The problem
Traverse Research investigates neural texture compression and its place in a rendering pipeline. We work on training, learned representations and the trade-offs between storage, reconstruction and evaluation cost.
Texture compression is a system problem. A smaller representation can require more computation, different sampling behaviour or a more involved asset pipeline. The quality target also varies by channel and material. We investigate the whole path from the source texture to the rendered result.
What we do
The Breda research environment includes texture-compression training code, multiple learned model representations and integrations with conventional compression tools. The work spans the training side and the GPU rendering context in which a representation must eventually be evaluated.
Our research builds on published neural texture-compression work. The focus of a partner engagement is the question that matters to its workload: which quality level, which data footprint, and what decoding or integration cost can be accepted.
Evidence and reporting
We evaluate compression against image quality, model size, training cost and runtime behaviour, using representative materials and target hardware.