Hunyuan-DiT is Tencent’s open-source text-to-image diffusion transformer, designed from the start for fine-grained understanding of both Chinese and English prompts. Unlike U-Net diffusion models, it uses a pure transformer architecture that generates images at multiple resolutions and follows detailed, paragraph-length descriptions unusually well. The project has about 4,300 stars on GitHub.
The model pairs a diffusion transformer with two text encoders — CLIP and a multilingual T5 encoder — which is what gives it its strong bilingual prompt comprehension, including Chinese idioms and culturally specific concepts that English-first models often miss. Tencent released the full pipeline: model weights, inference code, a Gradio demo, ControlNet and LoRA support, and distilled versions for faster sampling. It supports multi-resolution generation, multi-turn dialogue-based image refinement through the companion DialogGen model, and evaluation tools including an automated prompt benchmark. Hunyuan-DiT became the foundation of Tencent’s Hunyuan image family; the weights are published under the Tencent Hunyuan Community License and the repository remains actively updated in 2026.
Tencent also published the evaluation side of the project: a human-preference study and an automated benchmark showing the model ranking at or near the top of open models for text-image alignment at release, particularly on Chinese prompts. The repository provides training and fine-tuning code in addition to inference, and its multi-resolution design lets a single model serve both quick drafts and high-resolution finals. It requires a reasonably capable GPU, and the community license asks commercial users above certain scales to obtain permission.
Hunyuan-DiT is free to download and use under the Tencent Hunyuan Community License; you only need your own GPU to run it.
Hunyuan-DiT is for researchers, developers and creators who need an open text-to-image model — especially those working with Chinese-language prompts or building custom diffusion pipelines.