libwaifu-one-obsession-v24

One Obsession v24 for libwaifu

One Obsession v24 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.

By maxfeifei8, converted from their own release, One obsession v24, file oneObsession_v24.safetensors, sha256 2fae33d7f0d23920d145897827e6015ac098a2b90cb682d07846ae762213100b -- the hash the author publishes, so these weights are verifiably theirs. It is an Illustrious fine tune, so it is prompted with danbooru tags rather than with sentences.

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.

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:obsession, so there is nothing to download by hand:

waifu draw -m sdxl:obsession

sdxl:obsession:v24 names this release and keeps meaning it; sdxl:obsession follows whatever is current.

hf download ling0322/libwaifu-one-obsession-v24 --local-dir one-obsession-v24
waifu draw -m one-obsession-v24/one-obsession-v24.yaml

Keep the files side by side -- a manifest names its neighbours by file name and will not follow a path anywhere else.

Prompting

Everything the manifest suggests is the author's, read out of the generation metadata of their own sample images rather than out of the description. The two disagree, and the samples win: the description recommends 25 ~ 35 steps at CFG 3 ~ 6, while every one of the ten samples was drawn at CFG 4.5, in 20 ~ 24 steps, with Euler a, at a portrait size. The quality tags suggested are the ones common to all ten.

The author's own advice, from the model page:

You can make a good picture without lora, without complicated tags, and it is extremely friendly to beginners. The drawing tags are hidden in the sample images' prompts -- finding the artist strings is part of the fun.

Their recommended resolutions are 1024x1536, 832x1216, 896x1152, 768x1344 and 640x1536 -- portrait, every one of them, which is why the manifest opens on 832x1216 rather than on a square.

Conversion

python tools/sdxl_exporter.py \
  -checkpoint oneObsession_v24.safetensors \
  -output one-obsession-v24.safetensors \
  -part-size 2GB

License

The weights are the author's; this repository only changes the container they are in, and the terms are theirs rather than libwaifu's MIT. On the model page the author:

• requires credit -- so: this is maxfeifei8's work
• allows derivatives, but requires that they carry these same terms
• permits no general commercial use -- only selling the images it generates, and running it on Civitai

See the model page for the terms as they stand, which are the ones that govern.