Tencent Hunyuan releases open source acceleration library, shortening image generation time by 75%

On June 8, Tencent released an acceleration library for the open source large model of Tencent Hunyuan text generation graph (referred to as the Hunyuan DiT model), which greatly improved the inference efficiency and shortened the graph generation time by 75%. .

#The threshold for using the Hunyuan DiT model has also been significantly reduced. Users can use the Tencent Hunyuan Wensheng graph model capabilities based on the graphical interface of ComfyUI. At the same time, the Hunyuan DiT model has been deployed to the Hugging Face Diffusers general model library. Users can call the Hunyuan DiT model with only three lines of code without downloading the original code library.

Previously, Tencent announced that the large-scale Hunyuan text generation model has been fully upgraded and open sourced, and can be used by enterprises and individual developers for free commercial use. This is the industry's first Chinese-native DiT architecture text generation graph open source model, supporting Chinese and English bilingual input and understanding; using the same DiT architecture as Sora, it can not only support text generation, but also be used as a multi-modal visual generation model such as video. Base.

Tencent HunyuanDiTModelOpen SourceAfter It has been recognized by many community developers. In less than one month, the number of Github stars of the project has exceeded 2100, located in The most popular DiT models in the open source community.

Hunyuan DiT Github page

In order to improve the developer experience , Tencent Hunyuan officially launched a dedicated acceleration library, which shortened the inference time by 75% and improved the efficiency of large model operation. Developers can download the inference acceleration tool through Hugging Face.

The project team achieved sampling step compression and efficient inference deployment of the DiT model through knowledge distillation and the TensorRT high-performance inference framework. . Distillation mainly refers to reducing the number of iteration steps of the diffusion model to achieve acceleration. The overall structure and parameter amount of the model remain unchanged. Users can reduce the number of iteration steps by 50% by using distillation weights without any additional operations and equipment requirements, and the time consumption can be halved. The TensorRT inference acceleration solution can further reduce time consumption through engineering optimization and is decoupled from model weights. Using both simultaneously for inference deployment can reduce inference time by 75%.

Users can directly use ComfyUI’s graphical interface to leverage the community’s collective efforts based on its latest news. At the same time, through cooperation with the Hugging Face team, the hybrid DiT model has been deployed to the Hugging Face official model library Diffusers, and the use and generation code of the model library have been re-adjusted. Users can directly call the hybrid DiT model through this channel, which greatly simplifies The cost of user use.

ComfyUI is a WebUI interface design in the field of Vincentian graphs. It modularizes and graphicalizes the diffusion algorithm in the field of Vincentian graphs, improves the generation efficiency and resource utilization, and also significantly reduces The threshold for developers to use it has been raised. Users can use the Hunyuan DiT Wensheng graph model through the graphical workflow to achieve the same effect as the official model.

ComfyUI usage interface of Hunyuan DiT Wensheng graph model

In addition, the usage ecology around ComfyUI , and also spawned a strong open source community. Hunyuan DiT's support for ComfyUI also allows community members to experience the Wensheng diagram model based on the latest DiT architecture.

As a well-known AI open source community, Hugging Face's Diffusers is a universal library that currently calls various mainstream large-scale Vincentian graph models, and has become the community standard for the use of large-scale Vincentian graph models today.

Adapting the Hugging DiT model into Hugging Face Diffusers can greatly improve the model's ease of use and user base. Users do not need to download and deploy the original code library to their own environment. Developers who have installed the Diffusers library only need to run a few lines of code to call the Hunyuan DiT model, which is very convenient to configure and call. At the same time, Hugging Face and Tencent Hunyuan team jointly optimized the algorithm framework to speed up image generation.

This is also equivalent to providing underlying support for all subsequent uses and developments based on Hunyuan DiT, covering any scenario where Hunyuan DiT needs to be called, including the above-mentioned ComfyUI method . At the same time, for developers, previously configured workflows and plug-ins based on Diffusers can be directly used in Hunyuan DiT with a small amount of modifications.

Lu Qinglin, head of Tencent Wenshengtu, said: "Tencent's Hunyuan Wenshengtu model has received support and feedback from many developers after it was open sourced. We are very happy and are also working on development We will use the feedback from developers to work with the community to improve and optimize the open source ecosystem based on Hunyuan DiT, so that more developers can more conveniently enjoy the latest research results. We also welcome everyone to join us in building the next generation of visual generation open source ecosystem and promote the development of the world. The model industry is accelerating its development.”

Attached is the Tencent Hunyuan Wenshengtu open source large model (Hunyuan DiT model) project link

Official website: //m.sbmmt.com/link/35817bda28b111aa49bd8fdf61878246

##Code: //m.sbmmt.com/link/bb4b90201e39e55c4a9ccfec8436cfb8

//m.sbmmt.com/link/5dacab03c06f42f75c3f21a2c9f98997

//m.sbmmt.com/link/a0b173044f2019316bebc411696e7d35

##