First OpenAI released GPT-4o, then Google released Imagen3, and now, theTencentDelivered his share as well:origin of the universeVincent FigureLarge ModelFully upgraded andOpen Source!
This is the industry's first Chinese native DiT architecture open source model , and supports bilingual input and understanding .
What is DiT architecture? Simply put, Stable diffusion3 and Sora are also using this architecture, but at present Sora is not open to the public, and Stable diffusion3 is not completely open source as previously said, while the hybrid big model is completely open source, from this point of view, I think Tencent hybrid team is still very sincere! (Kudos)
Note: the official release of the architecture diagram, interested in can look at, do not understand you can let Kimi or GPT4 teach you, the test is valid!
So how does the Hybrid-DiT model perform? Allow me to tell you how
1、Support Chinese prompt word
As mentioned earlier, the Hybrid-DiT model supports both Chinese and English inputs, so it's a relatively big plus for domestic friends who don't have to go through the process of converting Chinese to English.
Note: Here are a couple of renderings that were released
2. Long text comprehension skills
Simply put, it is able to analyze and understand the information in long texts and generate corresponding artworks, and this is the official effect image released
3、Support multi-round dialog
This means that it is possible to keep modifying the image through multiple conversations to achieve our requirements, after all, sometimes a single conversation doesn't work well enough to generate a satisfactory image.
If you read my previous post introducing the Vincennes video tool Pika, you probably won't be unfamiliar with this feature, as Pika also supports multiple rounds of dialog to modify videos.
So how do you experience the Hybrid-DiT model? Unfortunately, at the moment, if at all, I have not found a place where I can experience it online!
Although the official website of the Hybrid-DiT model mentions that you are welcome to experience it in the Tencent Hybrid Assistant, I logged in and found that the model in there is not the new open source one (or am I not in the grayscale?). , for three reasons:
The first is that it's marked at the bottom as being based on Tencent's Hybrid Grand Model V1.7.6, and there's no latest open source news in the message center either
The second is that in the official video put out, the version I saw demoed was actually 2.0
The third is by feel, in the hybrid assistant generated in the picture obviously feel not as good as the official website put out, and also need to "generate a picture" and other prompt words trigger.
So if you want to experience it, you can only refer to the instructions on Github to install and experience it, and it just so happens that my computer configuration meets the requirements, which will be followed up by a separate installment of the instruction
Finally in terms of how it compares to other Venn diagram models, here's a test comparison put up on Github:
Note: This is the result of a review conducted by more than 50 professional reviewers
In general, compared to other open source models is improved, but for some closed source models there is still a gap, I hope to become better under the power of open source!
Related Addresses:
The official website of Hybrid-DiT:https://dit.hunyuan.tencent.com/
Hybrid-DiT Github Address:https://github.com/Tencent/HunyuanDiT
Hybrid Assistant Address:https://www.1ai.net/6765.html