Anima is a 2 billion parameter text-to-image AI model created via a collaboration between CircleStone Labs and Comfy Org, built on the NVIDIA Cosmos architecture. It is focused mainly on anime concepts, characters, and styles, and was trained on several million anime images alongside roughly 800k non-anime artistic images.
Notably, Anima features extremely good prompt adherence and artist style adherence. The model is trained on a mix of standard Danbooru-style tags and natural language captions, allowing users to combine both methods seamlessly. It is intentionally designed for illustrations and artistic images, and does not excel at realism or lengthy text rendering.
Prompting and Tag Usage
Artist Tags: To trigger a specific artist's style, the artist tag must be prefixed with an "@" symbol (e.g., "@artist_name"). The styling effect will be very weak without this prefix.
Quality Tags: The model understands both human score-based tags (e.g., masterpiece, best quality, worst quality) and PonyV7 aesthetic model tags (e.g., score_9, score_5). These can be used separately or together.
Time Period Tags: You can use specific year tags (e.g., year 2025) or general period tags (e.g., newest, recent, old).
Tag Order: A recommended prompt structure is: [quality/meta/year/safety tags] [1girl/1boy/etc] [character] [series] [artist] [general tags].
Natural Language: When using natural language instead of pure tags, aim for at least two descriptive sentences. Mixing tags and natural language in an arbitrary order works well.
