Generative Vision Atlas

Design axis

Training signal

What besides the base objective shapes what the model learns — data, auxiliary losses, reward models?

Online RL / policy gradient

Fine-tune a generator against a reward model with online policy-gradient methods such as GRPO, rather than only maximizing likelihood on a fixed dataset.

Representation alignment (REPA-style)

An auxiliary loss during generator training that pulls its intermediate hidden states toward a frozen vision foundation encoder's features, without changing the generator's own latent space.