Lumina-T2X is an open-source family of diffusion transformer models for generating images, video, and other modalities from text โ a unified “text to any modality” framework from the Alpha-VLLM team. The GitHub repository has around 2,200 stars and collects the Lumina model series, including Lumina-Next and Lumina-mGPT.
Lumina-T2X is built on the Flag-DiT architecture, a diffusion transformer design that scales across resolutions, aspect ratios, and durations without separate models for each setting. The family covers text-to-image generation, multimodal autoregressive generation with Lumina-mGPT, and extensions toward video and speech, giving researchers one codebase for experimenting with modern DiT-based generation.
A central idea in Lumina-T2X is that one diffusion transformer, given the right positional encoding and unified conditioning, can handle images of any aspect ratio and extend naturally to sequences for video. The repository collects the family’s checkpoints and training recipes in one place, which makes it a useful reference for researchers comparing DiT variants or bootstrapping their own text-to-image and text-to-video experiments without starting from scratch.
The project sits within a fast-moving research line: members of the family have been presented at major ML venues and are frequently used as baselines in text-to-image research. Its checkpoints cover multiple sizes, so smaller GPUs can run the lighter configurations while research clusters reproduce the full-scale results, and the codebase documents the training configurations needed to do so.
Lumina-T2X is free and open source under the MIT license, including the released model checkpoints.
Generative-AI researchers and engineers exploring diffusion transformers and text-to-any-modality generation with open checkpoints.