Meta's GenIA turns regular photos and videos into complete 3D assets

GenIA is an open-source 3D reconstruction framework from Meta that turns standard photos, sparse multi-view images, or dynamic video footage into complete 3D digital assets.

While existing generative 3D foundation models can hallucinate unobserved angles, they constantly struggle to faithfully match the geometry, texture, and pose of the original input. GenIA fixes this gap.

How it actually works

Instead of retraining an expensive base foundation model from scratch, GenIA introduces a test-time input alignment framework that anchors generative priors directly to real observations.

It builds on top of SAM 3D. During the denoising stage, it applies visibility-biased attention, cross-observation fusion, and differentiable rendering guidance to lock the appearance strictly to what was captured.

For spatial positioning, GenIA computes translation and scale directly from the observed geometry while preserving learned rotation priors, preventing generated objects from drifting or scaling randomly.

It also includes an optional post-denoising refinement step. This continuously compares the generated mesh against the source imagery, tuning lightweight decoder adapters and appearance latents to eliminate discrepancies.

When fed dynamic objects over time, the system can even extract a unified, input-aligned canonical appearance with stable placement in world space.

Performance and open-source availability

Across both synthetic benchmarks and real-world datasets, GenIA outpaces recent optimization-based tools, per-frame image-to-3D models, and video-to-4D pipelines in both pose accuracy and reconstruction fidelity.

Meta has released the full source code alongside setup instructions for local machine deployment on its project page.

One big catch to keep in mind: the release falls under a Creative Commons Non-Commercial license, meaning you can tinker and run experiments locally, but commercial deployment is off the table for now.

Why it matters

GenIA takes a very smart shortcut by fixing output drift at inference time instead of burning massive compute on retraining base weights. For developers and 3D creators, test-time alignment is proving to be a much more practical way to turn 2D media into usable meshes.

If you work in digital asset creation, robotics, or computer vision, it is definitely worth pulling the repo to test on messy monocular captures. Just remember that the license keeps it strictly in the personal project and prototyping bucket for now.