Using "a pre-trained T5-XXL encoder" for text encoding, and "operating on latents from the pre-trained VAE from Flux Schnell". To my understanding they've trained the denoising U-net using licensed data, but it is still largely relying on web-crawled data - just that those components of the model were trained by Google and Black Forest Labs respectively. The demo does, for instance, give you the incredible hulk when you ask for the incredible hulk.