kiwi / README.md
demibit
Adding v1b embeds
ad8f76f
---
language:
- en
tags:
- cyberpunk
- anime
- stable-diffusion
- kiwi
---
# KIWI
_Shameless plug: If you want a *free open source* fully local booru alternative, I have [created one](https://github.com/demibit/stable-toolkit). It automatically tags your images and let's you easily filter by generation parameters. The [wiki](https://github.com/demibit/stable-toolkit/wiki) has some pictures of how it looks like._
## DATASET
- **30a**: 30 images of Kiwi, with dimensions of 640x640.
## VERSION
- **v1**: Trained with clip skip 1 (off) (only applicable for embedding/hypernetwork, as I don't know the status for dreambooth). NAI-analog base model.
## DREAMBOOTH
[Trained via this GUI](https://github.com/n00mkrad/text2image-gui), but note that it's quite restrictive in training options.
- **kiwi-30a-v1**: Trained with "very high quality" preset and "normal" learning rate. [Examples](https://imgur.com/a/sZ3nbrP)
## EMBEDDINGS
_Note: You'll have to rename your chosen embedding to "kiwi"_
- **kiwi-30a-v1a-embeddings**: Embedding trained with voldy's web-ui with 8 tokens per vector, using a learning rate of _0.02:200, 0.01:1000, 0.005:2000, 0.002:3000, 0.0005:4000, 0.00005_, for 30k steps.
- [Examples with the 5k embed](https://imgur.com/a/58JAJ8L)
- [Examples with the 10k embed](https://imgur.com/a/ADmLPfj)
- [Examples with the 15k embed](https://imgur.com/a/cMXAwFF)
- [Examples with the 20k embed](https://imgur.com/a/qaG48Yh)
- [Examples with the 30k embed](https://imgur.com/a/dJHh85F)
- **kiwi-30a-v1b-embeddings**: Embedding trained with voldy's web-ui with 8 tokens per vector, using a learning rate of _0.005_, for 30k steps.
- [Examples with the 5k embed](https://imgur.com/a/yHuWNCz)
- [Examples with the 10k embed](https://imgur.com/a/wHuoHRQ)
- [Examples with the 15k embed](https://imgur.com/a/knd8cpn)
- [Examples with the 20k embed](https://imgur.com/a/Aoir2Ih)
- [Examples with the 30k embed](https://imgur.com/a/FrrC0YT)
## HYPERNETWORKS
_Apparently voldy's hypernetwork training is broken at the moment. Will either rollback, or wait for it to get fixed._