goku

Why we included this project

Goku comes from a joint HKU and ByteDance team whose paper was a CVPR 2025 Highlight, and the model itself is built on rectified flow Transformers. The repo walks through the whole pipeline, from data curation to model design and flow formulation, which makes it handy both as a training recipe and as a baseline for judging generative quality against other open and commercial video generators. The README reports strong results on standard text-to-image and text-to-video benchmarks, so practitioners get concrete points of comparison rather than just claims. Because one architecture handles image, video, and text-to-image generation, researchers can explore several tasks without juggling separate codebases. Developers comparing open alternatives to closed video models will also find the training approach and weights documented in one place.