holotard / README.md
hollowstrawberry's picture
Update README.md
6143346
|
raw
history blame
16.1 kB
metadata
license: creativeml-openrail-m
language:
  - en
pretty_name: Hololive vtuber LoRAs and TIs
task_categories:
  - question-answering
tags:
  - stable-diffusion
  - vtuber
  - hololive
  - stable diffusion 1.5
  - textual-inversion
  - lora
  - character
  - text-to-image
pipeline_tag: text-to-image
metrics:
  - character

Preamble

These resources are intended to be used with stable-diffusion-webui. If you don't know what it is or how to use it effectively, here's an extensive guide about it.

Recommended additional resources:

  • blessed2.vae.pt - without a VAE selected in the settings your colors will look faded.
  • TagComplete to view all available anime tags as you write your prompt.
  • Image Browser (updated fork) to view your past images and their prompt metadata.
  • Lycoris extension to be able to use locons and lohas, which are new types of lora.
  • EasyNegative - put it in your stable-diffusion-webui/embeddings folder and write EasyNegative in your negative prompt to drastically improve your images.

I make Loras of my own, particularly multi-outfit or multi-character Loras. Many of the Loras found in this repo are made by me or my peers. Here are some other Loras I made.

 

Models

I merged these with hll4-p3-ep8, which is the newest vtuber-finetuned model. They have blessed2 VAE baked in. Many more mixes can be found here.

Click here for a comparison

A comparison of different vtuber checkpoints

 

Loras

Most useful holo Loras will be linked to and backed up here. Put them in your stable-diffusion-webui/models/Lora folder. Use them in your prompt like this: <lora:filename:1>
Most Loras work well with a weight of 1.0, but some older ones work best at 0.7.

Learn to make your own Loras with my guide.

 

Multi Outfit Loras

These are the most modern and have several outfits of each talent in a single Lora, or the main outfit if no other options are available. Check the model page / text files for examples.

HoloEN

HoloID

HoloJP

Others

 

Small Loras

Some older Loras were manually scaled down from 144 MB to 18 MB or 36 MB and perform almost the same. They need a higher weight than the original to preserve detail.

This works because dim 128 used to be the default setting for making Loras but that was completely overkill and dim 8/16/32 work just as well.

Click here for a comparison

A comparison between a vtuber made with the original Lora and the compressed Lora A comparison between a vtuber made with the original Lora and the compressed Lora A comparison between a vtuber made with the original Lora and the compressed Lora A comparison between a vtuber made with the original Lora and the compressed Lora A comparison between a vtuber made with the original Lora and the compressed Lora A comparison between a vtuber made with the original Lora and the compressed Lora

 

Legacy Loras

Click here for old outfit loras

Main purpose of this repo. Format: (Source 1) (Source 2)... where each source can have multiple outfits.

The default outfit is the first/main one for that talent. Generic means it's not trained on a particular outfit and can do various clothing. In any case, most Loras can change clothes.

Example usage of each outfit can be found in the metadata of the image previews or in the text files next to each Lora.

HoloEN

HoloID

HoloJP

Others

(Legacy) Embeddings

TIs or textual inversion embeddings are an older but lighter alternative. They go into your stable-diffusion-webui/embeddings folder and are used like a normal word in your prompt.

Nobody makes these anymore because Loras are easier to make and work better for most applications.