AI NEWS 🤖 2x faster AI image generation in Automatic1111 (Patreon)
Content
There's been a breakthrough recently in regard to stabledifussion image generation with what are known as LCM (Latent Consistency Models). This new method of generating images should allow accelerating the process of generating image making it considerably faster. There are many ways to put it to good use, there are new models that have it already in them (they have LCM in their names), but here's the simplest way to use it with any of the existing (SD1.5) AI models:
SPEEDING UP AUTOMATIC1111
1.Install sd-webui-animatediff
In Automatic1111 go to Extensions > Available > Click Load From > search for "sd-webui-animatediff" and install it. I covered this a bit in my Animation guide. sd-webui-animatediff is a pack that allows doing animation with AUTOMATIC1111, but it also includes something that will help with this LCM process here (a LCM sampler).
2. Download the LCM LoRA
Download the LCM LoRA from civitai and add it to your [Automatic1111 folder]/models/Lora folder and restart Automatic1111
HOW TO USE IT
To speed things up, when generating images you need to:
1. select the LCM sampling method, not Euler or the others. If you don't have it, step 1 above wasn't completed successfully
2. Reduce CFG scale to something between 1 and 3, if it's bigger than that the images will have lots of artefacts
3. Reduce the number of Sampling steps. You can cut in half, or even more, what you normally used, to get the same kind of quality as before
4. Go to the LoRA section and add the LCM LoRA to the prompt:
When you click on that it will add this to the prompt "<lora:LCM_LoRA_Weights_SD15:1>" which will activate the LCM LoRA when generating the images. That part needs to be in the prompts for this to work.
Doing this should allow generating images considerably faster, or allow you to generate images twice or three times as big or detailed in the same amount of time as before!