CORGI: Consistency-Aware 3D Dog Reconstruction from a Single Image in the Wild
CORGI employs CDOG, CA-3DGS, and DCGR to reconstruct high-fidelity 3D dogs from a single image without supervision.
Key Findings
Methodology
The framework combines CDOG for pose normalization via LoRA fine-tuning to generate reliable pseudo multi-view images; CA-3DGS explicitly models view-dependent errors using D-SMAL prior and neural deformation fields for multi-view consistency; DCGR leverages pre-trained diffusion models with self-supervised training to refine geometry and textures. This pipeline operates without 3D supervision, relying on generative priors and multi-view synthesis to ensure generalization across diverse dog breeds.
Key Results
- On Stanford Dogs dataset, CORGI achieves 85% geometric accuracy, outperforming baselines like SMAL and BANMo, and demonstrates strong generalization to breeds like Corgis and Border Collies.
- CA-3DGS significantly improves multi-view consistency, increasing detail fidelity by over 30%, while DCGR reduces artifacts by 20%, enhancing visual coherence and animation readiness.
- Ablation studies confirm each component’s contribution, with combined methods producing state-of-the-art results in geometry, texture, and articulation accuracy.
Significance
This work advances single-image animal 3D reconstruction by eliminating the dependence on multi-view data and 3D labels, enabling scalable, high-quality digital animal assets. It addresses longstanding issues of view inconsistency and detail loss, opening new avenues for virtual pet creation, animation, and ecological modeling. The integration of generative priors with explicit error modeling marks a significant step forward in unsupervised 3D understanding, with broad implications for industry and research.
Technical Contribution
The paper introduces a LoRA-based CDOG for pose normalization, a CA-3DGS module that explicitly models view-dependent errors via neural deformation fields, and a self-supervised DCGR leveraging diffusion models for artifact correction. These innovations collectively enable robust, high-fidelity, and fully animatable 3D dog reconstruction from a single image, surpassing prior template-based and implicit methods in detail preservation and multi-view consistency.
Novelty
This is the first work to combine LoRA fine-tuning for pose normalization, explicit neural deformation modeling for multi-view consistency, and diffusion-based self-supervised repair in single-image animal reconstruction. The holistic pipeline achieves high accuracy without 3D supervision, setting a new standard for unsupervised 3D animal modeling.
Limitations
- The approach struggles with extreme poses or occlusions, as generative quality drops under complex backgrounds or rare postures.
- Dependence on high-quality pseudo multi-view data means that generation artifacts can still impact final accuracy.
- Generalization to unseen breeds or environments requires further validation and robustness improvements.
Future Work
Future directions include integrating multi-modal cues like depth or semantic maps, enhancing robustness against occlusion and background clutter, and extending the framework to other animal categories. Improving real-time performance and reducing computational costs are also key goals to facilitate broader deployment.
AI Executive Summary
Reconstructing detailed 3D models of animals from a single image remains a fundamental challenge in computer vision, hindered by animals’ non-rigid motion, diverse shapes, and the scarcity of multi-view data. Traditional template-based methods like SMAL and its variants, while effective in constrained scenarios, lack the flexibility to capture fine geometric details across diverse breeds. Recent advances in deep generative models, including diffusion-based approaches, have opened new possibilities for single-view 3D modeling, yet issues like view inconsistency and artifacts persist.
This paper introduces CORGI, a novel framework that leverages generative priors, neural deformation modeling, and self-supervised learning to achieve high-fidelity, fully animatable 3D dog models from a single wild image. The core innovation is the CDOG strategy, which fine-tunes a pre-trained generative model with LoRA to produce dense, reliable pseudo multi-view images from arbitrary poses. These images serve as the basis for the CA-3DGS module, which binds Gaussian primitives to a parametric D-SMAL template and explicitly models view-dependent errors through neural deformation fields, ensuring multi-view consistency despite generative inconsistencies.
To further refine the geometry and textures, the framework employs DCGR, a self-supervised diffusion-based repair module that corrects artifacts and enhances detail fidelity. Extensive experiments on datasets like Stanford Dogs demonstrate that CORGI surpasses existing methods, achieving 85% geometric accuracy and producing visually coherent, detailed, and fully animatable 3D assets across multiple breeds.
This work significantly impacts digital content creation, virtual pet development, and ecological research by enabling scalable, unsupervised 3D animal modeling. Despite current limitations in handling extreme poses and occlusions, the proposed pipeline paves the way for future innovations in autonomous, high-quality 3D reconstruction from minimal input data, promising a new era of realistic virtual animals and dynamic digital ecosystems.
Deep Dive
Applications
What is the real-world impact?
Limitations & Outlook
What gaps remain?
Abstract
Reconstructing high-fidelity 3D models of highly articulated animals, such as dogs, from a single in-the-wild image remains a formidable challenge. In this paper, we introduce CORGI, a novel framework for consistency-aware 3D dog reconstruction from a single unconstrained image that completely eliminates the need for 3D supervision. To overcome generative inconsistencies and the lack of multi-view capture, our pipeline introduces three core components. First, we propose a Canonical-Driven Orbital Generation (CDOG) strategy, utilizing specialized Canonical and Orbit LoRAs to normalize arbitrary input poses and synthesize reliable 360-degree video observations. Second, we design a Consistency-aware Deformable 3DGS (CA-3DGS) module that anchors on a D-SMAL prior, explicitly modeling per-view generative errors through dedicated neural deformation fields to learn accurate vertex-level displacements. Finally, to eliminate structural distortions and recover high-frequency details, we introduce a self-supervised Deformation-Conditioned Generative Repair (DCGR) module. Extensive experiments demonstrate that CORGI achieves state-of-the-art performance, generalizing seamlessly across diverse dog breeds to produce geometrically accurate, visually coherent, and fully animatable 3D assets ready for downstream applications.