Pith. sign in

REVIEW 3 major objections 5 minor 32 references

ViP-Rig: Visual-Prompted Controllable Rigging

T0 review · 3 major / 5 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read 2D skeletal and rigidity sketches can steer frozen rigging models to the particular skeleton and skinning an artist wants, not just a plausible one.

desk verdict Solid systems paper: 2D skeletal/rigidity prompts into frozen UniRig/Puppeteer via gated adapters; tables win under GT-derived prompts, free-form control is qualitative only. read the letter →

arxiv 2607.27982 v1 pith:XWIS2WS4 submitted 2026-07-30 cs.CV

Zihan Qin , Mingze Sun , Yifan Mao , Jialei Xu , Jingfeng Guo , Changrong Hu , Wenbo Zhao , Junjun Jiang
show 1 more author
Xianming Liu
This is my paper
classification cs.CV
keywords characterriggingskeletongenerationskinningweightsvisualpromptinggatedadapterscontrollable3Danimationautoregressivemodels
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

The same 3D character mesh can need very different bones and deformation behavior depending on the animation task, so artists routinely edit an automatic rig by hand. Most learning systems only produce one geometry-driven rig and give little direct control over joint layout or regional stiffness. ViP-Rig lets a user draw or edit a 2D bone sketch and a color-coded rigidity map on a mesh rendering, then injects compact features from those images into frozen pretrained skeleton and skinning networks through gated adapters. Under evaluation that reconstructs target rigs from such prompts, it recovers joints, branches, and skinning weights more accurately than geometry-only baselines on a large articulation dataset and in zero-shot transfer, while qualitative examples show localized prompt-first rigging and iterative edit loops.

What carries the argument

Dense-to-Compact Visual Prompt Encoding plus task-specific gated injection: a frozen vision encoder and Perceiver Resampler compress the 2D prompt into fixed-length tokens; layer-wise gated cross-attention adapters steer autoregressive skeleton decoding, and symmetric pre/post adapters condition both point and joint streams before final skinning matching.

What would settle it

Give artists freehand skeletal and rigidity sketches on held-out meshes (not projections of the ground-truth annotation), measure whether the output joints, hierarchy, and animated deformations match the intended edits better than geometry-only baselines and remain stable under sparse or multi-view-ambiguous drawings.

Watch

Extended reading notes

Core claim

ViP-Rig establishes that compact visual tokens from user 2D skeletal sketches and rigidity maps, injected via zero-initialized gated adapters into frozen pretrained skeleton generators and dual-stream skinning matchers, can explicitly control joint placement, branching, and regional deformation while preserving geometric priors—and that this recovers target skeletons and skinning weights better than geometry-conditioned baselines under prompt-guided evaluation.

Load-bearing premise

That success at recovering a rig from prompts built by projecting the same ground-truth skeleton and entropy-derived rigidity labels is a fair stand-in for real freehand artist control on ambiguous meshes.

Editorial extensions

If this is right

  • Artists can specify different task-specific rigs for one mesh by drawing 2D bone layouts and rigidity regions instead of editing a full 3D hierarchy or painting dense weights.
  • Frozen state-of-the-art skeleton and skinning backbones can be reused for controllable rigging without full retraining, via small gated adapters.
  • Prompt-first rigging extends to unrigged and image-to-3D assets when a single aligned rendering is available.
  • Result-guided loops—edit the projected skeleton or derived rigidity map, regenerate, repeat—become a practical workflow for elaborating or simplifying rigs.
  • Regional rigidity cues can separate adjacent parts that geometry alone would skin similarly (e.g., rigid armor next to a flexible shoulder).

Reading between the lines

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

  • If 2D visual prompts are enough to disambiguate underdetermined rigs, similar compact sketch conditioning may transfer to other underdetermined 3D authoring steps such as part segmentation, blendshape design, or cloth pinning.
  • The gap between annotation-projected benchmark prompts and true freehand sketches is the main open measurement problem; a public freehand-prompt test set would stress-test whether adapter control is truly user-facing.
  • Symmetric conditioning of both sides of a matching function is a general pattern that may apply whenever a user wants to bias pairwise assignments (e.g., correspondence or binding) without rewriting the matcher.
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. ViP-Rig proposes a two-stage framework for controllable character rigging conditioned on user 2D visual prompts: a skeletal sketch for joint layout/branching and a color-coded rigidity map for regional deformation. Compact tokens from a frozen DINOv2 encoder plus stage-specific Perceiver Resamplers are injected into frozen pretrained backbones—layer-wise zero-initialized gated adapters into UniRig’s autoregressive skeleton generator, and symmetric pre-/post-interaction gated adapters into Puppeteer’s point–joint skinning streams—while preserving geometric priors. Training synthesizes prompts by projecting GT skeletons and by discretizing normalized Shannon entropy of GT skinning rows. On Articulation-XL2.0 and zero-shot ModelsResource, under prompt-guided evaluation, the method reports lower skeleton Chamfer distances (Table 1) and better skinning precision/recall/ℓ1 (Table 2) than geometry-only baselines, with ablations (Tables 3–4) supporting full layer-wise fusion and dual-path injection. Qualitative Figure 4 illustrates prompt-first rigging and result-guided editing.

Significance. If the controllability claims hold under realistic user prompts, the work addresses a genuine production gap: automatic riggers produce a single plausible rig, whereas artists need task-specific skeletons and regional skinning without heavy 3D hierarchy editing or per-joint weight painting. Conditioning frozen SOTA backbones via gated adapters is a clean, reusable design, and supporting both prompt-first and iterative edit loops is practically valuable for AI-generated assets. The zero-shot ModelsResource transfer and the ablation evidence for persistent and symmetric injection are concrete technical strengths. Significance is currently tempered by how much of the quantitative gain is attributable to privileged GT-derived prompts rather than free-form control.

major comments (3)
  1. [§4.1, Tables 1–2] §4.1 Evaluation protocol together with §3.1–3.2 Training: benchmark I_skel is built by projecting the ground-truth joints/connections onto a VLM-selected view, and I_skin by discretizing normalized Shannon entropy of the GT skinning rows; Tables 1–2 then score recovery of those same annotations. Geometry-only baselines (RigNet, MagicArticulate, UniRig, Puppeteer) never receive this channel. The reported margins therefore mix genuine adapter control with privileged conditioning on a 2D rendering of the target. Free-form, sparse, or intentionally non-GT prompts appear only qualitatively in Figure 4. The central claim of controllable rigging needs a quantitative protocol that measures fidelity/localization when prompts are user-like (edited, incomplete, multi-view-ambiguous, or deliberately different from GT), and/or baselines given comparable structural side information. Without that, Tabl
  2. [§3.2, Eq. (12), Table 2] §3.2 Training, Eq. (12): I_skin is synthesized by thresholding normalized Shannon entropy of each point’s GT weight distribution into rigid/transitional/flexible. This is a convenient proxy, not an artist-authored rigidity intent, and it is derived from the same W that the model is trained to predict. The paper should justify the three-level discretization (thresholds, sensitivity) and show that control still holds when rigidity maps are painted independently of GT entropy (e.g., user edits in Figure 4 quantified, or held-out artist maps). Otherwise skinning gains in Table 2 may partly reflect reconstructing the entropy field that generated the prompt rather than general regional control.
  3. [Tables 1–2] Tables 1–2 report point estimates only (no standard deviations, multiple seeds, or significance tests) on a 2K Art-XL2.0 split and 270 MR characters. Given modest absolute margins in several cells (e.g., skinning ℓ1 0.335→0.309 vs Puppeteer on Art-XL2.0; precision 87.6%→88.3%), variability matters for the claim of consistent superiority. Please add error bars or multi-run statistics, at least for the main comparisons.
minor comments (5)
  1. [§3.1–3.2, Figure 2] Figure 2 and Eqs. (3)–(4): gate parameterization (per-layer α_ℓ, β_ℓ; four independent skinning adapter copies) is clear, but the number of Perceiver tokens K and feature dimension d are never stated in the main text; please report them for reproducibility.
  2. [§4.1] §4.1 says detailed sampling/augmentation and prompt perturbation settings are in the supplementary material; for a methods-heavy systems paper, a short main-text summary of perturbation magnitudes (landmark jitter, missing branches, color bleed) would help readers judge robustness claims without the supplement.
  3. [§2] Related work cites Animator-Centric Skeleton Generation (Sun et al. 2026) for global bone-density control; a brief quantitative or qualitative comparison against that form of weak structural conditioning would clarify the incremental value of dense 2D sketches.
  4. [§5 Conclusion] Single-view prompts are inherent to the interface; failure modes under heavy self-occlusion or wrong chosen view are not analyzed. A short limitation paragraph or failure-case figure would set expectations.
  5. [Abstract, §1] Minor prose issues: missing spaces in compounds (“Existingautomatic”, “prompt-firstrigging”, “resultingskinningweights”) appear in the abstract and introduction; clean for camera-ready.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circular derivation: GT-projected prompts make the quantitative protocol answer-conditioned, but 2D→3D skeleton and entropy-map→weights are not forced by construction.

  1. other [§3.1 Training; §3.2 Training; §4.1 Evaluation protocol; Tables 1–2]
    "benchmark prompts are constructed from ground-truth annotations using the same projection and rendering procedures employed during training. They specify which valid rig is intended for an otherwise ambiguous input geometry, so the benchmark measures prompt-guided target recovery. Because existing baselines do not accept visual prompts, they retain their standard inputs and serve as geometry-conditioned references."

    Quantitative ‘target recovery’ feeds the model 2D projections (or entropy renderings) of the same GT skeletons/weights it is scored against, while baselines get only geometry. Margins therefore mix genuine prompt control with privileged answer-side conditioning. This is evaluation circularity-of-comparison, not Eq.-by-construction: outputs remain underdetermined from the prompt alone (2D→3D hierarchy; coarse rigidity→full W). Free-form control is only qualitative (Fig. 4).

full rationale

ViP-Rig is an engineering conditioning paper, not a first-principles derivation. Its chain is: render/project user or GT-derived 2D skeletal and rigidity images → Dense-to-Compact tokens (DINOv2 + Perceiver) → gated adapters into frozen UniRig/Puppeteer backbones → autoregressive skeleton tokens or point–joint skinning weights. That is ordinary supervised conditional generation. The only circularity-adjacent issue is evaluation design, not mathematical reduction. Sec. 3.1–3.2 and 4.1 build benchmark I_skel by projecting GT joints/bones and I_skin by discretizing normalized Shannon entropy of GT skinning rows, then score recovery of those same annotations while geometry-only baselines never see the prompt channel. That privileges the proposed method and overstates how much Tables 1–2 isolate “controllable rigging” versus “answer-conditioned reconstruction.” It is not, however, self-definitional or fitted-input-as-prediction in the strong sense: a single-view 2D sketch does not determine 3D joint depths or DFS token sequences, and a three-level rigidity map does not determine per-joint weight vectors—geometry and the frozen priors still do real work, as ablations (Tables 3–4) and free-form/edited qualitative prompts (Fig. 4) indicate. No uniqueness theorem or ansatz is smuggled in via overlapping-author citation; UniRig and Puppeteer are external backbones used as frozen modules. Score 2 for one mild evaluation-conditioning concern that is not load-bearing on the method’s internal logic.

Assumptions & free parameters 6 free parameters · 5 assumptions · 3 invented entities

The central empirical claim rests on pretrained geometric priors (UniRig, Puppeteer, DINOv2, 3DShape2VecSet encoding), on the modeling choice that compact 2D single-view tokens plus gated residual adapters suffice for 3D hierarchy and weight control, and on an evaluation/training pipeline that synthesizes prompts from GT structure and entropy of GT weights. Free knobs include token count K, gate scalars, training hyperparameters, and entropy binning. No new physical entities; invented pieces are architectural modules.

free parameters (6)
  • Number of Perceiver Resampler tokens K and token dim d
    Fixed-length conditioning size chosen by design; controls how much prompt detail survives compression. Exact K not numerically stated in main text.
  • Layer-wise gate scalars α_ℓ, β_ℓ (and four skinning adapter gate sets) = initialized to 0; learned
    Zero-initialized then learned; determine how strongly visual tokens override each frozen layer.
  • AdamW one-cycle peak LR 1e-4, 100 epochs, batch 64, bf16 = 1e-4 peak, 100 epochs, bs 64
    Optimization hyperparameters that affect reported metrics; stated in §4.1.
  • Rigidity entropy discretization thresholds (3 levels) = three levels; edges not specified in main text
    Maps continuous normalized Shannon entropy of GT weights into rigid/transitional/flexible colors for I_skin; bin edges are a design choice affecting skinning supervision.
  • Frozen cosine similarity scale γ in skinning softmax = frozen from Puppeteer
    Inherited from pretrained Puppeteer matching; scales sharpness of predicted weights.
  • Active-influence threshold 1e-4 for precision/recall = 1e-4
    Metric definition choice following Puppeteer; changes P/R numbers.
assumptions (5)
  • domain assumption Frozen UniRig autoregressive skeleton prior and Puppeteer point–joint matching prior remain valid bases that adapters can steer without unfreezing.
    Entire method freezes these backbones (§3.1–3.2); claim of prior-preserving control depends on their quality and compatibility with residual visual injection.
  • domain assumption DINOv2 dense features on a single aligned rendering plus Perceiver compression capture user skeletal/rigidity intent well enough for 3D rig control.
    Dense-to-Compact encoding (§3.1 Eq. 2, §3.2 Eq. 7) is the sole prompt pathway; single-view occlusion is acknowledged via VLM view selection.
  • ad hoc to paper Normalized Shannon entropy of a point's GT skinning distribution is a valid proxy for artist-intended regional rigidity when synthesizing I_skin.
    §3.2 Training defines H_i and discretizes it because datasets lack rigidity maps; this labels the skinning stage's conditioning signal.
  • ad hoc to paper Benchmark prompts projected from GT annotations fairly measure controllable recovery versus geometry-only baselines that cannot take prompts.
    §4.1 Evaluation protocol explicitly builds prompts from GT and compares to unprompted baselines; load-bearing for Tables 1–2 interpretation.
  • standard math Standard next-token CE and row-wise KL objectives are appropriate for skeleton sequences and skinning distributions.
    Eqs. 6 and 13; conventional losses for AR tokens and simplex-valued weights.
invented entities (3)
  • Visual-Guided Skeleton Adapter (per-layer gated cross-attn + gated FFN on Z_skel)
    purpose: Inject skeletal prompt tokens into every frozen UniRig layer while defaulting to identity at init.
    Core control mechanism Eqs. 3–5; architectural module rather than physical entity; design is Flamingo/adapter-like but task-specific.
  • Symmetric pre/post JPCA rigidity adapters (G_pre_P/J, G_post_P/J)
    purpose: Condition both point and joint streams before and after frozen joint–point cross-attention so rigidity modulates compatibility.
    Eqs. 8–11; ablation Table 4 argues necessity; no external falsifiable handle beyond this paper's metrics.
  • Dense-to-Compact Visual Prompt Encoding (shared pattern: frozen DINOv2 + stage-specific Perceiver Resampler)
    purpose: Turn dense 2D prompt images into fixed-length conditioning tokens for both stages.
    Named contribution in intro/method; composition of known components for this task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ViP-Rig: Visual-Prompted Controllable Rigging." pith.science (2026). https://pith.science/paper/XWIS2WS4

@misc{pith2026260727982,
  author       = {Pith},
  title        = {Pith review of: ViP-Rig: Visual-Prompted Controllable Rigging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XWIS2WS4}},
  note         = {Machine review of arXiv:2607.27982}
}
read the original abstract

Rigging is inherently task-dependent because the same mesh may require different skeletons and deformation behaviors across animation tasks. In practice, artists often inspect an initial rig and repeatedly edit its skeletal structure and deformation behavior to meet specific animation requirements. Existing automatic methods primarily generate a plausible rig from geometry, offering limited explicit control over the resulting skeleton and deformation behavior. In this work, we present ViP-Rig, a visual-prompted framework that supports both prompt-first rigging and result-guided editing by injecting features extracted from user-drawn or edited 2D skeletal and rigidity prompts into frozen pretrained backbones. Specifically, ViP-Rig consists of two stages, Skeleton Generation and Skinning Prediction. In the first stage, the skeletal sketch is processed by the Dense-to-Compact Visual Prompt Encoding to produce compact, fixed-length conditioning tokens. The resulting tokens are injected into a frozen pretrained autoregressive generator through gated adapters to control joint placement and branching structure while preserving the generator's geometric prior. In the second stage, the rigidity map is processed using the same visual encoding design, while the pretrained skinning backbone remains frozen. The resulting tokens are symmetrically injected into the point and joint streams to modulate point-joint compatibility and the resulting skinning weights. Experiments on Articulation-XL2.0 and zero-shot evaluation on ModelsResource show that ViP-Rig more accurately recovers target skeletons and skinning weights than geometry-conditioned baselines under prompt-guided evaluation. Qualitative results further demonstrate explicit and localized control in both prompt-first rigging and result-guided editing.

Figures

Figures reproduced from arXiv: 2607.27982 by the authors.

Figure 1
Figure 1. Overview of ViP-Rig. Given an input mesh, a 2D skeletal sketch controls joint layout and branching structure, while a color-coded rigidity map controls regional deformation behavior, enabling different task-specific rigs for the same geometry. prompted framework for controllable rigging of a given mesh, with localized control over skeletal structure and regional skinning in both prompt-first and iterative editing wo… view at source ↗
Figure 2
Figure 2. Architecture of ViP-Rig. Dense visual features are compressed into conditioning tokens and injected into frozen pretrained backbones. Layer-wise gated adapters guide skeleton generation, while symmetric pre- and post-interaction adapters condition point–joint matching for skinning prediction. costly and retains image content unrelated to skeleton con￾trol. We further employ a trainable skeleton-stage Perceiver Resam… view at source ↗
Figure 3
Figure 3. Qualitative comparisons under the benchmark protocol. Top: ViP-Rig better follows prompt-specified joint layouts and branches than geometry-conditioned baselines. Bottom: Skinning-weight predictions and error maps show that rigidity prompts improve local influence assignment in adjacent regions with different deformation behavior [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Controllable rigging applications. Left: Users edit the projected skeleton of an unconditional UniRig result obtained with the visual adapters disabled to elaborate or simplify its structure, or provide a skeletal blueprint to rig an unrigged mesh directly, including m…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 6 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2508.10898 , year=

    Puppeteer: Rig and animate your 3d models , author=. arXiv preprint arXiv:2508.10898 , year=

    arXiv

  2. [2]

    2019 international conference on 3D vision (3DV) , pages=

    Predicting animation skeletons for 3d articulated models via volumetric nets , author=. 2019 international conference on 3D vision (3DV) , pages=. 2019 , organization=

    2019

  3. [3]

    ACM Transactions on graphics (TOG) , volume=

    Automatic rigging and animation of 3d characters , author=. ACM Transactions on graphics (TOG) , volume=. 2007 , publisher=

    2007

  4. [4]

    arXiv preprint arXiv:2005.00559 , year=

    Rignet: Neural rigging for articulated characters , author=. arXiv preprint arXiv:2005.00559 , year=

    arXiv 2005

  5. [5]

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

    Magicarticulate: Make your 3d models articulation-ready , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  6. [6]

    ACM Transactions on Graphics (TOG) , volume=

    One model to rig them all: Diverse skeleton rigging with unirig , author=. ACM Transactions on Graphics (TOG) , volume=. 2025 , publisher=

    2025

  7. [7]

    , author=

    L1-medial skeleton of point cloud. , author=. ACM Trans. Graph. , volume=

  8. [8]

    ACM transactions on graphics (TOG) , volume=

    Skeleton extraction by mesh contraction , author=. ACM transactions on graphics (TOG) , volume=. 2008 , publisher=

    2008

Show all 32 references
  1. [9]

    2010 Shape Modeling International Conference , pages=

    Point cloud skeletons via laplacian based contraction , author=. 2010 Shape Modeling International Conference , pages=. 2010 , organization=

    2010

  2. [10]

    Computer Graphics Forum , volume=

    Mean curvature skeletons , author=. Computer Graphics Forum , volume=. 2012 , organization=

    2012

  3. [11]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Point2skeleton: Learning skeletal representations from point clouds , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  4. [12]

    Computers & Graphics , volume=

    TARig: Adaptive template-aware neural rigging for humanoid characters , author=. Computers & Graphics , volume=. 2023 , publisher=

    2023

  5. [13]

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

    Drive: Diffusion-based rigging empowers generation of versatile and expressive characters , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  6. [14]

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

    Humanrig: Learning automatic rigging for humanoid character in a large scale dataset , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  7. [15]

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

    Make-it-animatable: An efficient framework for authoring animation-ready 3d characters , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  8. [16]

    arXiv preprint arXiv:2506.11430 , year=

    Auto-connect: Connectivity-preserving rigformer with direct preference optimization , author=. arXiv preprint arXiv:2506.11430 , year=

  9. [17]

    ACM Transactions on Graphics (TOG) , volume=

    Voxel cores: Efficient, robust, and provably good approximation of 3d medial axes , author=. ACM Transactions on Graphics (TOG) , volume=. 2018 , publisher=

    2018

  10. [18]

    SIGGRAPH Asia 2022 conference papers , pages=

    Morig: Motion-aware rigging of character meshes from point clouds , author=. SIGGRAPH Asia 2022 conference papers , pages=

    2022

  11. [19]

    arXiv preprint arXiv:2604.20539 , year=

    Animator-Centric Skeleton Generation on Objects with Fine-Grained Details , author=. arXiv preprint arXiv:2604.20539 , year=

    arXiv

  12. [20]

    , author=

    Bounded biharmonic weights for real-time deformation. , author=. ACM Trans. Graph. , volume=

  13. [21]

    Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation , pages=

    Geodesic voxel binding for production character meshes , author=. Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation , pages=

  14. [22]

    ACM Transactions on Graphics (ToG) , volume=

    Neuroskinning: Automatic skin binding for production characters with deep graph networks , author=. ACM Transactions on Graphics (ToG) , volume=. 2019 , publisher=

    2019

  15. [23]

    Proceedings of the ACM on computer graphics and interactive techniques , volume=

    HeterSkinNet: A heterogeneous network for skin weights prediction , author=. Proceedings of the ACM on computer graphics and interactive techniques , volume=. 2021 , organization=

    2021

  16. [24]

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

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

  17. [25]

    ACM Transactions On Graphics (TOG) , volume=

    3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models , author=. ACM Transactions On Graphics (TOG) , volume=. 2023 , publisher=

    2023

  18. [26]

    arXiv preprint arXiv:2205.01068 , year=

    Opt: Open pre-trained transformer language models , author=. arXiv preprint arXiv:2205.01068 , year=

    arXiv

  19. [27]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    arXiv

  20. [28]

    Advances in neural information processing systems , volume=

    Flamingo: a visual language model for few-shot learning , author=. Advances in neural information processing systems , volume=

  21. [29]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    arXiv

  22. [30]

    Proceedings of the 31st ACM International Conference on Multimedia , pages=

    Control3d: Towards controllable text-to-3d generation , author=. Proceedings of the 31st ACM International Conference on Multimedia , pages=

  23. [31]

    Proceedings of the 32nd ACM international conference on multimedia , pages=

    Sketch3D: Style-Consistent Guidance for Sketch-to-3D Generation , author=. Proceedings of the 32nd ACM international conference on multimedia , pages=

  24. [32]

    ACM Transactions on Graphics (TOG) , volume=

    Sketchdream: Sketch-based text-to-3d generation and editing , author=. ACM Transactions on Graphics (TOG) , volume=. 2024 , publisher=

    2024

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.