uri-eliabayev talks with misha-feinstein, CTO of Bria, an Israeli company training visual foundation models from scratch for “professional pixels” (enterprise marketing and media). Three pillars: fully licensed training data (no scraping, with indemnity and IP ownership for customers); fine-grained control; and open weights that enterprises can keep training on their own data. Training needs data diversity, a large interconnected cluster (one failing machine can sink a run), and architecture. Instead of buying catalogs up front, Bria pays data owners revenue share by attribution: each generated image is decomposed (color, composition, semantics) into embeddings matched to training images, capped at ~200 owners per image, which steers contributors toward under-covered concepts. Control: training on images paired with long structured JSON (objects, attributes, relations, lighting) instead of short captions. A reasoner turns a prompt into JSON, and users edit one field (chair color) without other things drifting. Customers can fine-tune the JSON vocabulary itself (a studio’s “eyebrow shape”). Frontier models wow in demos but commoditize everyone’s output and often fail production on cost and latency (“a 5-kilo hammer for snowflakes”). Next: cost, on-prem, and from pretty images to production-ready assets.
Show: hidden-layers
Concepts: image-generation · ai-moats · open-weight-models · model-selection