Meta research framework that aligns a frozen SAM 3D model to input images or video at inference time for 3D reconstruction.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
GenIA (Generative Reconstruction with Test-Time Input Alignment) is a research framework from Meta and collaborators that reconstructs complete 3D objects from monocular images, multi-view images and monocular videos. It works on top of the frozen SAM 3D generative model without retraining it, extending it beyond its single-view, static setting to multi-view and dynamic inputs.
At inference time it estimates object pose by deriving translation and scale from geometry while keeping the learned rotation prior. It then aligns appearance through visibility-biased attention, cross-observation fusion and differentiable rendering guidance during denoising. An optional test-time refinement step further adapts the appearance latent, lightweight decoder adapters and object placement. For dynamic objects it can take externally supplied geometry (ActionMesh is used) and recover a shared canonical appearance with stable world-space placement.
The project page reports improvements over optimization-based, per-frame image-to-3D and video-to-4D baselines on synthetic and real benchmarks, with minimal overhead over SAM 3D (SAM 3D median of 25.3 s on an A100). Listed limitations: dynamic geometry depends on an external predictor, and rotation remains prior-driven and is corrected only by noise-sensitive ICP registration. A YouTube creator reports the code is released under a non-commercial license. The provided material does not include a repository link, so that is unverified here.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Turns a photo or video into 3D objects by building on SAM 3D and iteratively checking the match against the original image. Code released under a non-commercial license.
Closest neighbours by meaning, not just by category.
Research · Preferred Networks
Generative method from Preferred Networks that reconstructs high-resolution 3D scene meshes from posed multi-view images.
Model
Research model that builds complete 3D assets from sparse, unposed images by jointly doing reconstruction and generation.
Research · Google DeepMind
Google DeepMind research that turns a text-to-video diffusion model into a single feed-forward, multi-task vision model.
Research
Research method that reconstructs PBR-ready indoor 3D scene meshes from multi-view images using a Trellis.2 generative prior.
Research
Agentic framework that reconstructs sim-ready 3D scenes with rigid, articulated and deformable objects from multi-view RGB.
Model · NVIDIA
NVIDIA's 117M-parameter looped transformer for multi-view 3D reconstruction of camera poses and dense geometry.
