Rigged Objects." pith.science (2026). https://pith.science/paper/QBWY3L3C

@misc{pith2026250112393,
  author       = {Pith},
  title        = {Pith review of: Towards Affordance-Aware Articulation Synthesis for Rigged Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBWY3L3C}},
  note         = {Machine review of arXiv:2501.12393}
}
read the original abstract

Rigged objects are commonly used in artist pipelines, as they can flexibly adapt to different scenes and postures. However, articulating the rigs into realistic affordance-aware postures (e.g., following the context, respecting the physics and the personalities of the object) remains time-consuming and heavily relies on human labor from experienced artists. In this paper, we tackle the novel problem and design A3Syn. With a given context, such as the environment mesh and a text prompt of the desired posture, A3Syn synthesizes articulation parameters for arbitrary and open-domain rigged objects obtained from the Internet. The task is incredibly challenging due to the lack of training data, and we do not make any topological assumptions about the open-domain rigs. We propose using 2D inpainting diffusion model and several control techniques to synthesize in-context affordance information. Then, we develop an efficient bone correspondence alignment using a combination of differentiable rendering and semantic correspondence. A3Syn has stable convergence, completes in minutes, and synthesizes plausible affordance on different combinations of in-the-wild object rigs and scenes.

Figures

Figures reproduced from arXiv: 2501.12393 by the authors.

Figure 1
Figure 1. Given arbitrary scene and open-domain rigged objects, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview. (Left) Our A3Syn takes four inputs: The scene geometry, a rigged object, a text prompt describes the desired articula￾tion, and an approximate location to perform the pose. The goal is to solve the object transformation and articulation parameters. (Middle) Our first stage aims to synthesize a course proposal posture, then optimizes the single-view pixel coordinate alignment with the current rest pose. The… view at source ↗
Figure 3
Figure 3. The affordance-aware articulation synthesized with our A3Syn. For each scene-prompt-location composition, we use three different objects to show that our algorithm can adapt to arbitrary open-domain objects, maintain the physical soundness, and be aware of the object semantics (e.g., the rabbit has a different jumping posture, the cat and dog has different tail signatures). Most importantly, the same object adapts t… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Comparisons. SDS has a limited pose change from the rest pose, or creates unnaturally distorted limbs (e.g., the legs of the shiba inu and rabbit). Our method produces more natural pos￾ture, while the added articulation better resembles the affordance. w/o BR Penalty w…
Figure 5
Figure 5. Figure 5: Ablation study. We show a sample of dog attempting to climb tree from two views. Removing bone rotation penalty (BR) causes unnatural limb bending, while omitting our second stage multi-view alignment (MV) leads to floating due to single-view depth ambiguity. Combining…
Figure 6
Figure 6. Figure 6: Our approach provides steady convergence. We visualize the bone rotation in degrees (y axis) to optimization iterations (x axis), each line represents a unique bone. All methods use similar hyperparameters. Our approach (no learning rate decay) has a clear converge dir…
Figure 7
Figure 7. Figure 7: Comparing inpainting strategies. Our strategy with grid prior maintains better cross-view consistency, compared to in￾dividually inpaint each view produces inconsistent limb locations. View 1 View 2 View 3 [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Additional out-of-domain objects. A3Syn can gener￾alize to cartoon characters with unusual appearance and geometry. affordance-aware articulation. Bone rotation penalty. We compare the full version of our method with a variant that removes our proposed Bone Ro￾tation P…
Figure 9
Figure 9. Figure 9: compares the generated quality across various inpainting models. Besides Flux-CN-Inp., DreamShaper￾8-Inpainting (DS-8) is fine-tuned on Stable-Diffusion v1- 5 inpainting model [36]. Stable-Diffusion-2-Inpainting (SD-2) is fine-tuned on Stable-Diffusion 2 [36]. Flux￾Inp…
Figure 10
Figure 10. Figure 10: Automatic VLMs verification process for the in￾painted images. We leverage advanced VLMs to automatically verify the quality of inpainted images and evaluate their alignment with the provided text prompts. In this example, the text prompt is: A Shiba Inu standing in a…
Figure 11
Figure 11. Figure 11: Effectiveness of Iterative Refinement of Multi-view fine-grained alignment. We demonstrate the effectiveness of multi-view fine-grained alignment at each stage. The depth am￾biguity issue is iteratively improved after each iteration of multi￾view alignment. Multi-view…
Figure 12
Figure 12. Figure 12: Visualization of the reverse head-to-tail postures. When the partial denoising rate is higher (e.g., 0.8, in this exam￾ple), the inpainted model sometimes generates objects with reverse head-to-tail postures. In this example, the inpainted images on view 0 and view 3 …
Figure 14
Figure 14. Figure 14: Intermediate steps for multi-view fine-grained alignment stage. Text prompt: A brown rabbit in mid-leap as it jumps down from a wooden chair. In this example, the left hind leg of the rabbit are wrongly placed after the single-view placement stage (initial pose), appe…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 50 canonical work pages

  1. [15]

    Beyond the contact: Discovering comprehensive affordance for 3d objects from pre-trained 2d diffusion models

    Hyeonwoo Kim, Sookwan Han, Patrick Kwon, et al. Beyond the contact: Discovering comprehensive affordance for 3d objects from pre-trained 2d diffusion models. In European Conference on Computer Vision, 2024. 2, 3, 4

    work page 2024

  2. [1]

    Blended latent diffusion

    Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics, 2023. 12

    work page 2023

  3. [2]

    Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B. Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In IEEE Conference on Computer Vision and Pat- tern Recognition, 2024. 1

    work page 2024

  4. [3]

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

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 8

    arXiv 2023

  5. [4]

    Virtual Pets: Animatable animal generation in 3d scenes

    Yen-Chi Cheng, Chieh Hubert Lin, Chaoyang Wang, Yash Kant, Sergey Tulyakov, Alexander G Schwing, Liangyan Gui, and Hsin-Ying Lee. Virtual Pets: Animatable animal generation in 3d scenes. arXiv preprint arXiv:2312.14154,

    arXiv

  6. [5]

    COLLAGE: Collaborative Human-Agent Interaction Generation using Hierarchical Latent Diffusion and Language Models

    Divyanshu Daiya, Damon Conover, and Aniket Bera. Col- lage: Collaborative human-agent interaction generation us- ing hierarchical latent diffusion and language models. arXiv preprint arXiv:2409.20502, 2024. 2

    work page Pith review arXiv 2024

  7. [6]

    Demo2vec: Reasoning object affordances from online videos

    Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In IEEE Conference on Computer Vi- sion and Pattern Recognition, 2018. 2

    work page 2018

  8. [7]

    The theory of affordances

    JJ Gibson. The theory of affordances. Perceiving, acting and knowing: Towards an ecological psychology/Erlbaum, 1977. 2, 6

    work page 1977

Show all 58 references
  1. [8]

    What makes a chair a chair? In IEEE Conference on Computer Vision and Pattern Recognition, 2011

    Helmut Grabner, Juergen Gall, and Luc Van Gool. What makes a chair a chair? In IEEE Conference on Computer Vision and Pattern Recognition, 2011. 2

    2011

  2. [9]

    Proposal flow: Semantic correspondences from ob- ject proposals

    Bumsub Ham, Minsu Cho, Cordelia Schmid, and Jean Ponce. Proposal flow: Semantic correspondences from ob- ject proposals. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 3

    2017

  3. [10]

    Instruct-nerf2nerf: Edit- ing 3d scenes with instructions

    Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Edit- ing 3d scenes with instructions. In IEEE International Con- ference on Computer Vision, 2023. 2

    2023

  4. [11]

    Populating 3d scenes by learning human-scene interaction

    Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes by learning human-scene interaction. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2

    2021

  5. [12]

    Clipscore: A reference-free evaluation met- ric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met- ric for image captioning. arXiv preprint arXiv:2104.08718,

    arXiv

  6. [13]

    Text2room: Extracting textured 3d meshes from 2d text-to-image models

    Lukas H ¨ollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. InIEEE International Conference on Computer Vision, 2023. 1

    2023

  7. [14]

    Skinning: Real-time shape deformation

    Alec Jacobson, Zhigang Deng, Ladislav Kavan, and John P Lewis. Skinning: Real-time shape deformation. In ACM SIGGRAPH Courses, 2014. 3

    2014

  8. [16]

    Eschernet: A generative model for scalable view synthesis

    Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xi- aojuan Qi, and Andrew J Davison. Eschernet: A generative model for scalable view synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2

    2024

  9. [17]

    Physically grounded spatio-temporal object affordances

    Hema S Koppula and Ashutosh Saxena. Physically grounded spatio-temporal object affordances. In European Conference on Computer Vision, 2014. 2

    2014

  10. [18]

    Putting people in their place: Affordance-aware hu- man insertion into scenes

    Sumith Kulal, Tim Brooks, Alex Aiken, Jiajun Wu, Jimei Yang, Jingwan Lu, Alexei A Efros, and Krishna Kumar Singh. Putting people in their place: Affordance-aware hu- man insertion into scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2

    2023

  11. [19]

    Flux, 2024

    Black Forest Labs. Flux, 2024. https://github.com/black- forest-labs/flux. 12

    2024

  12. [20]

    One-shot open affordance learning with foundation models

    Gen Li, Deqing Sun, Laura Sevilla-Lara, and Varun Jampani. One-shot open affordance learning with foundation models. In IEEE Conference on Computer Vision and Pattern Recog- nition, 2024. 2

    2024

  13. [21]

    Genzi: Zero-shot 3d human-scene interaction generation

    Lei Li and Angela Dai. Genzi: Zero-shot 3d human-scene interaction generation. In IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 4, 6, 13

    2024

  14. [22]

    Putting humans in a scene: Learning affordance in 3d indoor environments

    Xueting Li, Sifei Liu, Kihwan Kim, Xiaolong Wang, Ming- Hsuan Yang, and Jan Kautz. Putting humans in a scene: Learning affordance in 3d indoor environments. In IEEE Conference on Computer Vision and Pattern Recognition ,

  15. [23]

    Magic3d: High-resolution text-to-3d content creation

    Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE Conference on Com- puter Vision and Pattern Recognition, 2023. 1

    2023

  16. [24]

    Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models

    Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fi- dler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In IEEE Conference on Computer Vision and Pattern Recog- nition, 2024. 1

    2024

  17. [25]

    Zero-1-to-3: Zero-shot one image to 3d object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tok- makov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3d object. In IEEE International Conference on Computer Vision, 2023. 1

    2023

  18. [26]

    Syncdreamer: Gen- 9 erating multiview-consistent images from a single-view im- age

    Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Gen- 9 erating multiview-consistent images from a single-view im- age. In International Conference on Learning Representa- tions, 2023. 2

    2023

  19. [27]

    Learning affor- dance landscapes for interaction exploration in 3d environ- ments

    Tushar Nagarajan and Kristen Grauman. Learning affor- dance landscapes for interaction exploration in 3d environ- ments. In Neural Information Processing Systems, 2020. 2

    2020

  20. [28]

    Grounded human-object interaction hotspots from video

    Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In IEEE International Conference on Computer Vi- sion, 2019. 2

    2019

  21. [29]

    Chatgpt, 2024

    OpenAI. Chatgpt, 2024. Accessed: 2024-11. 4, 12

    2024

  22. [30]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Neural Information Processing Systems, 2019. 6

    2019

  23. [31]

    Expressive body capture: 3d hands, face, and body from a single image

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In IEEE Conference on Com- puter Vision and Pattern Recognition, 2019. 3

    2019

  24. [32]

    Dreamfusion: Text-to-3d using 2d diffusion

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. In Inter- national Conference on Learning Representations, 2023. 1, 2

    2023

  25. [33]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In International Conference on Machine Learning ,

  26. [34]

    Dreambooth3d: Subject-driven text-to-3d generation

    Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Ben Mildenhall, Nataniel Ruiz, Shiran Zada, Kfir Aberman, Michael Rubenstein, Jonathan Barron, Yuanzhen Li, and Varun Jampani. Dreambooth3d: Subject-driven text-to-3d generation. In IEEE International Conference on Computer...

    2023

  27. [35]

    Accelerating 3d deep learning with pytorch3d

    Nikhila Ravi, Jeremy Reizenstein, David Novotny, Tay- lor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 4

    2007 arXiv

  28. [36]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 12

    2022

  29. [37]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Neural Information Pro- ...

    2022

  30. [38]

    Realmdreamer: Text-driven 3d scene gener- ation with inpainting and depth diffusion

    Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ra- mamoorthi. Realmdreamer: Text-driven 3d scene gener- ation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 1

    2024 arXiv

  31. [39]

    Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022

    Jiaxiang Tang. Stable-dreamfusion: Text-to-3d with stable-diffusion, 2022. https://github.com/ashawkey/stable- dreamfusion. 6

    2022

  32. [40]

    Flux-controlnet-inpainting, 2024

    AlimamaCreative Team. Flux-controlnet-inpainting, 2024. https://github.com/alimama-creative/FLUX-Controlnet- Inpainting. 12

    2024

  33. [41]

    Synthesizing long-term 3d human motion and interaction in 3d scenes

    Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiao- long Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2

    2021

  34. [42]

    Scene- aware generative network for human motion synthesis

    Jingbo Wang, Sijie Yan, Bo Dai, and Dahua Lin. Scene- aware generative network for human motion synthesis. In IEEE Conference on Computer Vision and Pattern Recogni- tion, 2021. 2

    2021

  35. [43]

    Binge watching: Scaling affordance learning from sitcoms

    Xiaolong Wang, Rohit Girdhar, and Abhinav Gupta. Binge watching: Scaling affordance learning from sitcoms. In IEEE Conference on Computer Vision and Pattern Recog- nition, 2017. 2

    2017

  36. [44]

    Humanise: Language-conditioned hu- man motion generation in 3d scenes

    Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned hu- man motion generation in 3d scenes. In Neural Information Processing Systems, 2022. 2

    2022

  37. [45]

    Nerfiller: Completing scenes via generative 3d inpainting

    Ethan Weber, Aleksander Holynski, Varun Jampani, Saurabh Saxena, Noah Snavely, Abhishek Kar, and Angjoo Kanazawa. Nerfiller: Completing scenes via generative 3d inpainting. In IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 5

    2024

  38. [46]

    Unified human-scene interaction via prompted chain-of-contacts

    Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In International Conference on Learning Representations ,

  39. [47]

    Interdreamer: Zero-shot text to 3d dynamic human-object in- teraction

    Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object in- teraction. In Neural Information Processing Systems, 2024. 2, 3, 4

    2024

  40. [48]

    Agent-to-sim: Learning interactive behavior models from casual longitudinal videos

    Gengshan Yang, Bajcsy Andrea, Shunsuke Saito, and Angjoo Kanazawa. Agent-to-sim: Learning interactive behavior models from casual longitudinal videos. arXiv preprint arXiv:2410.16259, 2024. 2

    2024 arXiv

  41. [49]

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

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiao- han Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 8

    2024 arXiv

  42. [50]

    Telling left from right: Identifying geometry-aware semantic corre- spondence

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic corre- spondence. In IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3

    2024

  43. [51]

    Towards text-guided 3d scene composition

    Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou, Sergey Tulyakov, and Hsin-Ying Lee. Towards text-guided 3d scene composition. In IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2, 7

    2024

  44. [52]

    Place: Proximity learning of articulation and con- tact in 3d environments

    Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. Place: Proximity learning of articulation and con- tact in 3d environments. In International Conference on 3D Vision, 2020. 2 10

    2020

  45. [53]

    Generating 3d people in scenes with- out people

    Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes with- out people. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2, 6

    2020

  46. [54]

    Compositional human-scene interaction synthe- sis with semantic control

    Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthe- sis with semantic control. In European Conference on Com- puter Vision, 2022. 6

    2022

  47. [55]

    HIFA: High- fidelity text-to-3d generation with advanced diffusion guid- ance

    Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. HIFA: High- fidelity text-to-3d generation with advanced diffusion guid- ance. In International Conference on Learning Representa- tions, 2024. 6, 8

    2024

  48. [56]

    DreamHOI: Subject-driven generation of 3d human- object interactions with diffusion priors

    Thomas Hanwen Zhu, Ruining Li, and Tomas Jakab. DreamHOI: Subject-driven generation of 3d human- object interactions with diffusion priors. arXiv preprint arXiv:2409.08278, 2024. 2, 3, 4, 7

    2024 arXiv

  49. [57]

    Reasoning about object affordances in a knowledge base representation

    Yuke Zhu, Alireza Fathi, and Li Fei-Fei. Reasoning about object affordances in a knowledge base representation. In European Conference on Computer Vision, 2014. 2

    2014

  50. [58]

    < PROMPT>

    Yixin Zhu, Chenfanfu Jiang, Yibiao Zhao, Demetri Ter- zopoulos, and Song-Chun Zhu. Inferring forces and learning human utilities from videos. In IEEE Conference on Com- puter Vision and Pattern Recognition, 2016. 2 11 Towards Affordance-Aware Articulation Synthesis for Rigged ...

    2016

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.