AI GUIDE: SDXL TURBO (Patreon)
Content
StabilityAI, developers of Stable Diffusion AI text-to-image models, launched recently a new model named SDXL Turbo. Until now, generating images required multiple sampling steps. By default 20 steps in Automatic1111. By being able to do things in a single step, the time to generate images considerably decreases. It can go as low as 200-300ms per image. There's already talk about being able to use AI for realtime processing in the near future. This is crazy quick development considering that last year around this time we were just discovering AI images for the first time.
A few catches:
- unlike other models, SDXL Turbo is for non-commercial use only
- the general quality is lower compared to what people are used to, it will take a while for some to train more realistic models
- because the process is different, with less steps, the images will look fine at their intended size (512x512) but will be more deformed when scaling up compared to older models (at least based on my few tests)
So at the time of writing this I don't really recommend it except for people that want to experiment with something new. The generated images will likely be worse than those generated with other models, in nearly every other way except speed. But people will likely train new models on top of this model quite soon. That combined with how fast it is might bring some great results in the near future.
HOW TO USE IT WITH AUTOMATIC1111
1. Download sd xl turbo model into automatic1111/models/stable-diffusion
2. Edit automatic1111/webui-user.bat and add "--no-half-vae --disable-nan-check" to the COMMANDLINE_ARGS line:
GENERATE IMAGES
You can then select the model like any other model. The settings need to be a little different though:
1. set Sampling steps: 1
2. set CFG Scale: 1
3. generate
Increasing the sampling steps will add more detail to the image, at the cost of speed:
1 step
6 steps