this post was submitted on 10 Jun 2024
1 points (55.6% liked)

Casual Perchance

87 readers
1 users here now

Casual Perchance

A nonspecific casual place for anything Perchance, including generator outputs, memes, prompts, casual discussion, advertising your generator, and anything else you wouldn't post in the more technical Perchance Forum.

This is where to post Community Events.

Resources:

Rules:

founded 8 months ago
MODERATORS
Alllo
edwardthefma99
[email protected]
AdComfortable1514
VioneT
[email protected]
Cocell
1
[T2i] What are the rules of this game? (lemmy.world)
submitted 5 months ago by AdComfortable1514 to c/casual_perchance
3 comments fedilink hide all child comments
 

Just a fun WTF result I got from a prompt recently

:

images

:

you are viewing a single comment's thread
view the rest of the comments
[–] j4k3 2 points 5 months ago (1 children)

Very cool. I've been working on learning this in Comfy. Any recommendations?

[–] AdComfortable1514 2 points 5 months ago* (last edited 5 months ago)

Thanks :) !

  1. Download the vocab.json and find some way to sample it : https://huggingface.co/runwayml/stable-diffusion-v1-5/raw/main/tokenizer/vocab.json

Note that the vocab.json is badly configured for emoji symbols. They read as a garbled mess like ĮãĤĬãģŁãģĦ

Ignore those.

Instead use emojis in prompt. Find some ways to sample these as well : https://perchance.org/fusion-t2i-emojis-2

All unicode symbols (not required , but can be fun to see what weird stuff you can create with these symbols) : https://perchance.org/fusion-t2i-unicode-2

You can try entering some words/emojis in this SD tokenizer , and you will see that they match the vocab.json list : https://sd-tokenizer.rocker.boo/

  1. Divide vocab.json list two categories:

Suffix tokens = tokens that end with , which represents whitespace. Like "banana " .

Prefix tokens = tokens without whitespace. Like "post" used in words like "postapocaplypse " or "postman "

Prefix tokens give properties to suffix tokens when placed in front of them.

  1. If you have a means to invoke random prefix- , suffix- and emoji- tokens you can now mix them to create crazy results

I recommend using Notepad++ to sort the tokens.

  1. Follow the cross attention rule: "SD reads the prompt left to right , one token at a time, finding association from the previous token to the current token (and the inage rendered thus far)"

You can try the -prefix . -suffix and the emoji-tokens out on the fusion generator if you like (I'm the dev) : https://perchance.org/fusion-ai-image-generator

There are also other datasets there which you may find useful.

There is a tutorial at the top, but its kinda wordy and haphazard.

I recently wrote a guide on Fusion Generator Discord for NSFW stuff : https://discord.gg/j6qAX3MWub

You can also find savefiles posted there with example prompts which can be loaded into the fusion-generator

I don't want to present it as a "how to prompt" guide since there is no correct way to prompt things.

It's more a way to show users that there are more options available to them , should they wish to use them.