PaperlayerView exact revision on arXiv

Source-linked AI summary

Make-It-Poseable: Feed-forward Latent Posing Model for 3D Characters

Zhiyang Guo, Ori Zhang, Jax Xiang, Alan Zhao, Zhenxun Yuan, Wengang Zhou, Houqiang Li

arXiv:2512.16767v3cs.CV

TL;DR

Posing flawed 3D assets is challenging for methods constrained by fixed topology and skinning, while pose-conditioned generation can struggle with pose conformance. Make-It-Poseable transforms shape latents using dense skeletal context and adaptive completion, and reports improved posing quality with promising zero-shot generalization on tested morphologies.

  • Problem

    Existing posing methods struggle with fixed mesh topology and pose conformance, limiting work on assets with flawed structures such as fused geometry.

    Source: §1 · pp. 1–2

  • Method

    Make-It-Poseable predicts target shape latents from shape and dense skeletal context, with anchor-guided completion trained using bipartite-matched latent loss.

    Source: §1 · p. 2, §3.5 · p. 5

  • Results

    The method significantly outperforms existing approaches in posing quality and shows promising zero-shot generalization to tested unseen morphologies, including quadrupeds.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The framework supports 3D authoring applications including body-part replacement and refinement.

    Source: §1 · p. 2

    Outputs are reconstructed per pose and do not preserve input mesh connectivity or guarantee temporal consistency, so the method is positioned as a static pre-rigging geometry refiner rather than a production animation system.

    Source: §5 · p. 9

Abstract

from arXiv · show

Posing 3D characters is a fundamental task in computer graphics. However, existing paradigms, ranging from traditional auto-rigging to recent pose-conditioned generative models, frequently struggle with inaccurate skinning weights, fixed mesh topologies, and poor pose conformance. These challenges have become particularly pronounced with the recent explosion of AI-generated 3D assets, which often exhibit flawed structures and fused geometry. To address these issues, we introduce \textbf{Make-It-Poseable}, a novel feed-forward framework that reformulates character posing as a skinning-free latent-space transformation problem. By decoupling shape deformation from the constraints of fixed mesh connectivity, our method directly operates on compact latent representations to reconstruct characters in target poses. To achieve this, our framework integrates a latent posing transformer for shape manipulation, a dense pose representation for fine-grained control, and an adaptive completion module optimized via a bipartite-matched latent loss to robustly handle topological changes. Extensive experiments demonstrate that our method significantly outperforms existing baselines in posing quality. Furthermore, our design shows promising generalization to diverse morphologies such as quadrupeds in our qualitative tests, and supports various 3D authoring applications such as part replacement and refinement.

1 Introduction

Make-It-Poseable reframes character posing as a feed-forward transformation of shape latents, addressing skinning and fixed-topology constraints that hinder posing flawed 3D assets. Dense skeletal context and adaptive completion support fine-detail control and newly exposed geometry; experiments report improved posing quality and promising generalization and authoring uses.

1 Introduction

  • Traditional rigging and skinning can be labor-intensive and suffer weight-leakage artifacts, while fixed-topology methods cannot resolve fused geometry or reveal newly exposed surfaces.Pose-conditioned generators can accommodate topology shifts but often use slow iterative denoising and sparse pose signals, with insufficient pose conformance and difficulty preserving character identity.Source: §1 · pp. 1–2
  • Make-It-Poseable directly transforms compact shape latents to reconstruct a character in its target pose, avoiding vertex deformation and explicit skinning weights.The paper describes it as the first feed-forward 3D character posing model operating on a native 3D latent representation.Source: §1 · p. 2
  • Dense, spatially aligned skeletal context guides shape latents, supporting precise deformation control and direct latent supervision that prevents high-frequency detail loss.This canonical transformation path addresses spatial alignment ambiguity in unordered VecSets.Source: §1 · p. 2
  • An anchor-guided adaptive completion module predicts spatial anchors for newly exposed regions and uses them to guide supplementary shape-latent generation.The module is fine-tuned with a bipartite-matched latent loss after initial latent-loss training; predicted and ground-truth anchors are aligned using the Hungarian algorithm.Source: §1 · p. 2
  • The method significantly outperforms existing approaches in posing quality and shows promising zero-shot generalization to tested unseen morphologies, including quadrupeds.It also supports body-part rescaling, segmentation, replacement, and refinement; the framework was trained exclusively on bipedal data.Source: §1 · p. 2

2 Related Works

Related work spans latent shape editing, rigging and generation approaches to posing, and feed-forward 3D models. The paper positions latent posing as a way to address topology, pose-control, and detail-preservation challenges that remain in these lines of work.

2.1 Latent 3D Generation and Editing

  • Latent 3D representations support composition and refinement, but Neural Generalized Cylinder cannot separate fused artifacts or synthesize newly exposed structures.The cited literature includes native 3D VecSets and sparse voxel grids, as well as latent-space editing and interactive exploration.Source: §2.1 · p. 2

2.2 Character Posing

  • Rigging and skinning remain prevalent, but fixed topology, complex input poses, and weight leakage can produce severe deformation artifacts.The framework pipeline encodes a source character and source and target skeletons, predicts target shape latents, then decodes the posed mesh; its second training stage addresses newly exposed structures.Source: §2.2 · pp. 2–3, Figure 2 · p. 3
  • Pose-conditioned generation can struggle with precise, generalizable control, while human-template methods restrict generalization to highly stylized characters.The literature also includes unified geometry, pose, and appearance generation and cross-topology motion transfer; current pose-conditioned models are described as balancing detail preservation and pose conformance imperfectly.Source: §2.2 · p. 3

2.3 Feed-forward Paradigms in 3D

  • Although feed-forward models have advanced static 3D reconstruction, dense geometry estimation, and novel-view synthesis, applying pose understanding and deformation control in native 3D feed-forward architectures remains open.HumanRAM is described as an early attempt to add animation capabilities to an LRM in the 2D image domain.Source: §2.3 · p. 3

3 Method

Make-It-Poseable poses characters by transforming compact shape latents rather than deforming mesh vertices, using dense skeletal context, latent supervision, and adaptive completion for newly exposed geometry.

3 Method

  • The framework reformulates character posing as a latent-space transformation and trains latent-space supervision before fine-tuning adaptive completion for newly exposed geometry.The goal is to retain fine geometric detail and shape plausibility while re-posing a mesh.Source: §3 · p. 3

3.1 Preliminaries

  • A 3D VAE encodes point clouds as VecSet latents, while an aligned skeleton prior guides latent transformations and the decoder reconstructs meshes from SDFs.MIA is retrained to align its latent space with the higher-capacity shape VAE.Source: §3.1 · pp. 3–4

3.2 Skeleton Encoder

  • The skeleton encoder projects points onto associated bones and shares subsampling with the shape encoder to produce one-to-one pose and shape latent correspondence.Target coordinates are bonewise transformed to preserve alignment; shuffled pairing embeddings discourage overfitting to specific bone semantics.Source: §3.2 · p. 4

3.3 Latent Posing Transformer

  • A decoder-only transformer processes pose-aware source and target tokens in one sequence, predicting target shape latents in a single forward pass.The sampled target latents are decoded through an SDF to reconstruct the posed mesh.Source: §3.3 · p. 4

3.4 Training with Latent-Space Supervision

  • Unordered latent tokens make direct SDF supervision ambiguous and can produce over-smoothed mean shapes, so latent supervision establishes a semantically aligned transformation path.Ground-truth target latents are encoded using source query anchors transformed by the ground-truth deformation.Source: §3.4 · pp. 4–5

3.5 Fine-tuning with Adaptive Completion

  • When posing reveals structures absent from flawed source geometry, adaptive completion predicts spatial anchors and converts them into supplementary target shape latents.Hungarian matching aligns predicted and ground-truth anchors for a latent L1 loss plus anchor-coordinate MSE; training pairs fuse source surfaces by remeshing, and fine-tuning freezes the main transformer.Source: §3.5 · pp. 5–6

4 Experiments

The experiments compare posing quality against space-warping, rigging, and generative baselines, then assess generalization, authoring applications, and key design choices. Results include qualitative tests on unseen morphologies and ablations of the model’s components.

4.1 Experimental Settings

  • Evaluation uses Mixamo and HumanRig humanoids, sampled source-target poses, four mesh-fidelity metrics, and inference time; baselines span space-warping, rigging, and pose-conditioned generation.The baseline comparison uses a 30-pose subset selected for proximal limbs and severe self-contact, while ablations use the full 1k-pose test set.Source: §4.1 · p. 6, Table 1 · p. 6, Table 2 · p. 8

4.2 Comparison Results

  • Our method significantly outperforms space-warping, rigging, and generative baselines in mesh fidelity on a 30-pose stress-test subset.Qualitative comparisons report improved pose conformance and detail preservation over the generative baseline; the method also avoids artifacts seen in rigging and space-warping approaches.Source: Table 1 · p. 6, §4.2 · p. 6, Figure 4 · p. 7

4.3 Generalization Capability

  • Trained only on bipedal humanoids, the method produces plausible poses for exaggerated proportions, novel appendages, and quadrupeds, while severe skeletal extraction errors can still cause failure.The authors report tolerance to moderate skeletal errors, but not to severe errors such as extracorporeal bones.Source: §4.3 · pp. 6–7, Figure 5 · p. 7

4.4 Applications

  • The method supports pre-rigging refinement and direct mesh sequences, plus zero-shot part rescaling, segmentation, replacement, and refinement through target-skeleton edits.The authors note that the generated animation sequences are not intended for production-level animation.Source: §4.4 · pp. 7–8, Figure 6 · p. 8

4.5 Ablation Study

  • Ablations show that cylinder-based pose encoding resolves rotational ambiguity, per-latent embeddings aid correspondence, and adaptive completion with bipartite-matched latent loss improves newly exposed regions.Geometry-space SDF supervision produces over-smoothed shapes; without adaptive completion, newly exposed regions can be malformed or missing.Source: §4.5 · pp. 8–9, Table 2 · p. 8, Figure 7 · p. 8

5 Conclusion and Discussion

Make-It-Poseable frames posing as latent-space transformation, with experiments supporting its posing quality and authoring potential. The discussion notes that latent resolution limits detail and that per-pose reconstruction does not preserve mesh structure or guarantee temporal consistency.

5 Conclusion and Discussion

  • The method is presented as a latent-space posing framework with promising generalization and potential for diverse 3D authoring applications.Its stated limitations include resolution-limited geometric fidelity, no native appearance preservation, and no preservation of input mesh connectivity or temporal consistency; it is positioned as a static pre-rigging geometry refiner.Source: §5 · p. 9

A.1 Shape VAE

  • The shape VAE encodes surface samples into 4096 shape latents of 64 dimensions and decodes them into signed distance function values.The encoder downsamples 32768 sampled points to 4096; the pretrained VAE is frozen during training.Source: §A.1 · p. 11

A.2 Skeleton Encoder

  • The skeleton encoder projects surface points onto bone cylinders and reuses shape-encoder sampling indices to align skeleton and shape latents spatially.Each cylinder radius is set heuristically to 25% of its bone length.Source: §A.2 · p. 11

A.3 Latent Posing Transformer

  • The latent posing transformer uses 24 blocks, a hidden token dimension of 512, and 8 attention heads, processing concatenated source and target tokens projected from source shape/skeleton and target skeleton latents.Source: §A.3 · p. 11

A.4 Adaptive Completion Module

  • The adaptive completion module predicts spatial anchors for newly exposed geometry, then uses them to generate supplementary shape latents.Its encoder-decoder uses pooled contextual features to predict 512 anchors, which are embedded as decoder queries.Source: §A.4 · p. 11

A.5 Re-trained Skeleton Extractor

  • The re-trained skeleton extractor uses the posing model’s frozen shape VAE, allowing direct operation on the same 4096 shape latents.Only decoder attention layers, learnable query tokens, and prediction heads are fine-tuned.Source: §A.5 · p. 11

A.6 Training

Training proceeds in two stages, with a bipartite-matched latent loss used to fine-tune adaptive completion. End-to-end posing takes about one second, substantially less than the reported generative and per-case overfitting baselines.

A.6 Training

  • The first training stage runs for 5 epochs using an L1 loss between predicted and ground-truth target shape latents aligned by dense skeleton correspondence.The subsequent 3-epoch stage freezes the main transformer and encoders/projectors while training adaptive completion with a Hungarian-matched loss combining anchor MSE and latent L1.Source: §A.6 · p. 11
  • Training uses AdamW with a learning rate warmed up to 1 × 10^-4 and cosine-decayed to 1 × 10^-5, taking approximately 120 hours on 4 NVIDIA H200 GPUs.The runtime comparison reports per-case inference on a single NVIDIA H200 GPU with batch size 1.Source: §A.6 · pp. 11–12, §A.7 · p. 12

A.7 Efficiency

  • ~1 second end-to-end posing time is comparable to feed-forward auto-rigging, versus HY3D-Omni at ~17s and NGC at several minutes.The latent posing model itself takes approximately 0.6 seconds per case, excluding skeleton extraction.Source: §A.7 · p. 12

B.1 Data

The appendix describes the data, evaluation procedures, baselines, and authoring workflows used in the study. Its applications include replacing and refining body parts through skeleton edits and fine-tuning.

B.1 Data

  • Training pairs are sampled from 10k humanoids and 20k Mixamo animation sequences; baseline comparison uses a preselected 30-pose stress-test subset focused on close limbs and severe self-contact.Alpha-wrapping remeshing creates fused or self-contacting input surfaces, while the original meshes serve as ground truth.Source: §B.1 · p. 12

B.2 Baselines

  • NGC is evaluated with per-case SDF overfitting, Puppeteer is qualitative-only because its non-semantic bones prevent precise quantitative pose alignment, and HY3D-Omni’s comparison is not perfectly fair.The appendix also describes MIA as an auto-rigging baseline that predicts skeletons and skinning weights.Source: §B.2 · p. 12

B.3 Identity Preservation Evaluation

  • Identity preservation is measured as DINOv2 descriptor cosine similarity between source and posed renders, averaged across three viewpoints and samples in the Table 1 subset.The renders use a frontal view and two side views at ±45° azimuth.Source: §B.3 · p. 12

B.4 Applications

  • Part replacement requires repositioning hand bones to match new hand geometry and aligning them with the arm bones; refinement fine-tunes the model with the selected part’s source latents zeroed.The described refinement experiment reconstructs articulated fingers under skeleton guidance.Source: §B.4 · pp. 12–13

C.1 Additional Ablation Studies

The additional ablations compare skeleton and shape encodings and test adaptive completion choices. The visualization describes how fine-tuning with a bipartite-matched latent loss addresses geometry missing after remeshing and re-posing.

C.1 Additional Ablation Studies

  • Per-latent embeddings outperform alternatives using predicted weights or displacement vectors and preserve local geometric detail without relying on skinning-weight quality.Ground-truth blend-weight embeddings score comparably, but are impractical for in-the-wild inputs.Source: §C.1 · p. 13
  • Mean-only and sampled shape latents are insufficient for effective pose transformation; both often produce fragmented surfaces, and marching cubes can fail to extract a valid mesh.The model uses both the VAE mean and variance vectors to represent shape.Source: §C.1 · p. 13
  • Without adaptive completion latents, geometry-space SDF fine-tuning collapses in performance, while latent-loss fine-tuning without new latents yields only marginal improvement.Injecting adaptive tokens after the posing transformer is more effective than injecting them before it.Source: §C.1 · p. 13

C.2 Visualization of Newly Exposed Structures

  • Alpha-wrapping removes internal or self-contacting surfaces, leaving regions such as armpits unsampled; fine-tuning with a bipartite-matched latent loss guides adaptive completion to generate plausible geometry there.The passage identifies this issue as common in in-the-wild assets, especially AI-generated characters.Source: §C.2 · p. 13

C.3 Data-Driven Learning of Complex Deformations

The model can learn deformation behavior from its training data, including volume-preserving wrist twists when fine-tuned on DQS data. Adaptive shape latents also help reconstruct newly exposed regions that lack source-shape counterparts.

C.3 Data-Driven Learning of Complex Deformations

  • The adaptive completion module reconstructs plausible newly exposed surfaces, whereas the model without adaptive shape latents fails to produce complete geometry.These regions can be absent from source-shape samples because of self-contact.Source: Figure 9 · p. 14
  • The supplementary experiment fine-tunes a model on data generated with DQS to test whether it can learn non-linear skinning behavior.The primary training data uses LBS, which can lose volume during significant joint twists; DQS naturally preserves it.Source: §C.3 · pp. 13–14
  • DQS-trained fine-tuning preserves wrist volume under a severe twist, while the LBS-trained model reproduces volume collapse, without changing the network architecture.The result indicates that the framework can simulate deformation characteristics by changing its training distribution.Source: §C.3 · p. 14

C.4 Results with Texture

The paper applies a diffusion-based material-generation framework to its posed shapes as a feasible texturing solution, while leaving appearance modeling outside its primary focus. The reference image can guide color and style even when its character pose differs from the input shape.

C.4 Results with Texture

  • Hunyuan3D-Paint generates albedo, roughness, and metallic maps conditioned on the 3D shape and a reference image.The paper presents this as a feasible texturing solution for posed shapes, while focusing primarily on shape transformation and leaving appearance modeling for future work.Source: §C.4 · p. 14
  • Textured results remain plausible when the reference image shows a different character pose from the input shape.The authors observe that the reference primarily guides color and style rather than geometry.Source: §C.4 · p. 14

C.5 Sequential Posing and Consistency Analysis

Independent per-frame posing produces plausible mesh sequences, including for challenging characters, but does not ensure topology, UV, or temporal consistency. The accompanying discussion contrasts skinning-based rigging artifacts with the method’s skinning-free posing and notes that geometry-based rigging cannot correct source topology defects.

C.5 Sequential Posing and Consistency Analysis

  • Independent per-frame posing produces plausible results across a motion sequence, including for a character with an exaggerated head-to-body ratio and hands fused into its torso.The close-ups also show that closely contacting limbs may merge in some poses and separate in others.Source: §C.5 · p. 14, Figure 11 · p. 15
  • Independent SDF reconstruction does not preserve mesh topology, UVs, vertex ordering, or temporal consistency, limiting the method to static refinement or pre-rigging rather than production animation.The authors suggest re-posing into a clean rest pose and then using standard rigging tools for downstream applications requiring fixed topology.Source: §C.5 · pp. 14–15

D Skinning Issue of Geometry-Space Rigging

E Analysis on Robustness to Rigging Quality and Failure Cases

The method tolerates moderate skeletal-prior deviations but can produce disconnected geometry when skeleton errors are severe. The discussion identifies current limits and several future directions, including generative completion, multi-source composition, broader controls, and spatially grounded latent representations.

E Analysis on Robustness to Rigging Quality and Failure Cases

  • Severe skeletal errors, such as extracorporeal bones, make dense correspondence unreliable and can produce disconnected limbs.The analysis evaluates sensitivity to automatically extracted skeletal priors.Source: E Analysis on Robustness to Rigging… · p. 16
  • With moderate skeletal deviations such as misaligned foot bones, the skinning-free method tolerates input noise and produces plausible poses, unlike traditional skinning that distorts the mesh.The method’s robustness is therefore bounded; severe prior errors can still cause geometric artifacts.Source: E Analysis on Robustness to Rigging… · p. 16

F More Discussion and Future Work

  • Adaptive completion currently recovers missing geometry deterministically from training patterns rather than generating diverse or novel structures.Proposed directions include generative, multimodally conditioned completion; combining multiple source shapes; using controls such as 3D bounding boxes; and adapting to sparse spatial representations, which would require redefining correspondence as voxel positions and features change.Source: F More Discussion and Future Work · p. 16

Related papers

Loading 2512.16767v3…