Anima Base v1.0 · 2B anime · text-to-image

Anima — anime text-to-image

A 2-billion-parameter anime illustration model by CircleStone Labs and Comfy Org. Trained on several million anime images plus ~800k non-anime artistic images (no synthetic data, anime knowledge cutoff September 2025). It reads Danbooru-style tags, plain captions, or a mix of both — and it is explicitly not a realism model.

Anima sample: anime girl in a neon-lit rainy street
768×768, 8 steps, CFG 4.0 — rendered locally on a Ryzen 780M iGPU via stable-diffusion.cpp.
Status of this Space Static page. The Gradio demo (app.py) is already in this repo but needs ZeroGPU, which requires a Hugging Face account at least 30 days old or a PRO plan. Flip sdk: static → sdk: gradio once GPU access is attached.
Runs anywhere bf16 weights are ~4 GB for the transformer plus 1.1 GB text encoder and 242 MB VAE — it fits on a 8 GB GPU, and even on CPU (slowly).
License Non-commercial (CircleStone Labs license). Check before any commercial use.

Architecture

TransformerCosmosTransformer3DModel · 28 layers · 16 latent channels · patch (1,2,2)
Text encoderQwen3Model (Qwen3-0.6B, 1024-dim)
Text conditionerAnimaTextConditioner + T5 tokenizer
VAEAutoencoderKLQwenImage (16-channel, same as Qwen-Image)
SchedulerFlowMatchEulerDiscreteScheduler (flow matching)
Base modelnvidia/Cosmos-Predict2-2B-Text2Image, finetuned for anime

Recommended settings

Resolution512² – 1536² (multiples of 32)
Steps / CFG30–50 steps, CFG 4–5 (base & aesthetic). Turbo: CFG 1, 8–12 steps.
Positive prefixmasterpiece, best quality, score_7, safe,
Negativeworst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration
Tag styleLowercase, spaces instead of underscores. Score tags are the exception (score_7). Prompt weighting works, but needs bigger weights than SDXL, e.g. (chibi:2).

Try it locally

pip install -U "diffusers>=0.43.0" transformers accelerate sentencepiece protobuf

python - <<'PY'
import torch
from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained("circlestone-labs/Anima-Base-v1.0-Diffusers")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")  # or "mps", "xpu", "cpu"

image = pipe(
    prompt="masterpiece, best quality, score_7, safe, 1girl, solo, upper body, "
           "looking at viewer, white hair, red eyes, black jacket, cyberpunk city, neon signs, rain",
    negative_prompt="worst quality, low quality, blurry, jpeg artifacts",
    width=1024, height=1024, num_inference_steps=32, guidance_scale=4.5,
).images[0]
image.save("anima.png")
PY
The 10 officially released variants (base, aesthetic v1.0 / 1.0b / 1.1, turbo v1.0 / 1.1) live in circlestone-labs/Anima as ComfyUI single-file checkpoints. Only base v1.0 is published in diffusers format, which is what this Space uses.

Links

Official LoRAsAnima-Official-LoRAs