NAI Diffusion: Various Tips & Tricks - TravelingRobot/NAI_Community_Research GitHub Wiki

A few collected tips & tricks from either myself or that I have seen floating around the NAI discord.

In general, I will be focusing this wiki on text gen, so this guide will not be updated very frequently.

Read The Docs!

Really, this should be the first thing you look at if you haven't already. There are some gems in there.

In addition, as with everything NAI, you want to read the corresponding guide on the UKB.

Use Tags!

Danbooru is part of the training data for NAI diffusion. Which means you can use its extensive tagging system to steer your generation. Proper usage of the tag system can give you amazing abilities to get the image just the way you want it!

I recommend browsing images on Safebooru (SFW version of Danbooru) to get a feel for how things are usually tagged. If you are interested in NSFW gens, browse Danbooru (NSFW, obviously).

The number of tags in the data can be a bit overwhelming, but this overview should help to give you an idea of how they are organized (link has no NSFW images, but still lists NSFW tags!). If you want to see image examples for NSFW tags as well, here is the Danbooru version.

Additionally, if you have an image similar to what you are trying to generate, there is a fantastic tool called deepdanbooru. It will try to estimate the most likely tags for any given image. You can also run it locally by downloading it from its repo.

Not all tags will work. The more images there are for a tag, the better the chance the AI โ€œgets itโ€. Sometimes strengthening a tag with a few {} can help.

Some obscure tags might work better on the full model, more common tags might work better on the curated one.

Many of the tags are somewhat counter-intuitive. Make sure to read the wiki entry for a tag if there is one!

Some Especially Useful Tags

Specify Number Of Characters

  • Specify the number of characters with <n>boy(s), <n>girl(s) and <n>other(s). So for example 2boys, 1girl (but more than 1 char in a scene remains tricky).
  • solo combined with the tag 1boy, 1girl or 1other if you want to make sure there is only one character in the image.
  • solo focus for putting one character in focus and the rest in the background (maybe character focus as well?)

Direct The Gaze

  • looking at viewer to direct the character's gaze towards the viewer. More useful tags to direct the gaze here (scroll down to โ€œgazeโ€).

Specify General Composition

  • Plenty of useful tags for directing image composition here
  • portrait to have an image focused on the face of a character

Paint Her Black

  • In an attempt to get more diverse gens, Alexei P discovered that African-American will generate characters with a โ€œlight mocha [skin] toneโ€
  • For darker skin, combine with dark black skin. Strengthen that term as needed (e.g., African-American, {dark black skin})

Background

Discord user Myeong_Chung is working on a collection of prompts for specific background here.

Use The Undesired Content Field!

The Undesired Content field is really useful to get rid of aspects of the image you do not want. Use it just like you would the prompt field, but in reverse โ€“ put in stuff you want less off. The same rules as for the prompt apply here. So browsing potential tags can be useful here as well.

Unwanted Content Presets

The presets already contain a selection of tags to subtract from your image.

Note that nsfw is part of those presets if you use the full model or the furry model! However, if you use nsfw as a tag in your prompt, it is automatically removed from the preset. This seems only true for nsfw. Other tags will remain in the preset. So if you use the Low Quality preset and also prompt for lowres, it would be present both in your prompt and in your UC.

Anime Models

  • Low Quality contains: nsfw, lowres, text, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
  • Low Quality + Bad Anatomy contains: Same as Low Quality + bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits

Note: nsfw is only part of the preset for the full model.

Furry Model

  • Low Quality contains: nsfw, {worst quality}, {bad quality}, text, signature, watermark

Prompt Emojis For Giggles!

As discovered by Zaltys, you can prompt emojis! Example from Zaltys from the NAI discord:

Prompt: ๐ŸŒ‡๐ŸŽ†๐ŸŽ†๐ŸŽ†๐Ÿ’‘

I have no idea why this should work! But apparently it does :shrug:.

Discord user A Mystery is working on an overview for which emojis work well here.

SAVE (Yourself From Losing Your Gens)!

Make sure to save, save, save! Your prompts and images are lost when you close the tab.

I recommend installing a browser extension that caches your entries to text fields (I use Typio Form Recovery). That way, you at least should never lose a prompt.