--- title: libwaifu-anima-turbo-v1.1 canonical_url: "https://www.modelscope.cn/models/ling0322/libwaifu-anima-turbo-v1.1" md_url: "https://www.modelscope.cn/models/ling0322/libwaifu-anima-turbo-v1.1.md" repository: ling0322/libwaifu-anima-turbo-v1.1 last_updated: 2026-09-22 license: other pipeline_tag: text-to-image-synthesis tasks: - text-to-image-synthesis base_model: - circlestone-labs/Anima base_model_relation: finetune parameters: 2.7B tensor_type: - F16 library_name: - lora - safetensors downloads: 7 stars: 1 tags: - text-to-image - anima - cosmos-predict2 - libwaifu - not-for-all-audiences --- # libwaifu-anima-turbo-v1.1 > libwaifu-anima-turbo-v1.1 - ling0322 在 ModelScope 开源的模型。Anima Turbo v1.1 for libwaifu ling0322/libwaifu-anima-turbo-v1.1 是 ModelScope 魔搭社区上的 2.7B 参数text-to-image-synthesis模型，采用 other 许可，基于 circlestone-labs/Anima 构建。 - **Repository**: ling0322/libwaifu-anima-turbo-v1.1 - **License**: other - **Tasks**: text-to-image-synthesis - **Parameters**: 2.7B - **Base model**: circlestone-labs/Anima - **Tags**: text-to-image, anima, cosmos-predict2, libwaifu, not-for-all-audiences - **Downloads**: 7 - **Stars**: 1 - **Last updated**: 2026-09-22 Source: https://www.modelscope.cn/models/ling0322/libwaifu-anima-turbo-v1.1 --- # Anima Turbo v1.1 for libwaifu Anima-Turbo v1.1 converted to the format [libwaifu](https://github.com/ling0322/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, and the largest weight in this model is 123 against a ceiling of 65504. 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. Converted from [circlestone-labs/Anima](https://huggingface.co/circlestone-labs/Anima): | source file | sha256 | |---|---| | `split_files/diffusion_models/anima-turbo-v1.1.safetensors` | `fba11953276b57edf59d1dc4f1857ac05aa079c56f982b4d7c20298d57d3f7eb` | | `split_files/text_encoders/qwen_3_06b_base.safetensors` | `cd2a512003e2f9f3cd3c32a9c3573f820bb28c940f73c57b1ddaa983d9223eba` | | `split_files/vae/qwen_image_vae.safetensors` | `a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f` | **This is the turbo release**, which its author distilled for **8-12 steps at CFG 1** -- no classifier-free guidance, so a step costs one model evaluation rather than two. The aesthetic and base releases are the same architecture and want 30-50 steps at CFG 4-5 instead; drawing one with the other's numbers is not an error anything reports, it is a picture that comes out burnt or unfinished. The manifest carries this release's numbers, so the defaults are right without being typed. 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. | file | size | what | |---|---|---| | `anima-turbo-v1.1.yaml` | 3 KB | the manifest: what the model is, and which files it is made of | | `anima-turbo-v1.1.qwen3.tokenizer.json` | 10.9 MiB | Qwen3's vocabulary, for the text encoder | | `anima-turbo-v1.1.t5.tokenizer.json` | 2.3 MiB | T5's, for the adapter -- see below | | `anima-turbo-v1.1-00001-of-00003.safetensors` | 1.87 GiB | | | `anima-turbo-v1.1-00002-of-00003.safetensors` | 1.87 GiB | | | `anima-turbo-v1.1-00003-of-00003.safetensors` | 1.35 GiB | | 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:turbo`, so there is nothing to download by hand: ```bash waifu draw -m anima:turbo ``` `anima:turbo:v1.1` names this release and keeps meaning it; `anima:turbo` follows whatever is current. ```bash hf download ling0322/libwaifu-anima-turbo-v1.1 --local-dir anima-turbo-v1.1 waifu draw -m anima-turbo-v1.1/anima-turbo-v1.1.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 model is trained on Danbooru-style tags, natural language captions, and combinations of the two, so either works and they can be mixed. The author's own conventions, from the source card: - Lowercase tags, spaces rather than underscores. Score tags are the only ones with underscores. - Recommended prefix: `masterpiece, best quality, score_7, safe, ` - Recommended negative: `worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration` - Tag order: `[quality/meta/year/safety] [1girl/1boy/...] [character] [series] [artist] [general]` - **Artist tags need an `@` in front** -- `@nnn yryr`. Without it the effect is very weak. - Prompt weighting works but wants higher weights than SDXL: `(chibi:2)`. - Where a tag differs between Danbooru and Gelbooru, prefer the Gelbooru spelling. - Resolutions between 512x512 and 1536x1536, in multiples of 16. The manifest carries the prefix, the negative, three sizes and this release's steps and guidance, so the boxes open on them. ## Conversion ```bash python tools/anima_exporter.py \ -dit split_files/diffusion_models/anima-turbo-v1.1.safetensors \ -text split_files/text_encoders/qwen_3_06b_base.safetensors \ -vae split_files/vae/qwen_image_vae.safetensors \ -output anima-turbo-v1.1.safetensors \ -part-size 2GB ``` The VAE is a video autoencoder whose convolutions are five-dimensional. For a single image the temporal extent is 1 throughout, and the causal convolution then collapses exactly onto a 2-D one, so the export folds every `Conv3d` to a `Conv2d` over its last temporal slice. That folding is exact rather than approximate: the decoder here agrees with ComfyUI's, running the real `Conv3d`, to 2.1e-06 on these weights. ## 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 the [CircleStone Labs Non-Commercial License](https://huggingface.co/circlestone-labs/Anima/blob/main/LICENSE.md), which permits non-commercial use only. **The restriction is on the model, not on the pictures**: the author states that generated images may be used commercially -- selling images, paid commissions, assets for a paid product -- while hosting the model behind a paid API, putting it on a paid generation platform, or embedding the weights in a monetized product need a separate license (`tdrussell@circlestone.ai`). Anima is also a Derivative Model of [nvidia/Cosmos-Predict2-2B-Text2Image](https://huggingface.co/nvidia/Cosmos-Predict2-2B-Text2Image) and is subject to the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/) insofar as it applies to Derivative Models. Built on NVIDIA Cosmos. See the [source repository](https://huggingface.co/circlestone-labs/Anima) for the full terms and the author's own model card.