DailyArt: Discovering Articulation from Single Static Images via Latent Dynamics

Abstract: Articulated objects are essential for embodied AI and world models, yet inferring their kinematics from a single closed-state image remains challenging because crucial motion cues are often occluded. Existing methods either require multi-state observations or rely on explicit part priors, retrieval, or other auxiliary inputs that partially expose the structure to be inferred. In this work, we present DailyArt, which formulates articulated joint estimation from a single static image as a synthesis-mediated reasoning problem. Instead of directly regressing joints from a heavily occluded observation, DailyArt first synthesizes a maximally articulated opened state under the same camera view to expose articulation cues, and then estimates the full set of joint parameters from the discrepancy between the observed and synthesized states. Using a set-prediction formulation, DailyArt recovers all joints simultaneously without requiring object-specific templates, multi-view inputs, or explicit part annotations at test time. Taking estimated joints as conditions, the framework further supports part-level novel state synthesis as a downstream capability. Extensive experiments show that DailyArt achieves strong performance in articulated joint estimation and supports part-level novel state synthesis conditioned on joints. Project page is available at https://rangooo123.github.io/DaliyArt.github.io/.

• The paper presents a synthesis-mediated pipeline that generates a maximally articulated state from a closed-state image to reveal latent kinematic cues.
• It leverages a frozen image encoder and 3D lifting with transformer-based estimation to accurately compute joint parameters while preserving object identity.
• Experimental results on datasets like PartNet Mobility and AKB-48 show improved success rates and robust generalization compared to prior methods.

Synthesis-mediated Articulated Joint Estimation from Single Static Images: DailyArt

Problem Formulation and Motivation

DailyArt addresses articulated joint estimation and controllable state synthesis from a single static image, directly tackling kinematic ambiguity when only the closed-state observation is available and crucial articulation cues are occluded. Existing approaches require auxiliary priors, multi-view or multi-state observations, explicit part graphs, retrieval from limited databases, or part masks to partially expose the underlying structure. Such priors are brittle and non-scalable, especially for unseen or open-world objects.

DailyArt reframes the problem as synthesis-mediated reasoning: rather than direct regression from the observed image, it synthesizes a maximally articulated state, revealing latent kinematic cues under the same camera view, and estimates joint parameters from the cross-state discrepancy. This separation removes dependency on auxiliary priors, forms a unified pipeline, and enables joint-conditioned controllable motion synthesis.

Figure 1: Overview of the DailyArt pipeline, illustrating maximally articulated state synthesis and subsequent articulation parameter estimation from dual-state discrepancy.

Technical Framework and Methodology

Stage I: Novel Articulated State Synthesis

The first stage employs a backbone built on a frozen image encoder (DINOv2) and a learnable VAE-based decoder. Given a single closed-state image, the model synthesizes the opened, maximally articulated state by injecting a scalar motion index $t$ via AdaLN modulation. No explicit joint index or part-level guidance is used; synthesis is prior-free, driven by latent semantics of the input image. This stage is crucial for revealing otherwise hidden articulation evidence while strictly preserving object identity and geometry.

Stage II: 3D-Aware Joint Estimation

Stage II lifts both closed-state and synthesized opened-state images into dense, confidence-aware 3D point maps using VGGT. Cross-state comparison yields per-point 3D displacement, and non-learned filtering retains only those motion seeds consistent with plausible articulation. These filtered seeds are embedded as queries into a transformer-based estimator which regresses the full set of joint parameters (type, origin, axis, and motion range) in object-centered world coordinates. Slot assignment is supervised by Hungarian matching on ground-truth joint attributes.

Figure 2: Detailed diagram of the DailyArt three-stage paradigm highlighting novel state synthesis, 3D lifting, and joint estimation.

Stage III: Joint-conditioned Controllable Synthesis

Using the estimated joints as explicit conditions, the synthesis backbone is extended to generate joint-conditioned controllable motion sequences. Stage III accepts a selected predicted joint and articulates the object to any state within the motion range, producing visually consistent images with precise structure and kinematic constraints.

Training and Objectives

Training follows a staged, decoupled regime: pixel-level alignment loss (L1 + LPIPS), synthesis loss on articulated state, joint estimation loss (classification + regression), and controllable synthesis loss supervised by ground-truth articulated images. The pipeline operates in a feed-forward progression at inference.

Experimental Results and Numerical Evaluation

DailyArt is benchmarked on PartNet Mobility and AKB-48 datasets, comparing both synthesis quality and joint estimation accuracy with prior methods. It achieves a 68.4% Overall Success Rate on PartNet-Mobility for joint estimation, surpassing Physx-Anything by 5.6 points; individual errors are minimized for all joint attributes. On real-world AKB-48, DailyArt achieves 54.4% success, showing robust domain transfer. For novel-state synthesis, it yields highest PSNR (25.5 PartNet, 19.6 AKB), SSIM, lowest LPIPS and FVD, and highest CLIP-T, consistently outperforming DragAPart, PartRM, Puppet-Master, and LARM.

Figure 3: DailyArt performance on unseen objects; novel state synthesis and joint estimation remains robust without auxiliary priors.

Ablation studies isolate key module contributions: removal of Stage I target-state synthesis causes a 24.2% drop in success rate, and replacing 3D lifting with a 2D encoder reduces success by 18.4%, visually confirming the criticality of cross-state synthesis and 3D geometric reasoning. Direct regression fails to plausibly align articulation axes, and sequential synthesis is less efficient with lower accuracy.

Figure 4: Ablation results highlight failure modes—misaligned joint estimation without target-state synthesis and incorrect geometry without 3D lifting.

Implications and Future Directions

DailyArt demonstrates synthesis-mediated reasoning as a robust strategy for articulated joint estimation from single static images, removing reliance on fragile priors, part annotations, and multi-state observations. Practically, the method enables scalable generation of actionable articulated assets suitable for simulation, control, and interaction in embodied AI contexts, without requiring human intervention or dense annotation. Theoretically, the results underscore the importance of constructing latent cross-state evidence and object-centric 3D geometric reasoning in fundamentally ambiguous cases.

The framework exposes novel gaps: segmenting articulated parts exclusively via foundation models or prompts is insufficient, reinforcing synthesis-based reasoning as a superior means for kinematic inference.

Figure 5: Foundation segmentation models fail to separate moving parts from static base, emphasizing the need for articulated synthesis and joint reasoning.

Future research directions include:

• Extending the pipeline to handle objects without canonical closed/open states or extreme articulations.
• Integrating more advanced 3D generative priors for improved fidelity in challenging viewpoints.
• Adapting synthesis-mediated reasoning for real-world online learning and interactive manipulation across diverse environments.
• Exploring multi-object and multi-joint generalization, leveraging joint-conditioned synthesis for fine-grained controllability in robotics and world models.

Conclusion

DailyArt introduces an overview-first pipeline that systematically exposes and leverages latent articulation cues from single closed-state images, enabling high-fidelity articulated state synthesis and precise joint parameter estimation without priors or auxiliary inputs. Results establish clear numerical improvements over prior baselines, with robust cross-domain generalization and significant gains in both controllable synthesis and kinematic estimation. The approach has broad implications for vision-based control, simulation, and asset generation in embodied AI, laying groundwork for scalable articulation reasoning from minimal image evidence.