May be worth knowing about Automatic1111 (https://github.com/AUTOMATIC1111/stable-diffusion-webui) for Windows/Nix environments. It's less important there than for the goofiness that is Mac, but convenient and has a lot of useful tweaks easily available, albeit without the low-spec compatibility of some command-line options like the basujindal fork.
Will also likely help a lot of people to know the underlying limits to CLIP; with a few specialized exception forks that break up prompts, a prompt can only have 75 tokens (the AI's equivalent of syllables, but not quite?) to work with for SD. Some UIs will warn about exceeding token max, but most will silently truncate.
Some UIs expose prompt highlighting and/or cross-attention, which can allow you to put more emphasis on a part of a prompt, or even change the emphasis at certain progress steps. This is a complex tool, but if it's supported can give a lot of good options for more subtle or smaller details.
Lastly, be aware that prompt order, punctuation, and even capitalization matter. Putting styles earlier seems to give more refined pictures but less control over content, while putting styles only at the end of a long prompt can have their effects be less severe -- sometimes it can be worth it to give a weak style guide (eg, "portrait", "photograph", "realistic", "cartoony") in the front half, and then more specific style or artist terms toward the tail end. If a prompt seems to be getting ignored, sometimes adding commas or periods before and after can help.
I've been using stable diffusion UI v2.195 . They update it very regularly with new features, saving txt with prompt and seed and all details along with an image to any folder automatically. Works on CPUs as well ! Very slow if using CPU though.
May be worth knowing about Automatic1111 (https://github.com/AUTOMATIC1111/stable-diffusion-webui) for Windows/Nix environments. It's less important there than for the goofiness that is Mac, but convenient and has a lot of useful tweaks easily available, albeit without the low-spec compatibility of some command-line options like the basujindal fork.
Will also likely help a lot of people to know the underlying limits to CLIP; with a few specialized exception forks that break up prompts, a prompt can only have 75 tokens (the AI's equivalent of syllables, but not quite?) to work with for SD. Some UIs will warn about exceeding token max, but most will silently truncate.
Some UIs expose prompt highlighting and/or cross-attention, which can allow you to put more emphasis on a part of a prompt, or even change the emphasis at certain progress steps. This is a complex tool, but if it's supported can give a lot of good options for more subtle or smaller details.
Lastly, be aware that prompt order, punctuation, and even capitalization matter. Putting styles earlier seems to give more refined pictures but less control over content, while putting styles only at the end of a long prompt can have their effects be less severe -- sometimes it can be worth it to give a weak style guide (eg, "portrait", "photograph", "realistic", "cartoony") in the front half, and then more specific style or artist terms toward the tail end. If a prompt seems to be getting ignored, sometimes adding commas or periods before and after can help.
Thanks, this is all super useful and will be good info for the prompt engineering guide.
How is Macintosh goofy? It's a FreeBSD implementation that's lightyears ahead of anything Microsoft has ever done to Windows.
Updated DiffusionBee now records the seed used (& all other settings used) in the history tab.
As a newbie, this overview was very helpful. Sincere thanks for sharing your knowledge and advice.
I've been using stable diffusion UI v2.195 . They update it very regularly with new features, saving txt with prompt and seed and all details along with an image to any folder automatically. Works on CPUs as well ! Very slow if using CPU though.
Thank you for this! So helpful!
how do you add negative prompts to playground ai?