EPFL's PRIMA Builds 3D Animal Shapes From a Single Photo With 0.98 Accuracy
PRIMA combines biological taxonomy embeddings and test-time optimization to reconstruct accurate 3D animal meshes from a single image, beating prior SOTA across all major benchmarks
Takeaways
- PRIMA is a new framework from EPFL for reconstructing accurate 3D meshes of quadruped animals from a single image, achieving state-of-the-art on all major benchmarks.
- It uses BioCLIP embeddings as biological priors to initialize SMAL shape parameters by animal taxonomy, preventing models from defaulting to dog-like shapes for all species.
- A test-time adaptation (TTA) loop refines predictions per-instance using only 2D keypoint reprojection constraints, with no 3D labels required at inference.
- The team used TTA to auto-generate Quadruped3D, a new pseudo-3D dataset of 9K high-confidence samples across 23 species, expanding 3D supervision from 13K to 23K examples.
- PRIMA achieves an ARI shape score of 0.98 (vs. 0.61 for prior SOTA AniMer+), meaning predicted shapes cluster almost perfectly by animal family.
- Code and models are available at github.com/AdaptiveMotorControlLab/PRIMA; integrates with the SuperAnimal pose estimation pipeline.
Getting a precise 3D mesh of an animal from a single photo sounds like a solved problem in 2026 -- but it is far from it. Unlike humans, where decades of motion capture data and a relatively consistent skeleton make 3D reconstruction tractable, animals present a combinatorial nightmare: wildly different body proportions across species, extreme pose diversity, and a chronic shortage of 3D ground-truth labels. The result is that most existing models quietly cheat -- they hallucinate a "dog-like" shape for a zebra, or collapse a sheep into something resembling a cat, because their training data is dominated by dogs and horses.
Researchers at EPFL's Mathis Lab of Adaptive Intelligence have now released PRIMA (PRIors for Mesh Adaptation), a framework that directly attacks this bias problem through three interlocking ideas: biological priors from a species-aware foundation model, a test-time adaptation loop that refines predictions per-instance using only 2D keypoints, and a new large-scale pseudo-3D dataset called Quadruped3D that the framework generates itself.
Why animal 3D reconstruction is still hard
Unlike humans, whose skeletal topology is largely consistent, animals exhibit substantial structural diversity across families. Inter-class variation manifests in pronounced differences in body proportions and limb configurations, while intra-class variation is also significant, particularly in breed-level differences observed in species like dogs. The standard parametric model for quadrupeds is SMAL (Skinned Multi-Animal Linear model) -- essentially the animal equivalent of SMPL for humans. SMAL parameterizes 3D geometry using shape parameters β ∈ R⁴¹ and pose parameters θ ∈ R¹⁰⁵, and is a differentiable function that outputs a posed 3D mesh with 3,889 vertices.
Existing 2D and 3D datasets exhibit highly imbalanced family-level distributions, where families such as Canidae and Equidae dominate the training data, while others like Felidae, Bovidae, and Hippopotamidae are significantly under-represented. Under such long-tailed distributions, models trained in a purely data-driven manner tend to bias predictions toward majority families. The practical consequence: give a state-of-the-art model a zebra, and it outputs a mesh shaped like a dog.

Getting 3D ground truth for animals is also fundamentally harder than for humans. While human reconstruction benefits from large-scale, high-fidelity Motion Capture datasets, obtaining 3D ground truth for animals in the wild remains highly challenging. Optimization-based fitting pipelines require manual inspection and are computationally expensive, severely limiting how much labeled data you can realistically create.