MiaoMiao Harem Anima v1.6 for libwaifu

MiaoMiao Harem Anima v1.6 converted to the format libwaifu reads, so it can draw from it directly. The weights are the original ones, narrowed from bfloat16 to float16 -- a narrowing of range and not of precision, since every bfloat16 value inside float16's exponent range converts exactly.

Anima is not a Stable Diffusion model. It is a Cosmos-Predict2 diffusion transformer with a Qwen3-0.6B text encoder and the Qwen-Image VAE, published as three separate files; this repository is those three exported together as one model. It is a rectified-flow model, not an epsilon-prediction one. This is the aesthetic-style release -- 28-50 steps at CFG 3.5-5.0, unlike the distilled turbo releases that draw at 8-12 steps and CFG 1.

Converted from MiaoMiao Harem, version Anima_1.6 by MIAOKA:

The text encoder and VAE hashes are identical to the ones circlestone-labs publishes alongside every Anima release -- this checkpoint bundles the stock shared files beside its own fine-tuned denoiser, and only the denoiser differs.

This model draws from a prompt only. The VAE's encoder is not part of this export, so waifu draw -image is not available for it.

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 tokenizers are the tokenizer.json of the two vocabularies this model is prompted through.

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.

Two tokenizers, because the prompt is tokenized twice. Qwen3-0.6B encodes it into hidden states; the same prompt is tokenized again with T5, and those ids are the query stream a six-block adapter runs against those hidden states. Only the adapter's output reaches the denoiser. This is inherited from Cosmos-Predict2, which was pretrained against T5, and it is why prompt weighting on this model has to be pushed harder than on SDXL -- the weights multiply the query stream rather than the encoder's output.

Usage

waifu knows this model as anima:miaomiao, so there is nothing to download by hand:

waifu draw -m anima:miaomiao

anima:miaomiao:v1.6 names this release and keeps meaning it; anima:miaomiao follows whatever is current.

hf download ling0322/libwaifu-miaomiao-harem-v1.6 --local-dir miaomiao-harem-v1.6
waifu draw -m miaomiao-harem-v1.6/miaomiao-harem-v1.6.yaml

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

Prompting

The card's own template opens with a quality-tag prefix in the Pony/Illustrious rating-tag style (score_7, score_9, sensitive) rather than plain danbooru tags:

best quality, score_7, score_9, sensitive, very aesthetic, ultra detailed, fair skin, high contrast, {your prompt}

and its negative prompt:

worst quality, low quality, score_1, score_2, score_3, artist name, shiny skin

The card also notes a simple prompt without the quality prefix gets a respectable result on its own; the prefix is what pushes it further. The manifest carries this prompt, this negative, three sizes and this release's steps and guidance, so the boxes open on them.

Conversion

python tools/anima_exporter.py \
  -dit  miaomiaoHarem_anima16.safetensors \
  -text miaomiaoHarem_anima16_txt.safetensors \
  -vae  qwen_image_vae.safetensors \
  -output miaomiao-harem-v1.6.safetensors \
  -part-size 2GB

License

The weights are the author's; this repository only changes the container they are in.

The model's page states it is only released on Civitai and SeaArt.Ai, and that releases on other platforms are not authorized by the author. This mirror exists because the author, MIAOKA, allowed the libwaifu maintainer to make this specific conversion, after this project and both hubs it would be published to were named to them. That exchange was a private one; this card states its outcome rather than reproducing it.

The terms, as given: non-commercial use, with attribution to MIAOKA and to the original model page. This is narrower than circlestone-labs' own Anima licence (which this checkpoint is a fine tune of) and narrower than what most of this project's other mirrors carry -- do not treat it as a general-purpose licence the way the others are.

Anima is also a Derivative Model of nvidia/Cosmos-Predict2-2B-Text2Image and is subject to the NVIDIA Open Model License Agreement insofar as it applies to Derivative Models. Built on NVIDIA Cosmos.

See the source page for the author's own model card and future versions.