| Oct 02, 2026 | We released our work on post-training frontier text-to-image models by combining human preference and rubric rewards. Our recipe improves FLUX.2-dev’s Arena Elo score by 69 points, and our post-trained Ideogram-4 ranks above every open-source model in the September 4, 2026 Arena leaderboard snapshot. |
| Sep 24, 2026 | Arena-T2I Hard: Benchmarking and Improving Faithfulness with Dependency-Aware Checklist was accepted to NeurIPS 2026 as a Spotlight! |
| Sep 11, 2026 | We released ReCAST, a method for assigning reward credit across diffusion timesteps. |
| Aug 20, 2026 | We released Stream4D, which uses 4D reconstruction rewards and motion priors to improve scene consistency and preserve natural motion in streaming video generation. |