MiniT2I Open-Sources Full Training Recipe and Models on HuggingFace
June 20, 2026
MiniT2I, a text-to-image model from collaborators including Kaiming He, is fully open-sourced with code in both JAX and PyTorch, model weights on HuggingFace, and the complete training recipe. No benchmark numbers are provided in this post, but comparisons to other text-to-image models are available via the linked paperswithcode page.
HOW THIS AFFECTS YOU
●
builderYou can pull weights from HuggingFace and integrate MiniT2I into image generation pipelines today using either JAX or PyTorch.
●
researcherFull training recipe and dual-framework code release lets you reproduce and extend the work directly.