diffusion-models
(4)
AxionOrbital Space
AxionOrbital Space is a San Francisco startup building foundation models that turn radar satellite data into analysis-ready optical imagery in real time. Its flagship model, ORION, translates Synthetic Aperture Radar (SAR) backscatter into photorealistic images so customers can see through clouds, smoke, and darkness - the roughly 70% of the time when conventional optical satellites go blind. A Y Combinator Winter 2026 company, it targets defense, commodities trading, and disaster response.
Doji
Doji is a New York-based AI fashion startup building an app that turns a handful of selfies into a photo-realistic personal avatar, then lets people try on real clothing from anywhere on the web. Founded in 2024 by Dorian Dargan and Jim Winkens - the name is a mash-up of their first names - Doji uses its own diffusion models to make virtual try-ons feel less like a gimmick and more like a fun, social way to shop. The company raised a $14M seed round led by Thrive Capital with participation from Seven Seven Six in May 2025, days after launching on the App Store in more than 80 countries.
Inception
Inception builds diffusion-based large language models (dLLMs) that generate text in parallel instead of one token at a time. Its flagship Mercury family is positioned as the first commercial-scale diffusion LLM, reporting 5–10x faster throughput than traditional autoregressive models at competitive quality, with Mercury Coder targeting code generation and Mercury 2 reasoning workloads.
Pika
Pika is a San Francisco-based generative AI company building a consumer video platform that turns text, images, and clips into short cinematic videos. Founded in 2023 by Stanford AI Lab alumni Demi Guo and Chenlin Meng, Pika has shipped a string of fast-iterating model releases (Pika 1.0, 1.5, 2.0, 2.1, 2.2) and viral features like Pikaffects, Scene Ingredients, and Pikaframes. The company has raised roughly $135M from Spark Capital, Lightspeed, Greycroft, and others.