The Atrium Sponza Palace in Dubrovnik, re-modeled by Frank Meinl at Crytek with inspiration from Marko Dabrovic's original. I took the OBJ version exported by Meinl, computed bump maps from the normal maps using normal2bump.cpp (since MTL files expect height bumps, not normals), put the "mask" textures into the alpha channel of the associated diffuse texture, cleaned up noise in the masks, corrected the material mapping for the octagonal vases, and removed the long untextured banner floating in the middle of the atrium that appears in the file but in none of the published images of the model. The zipfile includes the banner as a separate object.
A 3D model of the real Lawrence Berkeley National Laboratory building 90 3rd floor conference room originally created by Anat Grynberg and Greg Ward, circa 1991. The model was created from manual measurements with coordinates directly typed into a text editor. It is commonly used for global illumination experiments.
This model was flipped inside-out at some point in history and appears backwards in some papers relative to the original, with the EXIT signs reading as if in a mirror. Kenzie Lamar at Vicarious Visions fixed the face winding directions on all surfaces to be consistent, re-modeled the trash cans and fire extinguisher to have reasonable normals and joints with no interpenetration, replaced the radiator vents with a single panel, and flipped the EXIT signs. The textures observable (especially in the wood grain) in Ward's original rendering are procedural and are not represented in this model.
Additional resources:
I converted the PLY from Georgia Tech to IFS, and then computed vertex normals and wrote the result to OBJ format using custom code. This model is originally from the Stanford 3D Scanning Repository. It was scanned from a real object by a Cyberware 3030 MS scanner. The model reconstructed from scans by Stanford contains fewer triangles and more vertices. The Georgia Tech version may have been reconstructed independently from the original points. As with many models on this page, I report the vertex count after merging co-located vertices with the same normals during load (i.e., the size that your vertex buffer will actually be in memory), so my vertex count is lower than the one reported by Georgia Tech, but my face count is identical. The icon image is from NVIDIA.
Martin Newell modeled a teapot using bicubic Bezier patches in 1975 as part of an entire tea set. It first appeared in his Ph.D. dissertation. Jim Blinn scaled the teapot horizontally by 1.3x to cancel the effect of the rectangular pixels on their display. The original had no bottom; most variations add one. Kenzie Lamar at Vicarious Visions created this version from the default teapot model in 3D Studio Max. He resized the top so that it fits snugly, created a shell and welded the spout and handle so that the teapot is a manifold without boundary, and created a unique and consistent-resolution texture parameterization for the entire model.
A cube with texture coordinates and explicit vertex normals that I created by hand. The cube is centered at the origin and has unit volume. I included the checkerboard texture and material mapping to make it easy to modify to place other textures on the cube.
Brian Curless and Marc Levoy created this model from multiple range scans of a real object using a Cyberware 3030 MS scanner. I converted the PLY from Georgia Tech to IFS, computed vertex normals, and then exported it OBJ format using G3D.
Samuli Laine and Tero Karras at NVIDIA Research created this mass of thin strands for Two Methods for Fast Ray-Cast Ambient Occlusion.
Donald Goldberg and students at Cornell University created the original Cornell Box radiometry test data and physical box. Guedis Cardenas, Morgan McGuire, and Michael Mara created these eight OBJ files that represent some of the most commonly used test scenes based on that data. These include the original box and many variations by Henrik Jensen (the icon image is by Jensen) such as a box full of water. Note that the original real box is not a perfect cube, so the faces are correspondingly imperfect in this data set.