Downloads · 30 days
22
32% of all-time downloads
jlweights/Anima-Base-MLX-Serve
Anima-Base-MLX-Serve is a text-to-image model from jlweights. Use it when you need an image from a text prompt. It is set up for mlx-serve. The card lists the license as other.
circlestone-labs/Anima (base-v1.0 variant) repackaged for mlx-serve's native Zig + MLX engine — anime/illustration text-to-image and image-to-image, generated locally on Apple Silicon, no Python/ComfyUI required.
Downloads · 30 days
22
32% of all-time downloads
All-time downloads
68
Public
Repo size
5.6 GB
Likes
0
Public
Click a slice to open those files.
.safetensors5.6 GB · 100%
From the Hugging Face model README
circlestone-labs/Anima (base-v1.0 variant) repackaged for mlx-serve's native Zig + MLX engine — anime/illustration text-to-image and image-to-image, generated locally on Apple Silicon, no Python/ComfyUI required.
NVIDIA Cosmos-Predict2 2B DiT + a 6-block Perceiver-style LLMAdapter,
conditioned by a Qwen3-0.6B text encoder, decoded by the Qwen-Image VAE.
Base schedule: 32 steps, CFG 4.5 (full classifier-free guidance — an
unconditional forward runs every step, unlike Turbo's single-forward
shortcut). Weights are copied byte-for-byte from the upstream ComfyUI split
files — no quantization yet (runs f32 end-to-end).
| File | Contents |
|---|---|
transformer.safetensors | Cosmos-Predict2 DiT + LLMAdapter |
text_encoder.safetensors | Qwen3-0.6B |
vae.safetensors | Qwen-Image VAE (encoder + decoder) |
tokenizer/ | Qwen BPE tokenizer |
t5_tokenizer/ | T5 SentencePiece unigram vocab (LLMAdapter query tokens) |
Download MLX Core.app, open the Image tab and pick Anima from "On This Mac" once this repo is downloaded into your models folder.
Over HTTP:
mlx-serve --serve
curl http://127.0.0.1:11234/v1/images/generations \
-H 'Content-Type: application/json' -o apple.png -d '{
"model": "Anima-Base-MLX-Serve",
"prompt": "a red apple on a wooden table, simple illustration",
"size": "512x512",
"steps": 32
}'
img2img (mode:"variation" + image + strength) and runtime LoRA
(lora_paths/lora_scales) are both supported. Instruction editing is
not — this checkpoint has no edit training.
Rebuild from the upstream repo with
scripts/convert_anima_weights.py.
Weights are covered by the CircleStone Labs Non-Commercial License
(see LICENSE.md) — non-commercial use only; this repository is itself a
Derivative under that license, redistributed under the same terms.
The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.
This repository modifies the CircleStone Model (repackaged into mlx-serve's flat safetensors layout; no weight values are changed).