A week after OpenAI's SORA domination, StabilityAI also released a new model, StableDiffusion3 (hereinafter referred to as "SD3"), last week. This model has ushered in a revolutionary improvement in the quality of generated images, multi-topic prompts, and text writing effects, and has become the "most powerful" Wensheng diagram model of StabilityAI.
Say goodbye to gibberish and render text more accurately.
In the figure above, the SD3 model not only generated a combination of virtual and real, natural light and shadow and visually comfortable picture, but also accurately wrote the English of "if you don't succeed, you will become benevolent", which changed the public's impression that it was difficult to output text from the previous Wensheng diagram model.
2. A more accurate understanding of physical rules.
Judging from the official example picture, the SD3 model seems to be striving to become the god of junior high school physics in ancient Greece, restoring the scene depicted by the prompt "a horse elegantly standing on a colorful ball".
3. Improve the ability of multi-topic prompting.
Now users can also enter multiple theme prompts at one time, in the past, how to accurately restore the attributes and positions of multiple prompt word objects, is a difficult problem to be solved by the Wensheng diagram model, from the official renderings, the current SD3 has been able to understand the elements of multiple prompt words such as "astronauts, pigs in tufties, pink umbrellas and robins".
Under the trend of curiosity, some netizens used the same multi-topic prompt words to generate images through the models of several other AI Wensheng diagrams, and started a battle for the king of volumes.
4. The generation effect is higher quality.
Compared to the previous version, the quality of the images generated by SD3 has been significantly improved, for example, the image produced by "Close-up of a chameleon on a black background" is shown in the image above, which is also in line with the journal magazine.
5. New functions such as image conversion are added.
In addition, the founder of Stability also said that, first, the SD3 model also supports the use of text to modify the content of the picture, and precisely control every element in the image, including replacement and deletion. Second, the image is seamlessly transferred, and the "grafting" without any traces of alteration makes people shout amazing.
These continuous improvement of functions, thanks to the model uses the same architecture as SORA Transformer technology and flowmatching technology, although from the release time seems to have a kind of "since you want to roll, simply roll hemp" rush, but the use of new technology is also an earlier decision, this architecture is also the same as SORA from last year's **.
It is reported that, like SORA, SD3 is not yet fully open, and the company's CEO said that the model will be open-sourced based on user feedback in the future. But even if it hasn't been opened yet, there have already been a number of netizens who have said that their computer configuration is almost unbearable.
The ΣΆ RTX 4080 SUPER Metal Master series accelerates your production and creation experience. Equipped with a full-blooded version of the AD103-400 core, 16GB GDDR6X large video memory and *** Tensorcores, the third-generation RTCres, its professional productivity and game performance have reached an impeccable level, and with the blessing of the TensorRT plug-in, the production efficiency of AI graphics can also be instantly improved.