Be sure to use them if you want to replicate the results. Certain models will provide a "prompt" that helps direct the style/character. For example "1girl" is not a word in english, but it's a tag used on the sites, and thus will behave accordingly, however it will not work in the base SD model (or it might, but with undesired results). This is especially true of anime-focused models trained on the booru sites. Some models understand "words" that are more like tags. Words that are earlier in the prompt are automatically emphasized more. You can decrease emphasis by using such as or (woman:0.8) (numbers lower than 1). IE ((woman)) is more emphasized than (woman). Or add extra parenthesis to add emphasis without that. Likewise you can do (woman:1.2) or some other number, to specify the exact amount. photo, (woman) will put more emphasis on the image being of a woman. You do this by putting them in parenthesis. You can also increase emphasis on particular words, phrases, etc. If you've ever used a booru site, or some other site that has tagged images, it works remarkably similar. So instead of a photo of a woman you can use photo, woman to get a similar result. In this sense, you can write your prompt as a list of tags. For example, "woman" is one, and so is "photo". Instead, your words are converted into "tags" or "tokens", and the AI understands each word as such. However, the AI doesn't "think" like that. You can use plain natural english to write out a prompt such as "a photo of a woman". The positive prompt is what you want the AI to draw, and the negative prompt is what you want it to avoid. You can train these using "dreambooth".īelow that you have two fields, the first is your "positive prompt" and the second your "negative prompt". Different models will have the AI draw differently and know about different things. Models are the "database" and "brain" of the AI. Use the "refresh" button next to the drop-down if you aren't seeing a newly added model. This is a drop down for your models stored in the "models/Stable-Diffusion" folder of your install. There's a variety of options here, that aren't exactly clear on what they do, so hopefully I can explain them a bit.Īt the top of the page you should see "Stable Diffusion Checkpoint". This lets you create images by entering a text "prompt". You may be instructed to, but this disables your "antivirus" that protects against malicious models. ![]() NOTE: Do not use "-disable-safe-unpickle". And -gradio-img2img-tool color-sketch lets you use colors in img2img. deepdanbooru is an additional captioning tool, -api lets you use other software with it like painthua. These three arguments are all "quality of life" stuff. deepdanbooru -api -gradio-img2img-tool color-sketch I did notice that after doing this I could make larger images (going up to 1024x1024 instead of limited to 512x512). ![]() The results I found aren't great, but some people swear by it. xformers is also an option, though you'll likely need to compile the code for that yourself, or download a precompiled version which is a bit of a pain. You can also use -lowvram instead of -medvram if you're still having issues. ![]() These are -precision full -no-half which appear to enhance compatbility, and -medvram -opt-split-attention which make it easier to run on weaker machines. There's a few things you can add to your launch script to make things a bit more efficient for budget/cheap computers. I only got into SD a couple weeks ago, so this might be wrong, but hopefully it can help some people? Commandline Arguments Hopefully this is alright to post here, but I see a lot of the same sorts of questions and basic how-to questions come up, and I figured I'd share my experiences.
0 Comments
Leave a Reply. |