Anima-Lightning

Anima-Lightning is a 4-step distilled text-to-image model derived from circlestone-labs/Anima. It is structurally identical to upstream Anima and keeps the same Anima/Cosmos Diffusers components, but uses distilled model weights.

The model is intended for anime, illustration, and other non-photorealistic art generation. It is not a realism-focused model.

This repository is not an official CircleStone Labs or NVIDIA release.

Architecture

The model is packaged in Diffusers modular format:

• Pipeline: AnimaModularPipeline
• Transformer: CosmosTransformer3DModel
• Text encoder: Qwen3Model
• Tokenizer: Qwen2Tokenizer
• Text conditioner: AnimaTextConditioner
• VAE: AutoencoderKLQwenImage
• Scheduler: FlowMatchEulerDiscreteScheduler

Installing and running

This model requires a Diffusers build that includes the Anima modular pipeline and Cosmos text-to-image components.

This is not the original 30-50 step Anima runtime. Use the distilled 4-step TDM runtime described in distilled_generation_config.json.

import torch
from diffusers import AnimaModularPipeline

repo_id = "YOUR_USERNAME/Anima-Lightning"

pipe = AnimaModularPipeline.from_pretrained(
    repo_id,
    torch_dtype=torch.bfloat16,
)
pipe.load_components(
    ["text_encoder", "tokenizer", "t5_tokenizer", "text_conditioner", "transformer", "scheduler", "vae"],
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

# Use the distilled runtime contract below. A normal pipeline call with
# num_inference_steps=40 and guidance_scale=5.0 is for upstream Anima, not this
# distilled checkpoint.
prompt = (
    "masterpiece, best quality, score_8, safe, anime illustration, "
    "solo shrine maiden standing in a rainlit courtyard, wide shot"
)

Generation settings

Use these defaults for the distilled model:

• Image size: 1024x1024
• Precision: bfloat16
• Max sequence length: 512
• Runtime CFG: 1.0
• Student steps: 4
• Total flow timesteps: 1000
• Timestep schedule: (999 -> 749), (749 -> 499), (499 -> 249), (249 -> 0)
• Scheduler family: FlowMatchEulerDiscreteScheduler

Important runtime notes:

• The distilled model uses one conditional student pass per step.
• Negative-prompt CFG from the upstream sampler is not used at inference.
• Upstream Anima's reference settings are 40 steps and CFG 5.0; those are useful for comparison, but not for running this distilled checkpoint.

Prompting

The model responds well to a mix of Danbooru-style tags and natural language. A practical prompt order is:

[quality/meta/safety tags], [subject], [character], [series], [artist], [scene and details]

Tips:

• Use lowercase tags and spaces instead of underscores, except for score tags such as score_7.
• Put safety and quality tags near the beginning of the prompt.
• Prefix artist tags with @.
• For natural language prompts, describe the character appearance and scene in complete sentences.
• For multiple characters, describe each character explicitly instead of only listing names.

Limitations

• This model is intended for anime, illustration, and stylized art. It is not a realism-focused model.
• It needs the distilled TDM runtime above. Running it as a normal full-step Anima checkpoint can produce incorrect output.
• Text rendering may be unreliable, especially for long phrases.
• Short or underspecified prompts can produce unwanted content or weak subject control.
• Behavior will differ from upstream Anima because this checkpoint uses distilled weights.

License

This model is licensed under the CircleStone Labs Non-Commercial License because it is derived from circlestone-labs/Anima. Anima is itself a derivative of nvidia/Cosmos-Predict2-2B-Text2Image, so NVIDIA's Open Model License Agreement may also apply to this derivative model.

See LICENSE.md and NOTICE.md before using, redistributing, or publishing this model. This summary is not legal advice.

See PUBLISH.md for the current Hugging Face CLI upload commands.

Built on Anima and NVIDIA Cosmos.