Alpha-VLLM/Lumina-T2X

Using flow-based diffusion transformers, the framework generates images, audio, video, and other modalities at variable resolutions and durations from text prompts with unified architecture. It integrates with Hugging Face Diffusers and supports inference/training workflows including DreamBooth, with pre-trained checkpoints available across multiple model sizes (2B-5B parameters). The approach leverages large transformer models as the core diffusion backbone, enabling compositional generation and multi-modal control capabilities beyond standard text-to-image pipelines.

2,254 stars. No commits in the last 6 months.