ling0322/libwaifu-illustrious-xl-v2.0

Illustrious XL v2.0-STABLE for libwaifu

Illustrious XL v2.0-STABLE converted to the format libwaifu reads, so it can draw from it directly. The weights are the original ones: the U-Net and both text encoders are stored as float16, and the VAE is left wider because it overflows float16.

Converted from OnomaAIResearch/Illustrious-XL-v2.0, sha256 c2a1a3eaa13d4c107dc7e00c3fe830cab427aa026362740ea094745b3422a331. This is the official Illustrious release rather than a fine tune of one -- WAI and NoobAI, published beside this for libwaifu, are both trained from Illustrious. It is prompted with danbooru tags rather than with sentences.

v2.0 was trained on a cosine annealing schedule, and STABLE is the checkpoint from its last annealing phase: the author describes it as subtly different from the earlier v2.0 weights and generally steadier at generation time.

This is an epsilon-prediction model, which is what libwaifu's Euler sampler reads.

Both halves of the autoencoder are here, so it draws from a picture as well as from a prompt.

Files

A model is a .yaml and the files it names, all of them in one directory. The manifest is a small text file you can read and edit; the weights are ordinary safetensors, readable by torch, numpy or the viewer on this page; the tokenizer is the tokenizer.json Stability published.

file	size	what
illustrious-xl-v2.0.yaml	2 KB	the manifest: what the model is, and which files it is made of
illustrious-xl-v2.0.tokenizer.json	3.5 MiB	the CLIP tokenizer
illustrious-xl-v2.0-00001-of-00004.safetensors	1.88 GiB
illustrious-xl-v2.0-00002-of-00004.safetensors	1.88 GiB
illustrious-xl-v2.0-00003-of-00004.safetensors	1.87 GiB
illustrious-xl-v2.0-00004-of-00004.safetensors	0.99 GiB

The weights are split because seven gigabytes is an awkward size to fetch: it cannot be downloaded in parallel and a failed transfer starts over. The split is between tensors, never inside one, so each file is a safetensors in its own right and can be read and checked alone. The picture the four draw is identical, to the byte, to the one an unsplit model draws.

Usage

waifu knows this model as sdxl:illust, so there is nothing to download by hand:

waifu draw -m sdxl:illust

sdxl:illust:v2.0 names this release and keeps meaning it; sdxl:illust follows whatever is current.

To fetch the files yourself instead, put them all in one directory and point waifu draw at the manifest. They must stay side by side -- a manifest names its neighbours by file name and will not follow a path anywhere else.

hf download ling0322/libwaifu-illustrious-xl-v2.0 --local-dir illustrious-xl-v2.0
waifu draw -m illustrious-xl-v2.0/illustrious-xl-v2.0.yaml

Prompting

Onoma's card publishes no recommended prompt or sampler settings, so what the manifest suggests is what the model inherits rather than what its author advises -- the manifest says which is which, in a comment above the block.

Illustrious is a danbooru-tag model rather than one prompted with sentences: a prompt is written as <1girl/1boy/1other/...>, <character>, <series>, <artists>, <general tags>, usually behind a few quality tags. libwaifu's defaults -- 30 steps at CFG 5 -- are a reasonable place to start.

License

The weights are the author's; this repository only changes the container they are in. The license is inherited from the source and is not MIT like libwaifu itself: it is creativeml-openrail-m, which permits commercial use but attaches use-based restrictions that follow the weights into any derivative. See the source repository for the full terms and model card.