SD Prompts - pmcculler/sd-dynamic-javascript GitHub Wiki

Here we talk about all things prompting. Some may just be an outline as a reminder.

Help

You could try asking politely in the prompt-help channel in the Civitai discord.

Here is a pretty comprehensive overview of the terms from photography that can be inserted into prompts to hint how you want the viewpoint oriented in terms of distance and angle to the subject. Note that it is written in Chinese, but machine translation seems to be pretty effective to English, anyway. https://blog.256pages.com/stable-diffusion-prompt-distance/

For this try an extension https://github.com/Haoming02/sd-webui-vectorscope-cc - there may be loras you could bring in as well.

Try "(multi-panel image:1.3)" in your prompt, or this embedding might help, dunno? https://civitai.com/models/71889?modelVersionId=76636

That's escaping the ( and ) characters so '\(style\)' is treated as the word '(style)' and not as if it was the word 'style' with emphasis. Full explanation here: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis

That syntax is used by wildcard extensions; it means 'go to the file named foo.txt and grab a random line and put it right here' and so on.

In Automatic1111 the AND keyword (all caps) is used to stick multiple prompts together, but give them different weight (emphasis). Full explanation here https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#composable-diffusion.
You might see AND_PERP or AND_SALT in the wild; those are from and for the Neutral Prompt extension.

In Automatic1111 the BREAK keyword (all caps) fills up the current chunk and starts a new one. What's a chunk? A set of 75 tokens, and tokens are what prompts are separated into for processing. Why would you want to do this? Possibly for emphasis, as prompt tokens near the beginning of a chunk have more impact, but that's something of an advanced topic. You may be better off using attention emphasis syntax. Some people use it to control color bleed and possibly other effects; see https://www.youtube.com/watch?v=ZIvwjKaP5DI.

So for example, '(bananas:0)' vs '[bananas::0]. In both cases the word has a negligible effect, but under testing we see that it's not the same negligible effect. Don't know why you would pick one over the other. Oh, why would you even do this? I dunno. I use it to identify prompt templates.

It's the parts that Stable Diffusion breaks your text prompt into before if affects the software; surprisingly, perhaps, some words become 2 tokens, and others are effectively synonyms. To visualize as short prompt as token relationships, try this extension: https://github.com/hnmr293/sd-webui-evviz2 and to break prompt words into tokens and look for synonyms and more, take a look at this extension, TokenMixer: https://github.com/Nekos4Lyfe/TokenMixer

You could try asking this tool to make the change in the image? https://www.timothybrooks.com/instruct-pix2pix/

Loosen up, practice, look around. Many keep snippet files with parts of prompts they've seen or written; some keep all their prompts. While I work on expanding this answer, take a look at this guide: https://stable-diffusion-art.com/prompt-guide/

There are many ways, but here's one: use latent mirroring (an extension) with "alternate steps" and method "rolling channels". Then you flip on the setting "drop next to last sigma" which is normally used to denoise second order samplers, apparently, and you should get cool rainbow effects, in theory. Good luck have fun.