Farm3D vs Text-To-4D

Farm3D is a research method for single-view 3D reconstruction of articulated animals, built at the University of Oxford and published at 3DV 2024. You feed it one photo of a horse, cow, or sheep and it returns a full 3D shape with texture in seconds, without training on real photographs. The project page documents the approach, shows reconstruction demos, and links to code and the Animodel benchmark dataset.

Most 3D generators need large labeled 3D datasets or multi-view captures. Farm3D trains entirely on synthetic views produced by Stable Diffusion, then uses the same diffusion model as a critic during learning. That lets it recover fine details like legs and ears on categories it never saw in real images, which is unusual for monocular reconstruction pipelines that rely on real-world supervision.

Researchers studying articulated 3D shape, computer vision engineers prototyping animal asset pipelines, and 3D artists exploring controllable synthesis will find the demos and open code useful. The method also supports relighting, texture swapping between same-category models, and skeletal animation once a shape is generated.

Text-To-4D is a Meta AI research project (MAV3D) that generates three-dimensional dynamic scenes from text descriptions. The method uses a 4D dynamic Neural Radiance Field optimized for appearance, density, and motion consistency by querying a text-to-video diffusion model.

Unlike static 3D generators that output a single mesh, Text-To-4D produces scenes you can view from any camera angle and composite into other 3D environments. The approach needs no 3D or 4D training data; the underlying text-to-video model trains only on text-image pairs and unlabeled videos.

The project page hosts demo samples for text-to-4D prompts like "a corgi playing with a ball" and image-to-4D conversions from still photos. It is a research showcase, not a commercial product with a public API or signup. Researchers and 3D artists interested in NeRF-based dynamic scene generation can explore the paper and sample outputs on the site.