NextStep-1.1 refines StepFun's image model with flow-based RL

Per StepFun's post, NextStep-1.1 is a re-engineered update to StepFun's image-generation model, improving image fidelity and training stability through extended training combined with flow-based reinforcement learning and a fix to numerical instability in the earlier autoregressive RL process.

The original NextStep-1 explains the underlying approach these changes build on: rather than the common method of converting an image into a fixed vocabulary of discrete tokens before generating it step by step, NextStep pairs a causal transformer with a lightweight flow-matching head that predicts continuous image tokens directly, and StepFun reported quality comparable to established models Flux and SD3.5, with the approach accepted as an ICLR 2026 oral presentation (NextStep-1 details). NextStep-1.1's two changes are refinements to that same continuous-token approach rather than a new architecture.