I am, somewhat, a fan of Pokemon. And fan of machine learning. Well, not only pokemon - any-mon works good, but I wanted a network that is capable of training actual, good fantasy creature images. And so I thought - why not make a tool that will actually help making specific datasets? It is not that hard; but will ease a lot of burden afterwards. So.

Initial idea:
Generate datasets, consisting of image + text files. Text files should have specific data in them.

Idea for prompt/data:
A digital art of gen1_pokemon water_typepokemon electric_typepokemon water2_egggrouppokemon resembling a anglerfish wearing a mask in nintendopokemonstyle style
Medium - favorite generation - pokemon types (to not intersect with existing tags like 'grass' or 'water') - egg type (optional, but who knows) - blip+human-edited description of pokemon - in nintendopokemonstyle style

That SHOULD allow us to preserve most 'ideas' from original network, by adding just a couple of new tags. But if we do not want to get any specific things from pokemon such as gen or types, we need to teach network different prompts. There is where generator comes in.

-------------------

First of all there is a 'modules'. Just a folders with subfolders (and possibly sub-subfolders, and sub-sub-subfolders...). Modules have names such as 'gen1', 'gen2', 'gen3', and allows to quickly add or remove a lot of data from dataset.
Inside a module there is any number of folders - actual data lines. in case of gen1 it would be gen1/1, gen1/2 ... gen1/151. Yeah, 151 folder corresponding to every pokemon in dex for this gen.
The last piece - data. inside each folder there is one image file, named image.jpg; and a bunch of .txt files, named like 'generation.txt', 'eggtype.txt', 'pokemontype.txt', 'description.txt' and so on.

When it is time to generate a dataset; you feed a yaml file to a script.

---
modules:
  - pokemon/gen1
  - pokemon/gen2
  - pokemon/gen3
  - pokemon/gen1megas
  - pokemon/gen8gigas
  - fakemon/alphaversions
  - fakemon/darkandwindie/gen1

prompts:
  - "a digital art of [[generation]] [[pokemontype]] [[eggtype]] resembling [[description]] in nintendopokemonstyle style"
  - "a digital art of [[generation]] [[pokemontype]] resembling [[description]] in nintendopokemonstyle style"
  - "a digital art of [[generation]] [[eggtype]] resembling [[description]] in nintendopokemonstyle style"
  - "a digital art of [[generation]] resembling [[description]] in nintendopokemonstyle style"
  - "a digital art of [[pokemontype]] [[eggtype]] resembling [[description]] in nintendopokemonstyle style"
  - "a digital art of [[pokemontype]] resembling [[description]] in nintendopokemonstyle style"
  - "a digital art of [[eggtype]] resembling [[description]] in nintendopokemonstyle style"
  - "a digital art of pokemon resembling [[description]] in nintendopokemonstyle style"


I guess you can see where it is going. When activated, script will go to each folder of a module, generate a specific name for file (modulename-foldername-promptid) and put image here, along with text file, containing one of the phrases. Every single square brackets words will be replaced with a text file contents with a same name. If there is no such file inside; script by default will halt; but instead of halting can just skip this prompt - quietly or with a warning.

output result will be either a folder full of files; or (in future) a dataset, compatible with diffusers.


to use main.py script; there is 1 required and 4 optional flags.

required:
  -c or --config - a YAML file containing modules used and prompts.

optional:
  -o or --output - will change output directory. by default - ./output
  -s or --skip - if prompt requires a specific txt file but there is none in folder - script wont halt, but instead, will just skip that prompt. If module does not exist; it will skip it. if image does not exist; it will be skipped.
  -q or --quiet - surpresses warnings from --skip.
  --dataset_filename image.jpg - will search for this file in every sub-directory
  --prefix 2 - will produce files like 2braixen_1_0.jpg instead of braixen_1_0.jpg, if prefix = '2'