Ceci est une ancienne révision du document !
In part seven of An Introduction to Stable Diffusion, we will look at new releases and more. The current release of Easy Diffusion is 2.5.48 beta (1 August), which should automatically update to the new version, or to beta if selected in the settings, whenever you restart. We will also compare Easy Diffusion with Automatic1111, which is an interface for Stable Diffusion.
The beta version interface of Easy Diffusion (ED), and its issues, will be addressed. The open image modifiers section at the lower-left of the interface panel is now at the top-right, and is replaced with a new +Image Modifiers button. It cleans up the look, but, for newbies, it may make options less obvious. If you select Image modifiers, you then choose from the listing previously found in Easy Diffusion.
One of the desirable changes is on the Image Size line. At the far right, just beyond the Height option, is another icon, and selecting it brings up a list of options that you have previously used. It’s a quick way to select 1024 x 1024, for example. In the beta release, even with the ‘Fix incorrect faces and eyes’ selected under Render Settings option, you might be disappointed by the results. Probably the 512 x 512 is the way to go until that rendering option is working correctly. At the bottom of the right-hand panel is also a note indicating the image modifier can be found at the top.
Missing is one helpful feature. Previously holding the cursor over an option would indicate the purpose, and a short suggestion about how to use it. For a newbie, that was very helpful and hopefully will be available in the non-beta release.
Both Easy Diffusion and Automatic1111 do have a similar problem. If you’re wanting to use the newly released SDXL 1.0 model, you are likely to be disappointed as the video ram suggested is 16 GB which is a dramatic upgrade for most user machines and even gamer computers. While I have also seen suggestions that as a bare minimum you need 4 GB of VRAM. when using eight, I still get errors running out of VRAM and stopping image generation. Perhaps as pared down versions of the SDXL model (it’s 6+ GB) are created, the extra RAM might not be required.
Automatic1111 is another interface to Stable Diffusion and seems to have a larger following although there are trimmer interfaces. It was upgraded to 1.5.1 and can now run the SDXL 1.0 model if you have enough VRAM and the suggested 24 GB of RAM. Without the extra VRAM, however, it means I have the same error problem as ED. We will delve into this interface in later articles because it’s still useful for smaller models.
Using the Automatic1111 interface of Stable Diffusion, I have blended two faces. For example one could blend Prince Charles with Joe Biden; King Charles the III is a relatively new name for the reigning monarch, and I am not sure that earlier models were created with that update. The prompt:
[prince charles : joe biden :0.9]
essentially says after 10% of the sampling steps, swap the sampling steps from Prince Charles to Joe Biden. Thus if there were 20 sampling steps, the first two were for prince Charles and the last 18 were for Joe Biden resulting in the graphic being more like Joe Biden. The three generated oil paintings below are:
Prince Charles : Joe Biden :0.9, Prince Charles : Joe Biden :0.5, Prince Charles : Joe Biden :0.1.
A similar prompt works with ED. Next time we will look deeper into the Automatic1111 interface unless other critical updates worthy of discussing are released.