Homam Bahnassi has posted an interesting article showcasing R&D work completed at INFramez games designed to improve the quality of texture baking using machine learning. According to Homam, a typical workflow involves using texel density to evaluate quality, but this misses the importance of human perception. Their approach promises significantly better results without having to increase texture sizes. Currently the tool works only in Softimage, but if there is sufficient interest it may be ported to Unreal and Unity. Find out more on the INFramez blog.
Baking textures with machine learning enhancements
Related Posts
CGPress is an independent news website built by and for CG artists. With more than 15 years in the business, we are one of the longest-running CG news organizations in the world. Our news reporting has gathered a reputation for credibility, independent coverage and focus on quality journalism. Our feature articles are known for their in-depth analyses and impact on the CG scene. “5 out of 5 artists recommend it.”
Wow, what kind of magic is this?
I have a feeling humanity is going to be obsolete soon with all this Machine learning/Deep learning….
Nah, humanity won’t be obsolete. Just largely jobless. lol
Interesting – it seems to be a kind of edge detection algorithm that increases texel ratio in higher contrast areas? If so, I wonder how does it achieve those results given the light map UV’s? Does it somehow split the mesh and create new UV’s on the fly? Also as a side note, you could argue the Sponza example looks better with fuzzy low res shadows rather than the more jaggy ‘enhanced’ higher res edges. But maybe that’s unfair – its a cool tech demo for sure.