The developers of the Stable Diffusion neural network, previously available only in beta, have released the project to the public. Now anyone can try the service for converting text queries into images.
The easiest way to test Stable Diffusion is on Hugging Face, but there are long queues on the site: one request is processed in about 5 minutes. Through the DreamStudio platform, everything works much faster, but you need an account (you can log in through your Google account or Discord).
In addition, in DreamStudio, each user is given a limited number of credits that can be spent on generation. The more pictures and the higher the resolution, the more expensive each iteration is. To restore previously generated pictures, you also need to spend credits, so you won’t be able to have fun for a long time.
The project is also posted on GitHub. For now, it requires a computer with an NVIDIA graphics card and 6.9 GB of VRAM, in the future it is planned to add support for AMD video chips, including the Apple M1/M2.
How well does everything work? Depends on the user. The more complex and precise the queries, the more chances for cool results. For example, here are the results for the phrase «Daenerys Targaryen in a wedding dress with a dragon in the style of Alphonse Mucha»:
This is Wizard Yoda in Cyberpunk City.
This is Harry Potter Horror. It scares you, but not in the way it should.
In other words, you can’t force a neural network to think instead of yourself: for a clear result, you need to think about what you want to get as an output.
Stable Diffusion allows you to customize the accuracy of matching the query, the number of steps, and other parameters. The maximum possible resolution is 1024 × 1024 pixels, up to 9 pictures can be displayed simultaneously. The neural network does not know how to increase the resolution, for this you will have to rely on third-party programs.
If you try the service, share the results in the comments!
Read also 🧐