Lumina-T2X

Why we included this project

Lumina-T2X is a good place to see one codebase handle text-to-image, text-to-video, text-to-audio, and 3D generation with the same flow-based diffusion transformer architecture and a shared training and inference harness. It ships working pipelines for the Lumina-Next image models, including a 2K-resolution variant and compositional generation from multiple region captions, plus text-to-music and audio modules, with checkpoints published on Hugging Face and WiseModel. Practical workflows are covered too: img2img translation, Dreambooth-style fine-tuning, and SD3 support, and the image pipeline plugs into the Hugging Face diffusers ecosystem and a ComfyUI wrapper. The repo also carries Flag-DiT and Next-DiT training and comparison material, which makes it handy for researchers comparing diffusion-transformer designs. If you want to study or reproduce a state-of-the-art flow-matching transformer rather than call an API, the code here is concrete and runnable.