Anima-2.9B continues the Anima line by expanding model depth and fine-tuning on fresh anime/illustration samples so prompts for stylized, tag-rich art yield more consistent composition and detail. Rather than a fresh architecture, it grows capacity by adding transformer layers to a proven Anima base and focusing training on high-resolution and tag-mixed captions, making it a pragmatic upgrade for artists and hobbyist pipelines.
What Sets It Apart
- Architecture expansion: transformer depth expanded (28 → 40) to ~2.9B parameters via layer duplication and careful initialization, so the model preserves base behavior while increasing capacity.
- Dataset freshness and scale: ~1.7M additional anime/illustration samples with a training/knowledge cutoff in July 2026, improving coverage of newer styles and assets.
- Caption strategy: mixed tokenization of Danbooru-style tags and natural-language captions (no score tags in training), which favors tag-driven prompt patterns common in illustration workflows.
- Integration-first: packaged for ComfyUI and Forge-Neo workflows with practical generation recommendations (samplers, schedulers, resolutions) and an ecosystem-conscious release (derivative under CircleStone Labs Non-Commercial License).
Who It's For & Tradeoffs
Great fit if you are an artist, hobbyist, or ComfyUI user who needs a tag-aware anime/illustration model with up-to-date training data and easy integration into existing compositor pipelines. It is also useful for experimentation with higher-resolution outputs and prompt engineering using Danbooru/Gelbooru-style tags.
Look elsewhere if you need commercially licensed weights, photorealistic generation, or a model explicitly trimmed for low-VRAM consumer devices—the license is non-commercial (derivative) and the model remains specialized for non-photorealistic/illustrative outputs.
Practical notes
- Prompting: follow Anima conventions (quality tags, @artist prefix, explicit character/series tags and appearance descriptions). More detailed prompts produce better composition and background detail.
- Recommended generation settings (empirical): Euler or res-multistep samplers, sgm-uniform or linear-quadratic schedulers, 28–50 steps, CFG ~3.5–5, resolutions like 812×1216 or 1152×1536 for good tradeoffs.
- License & usage: released as a derivative under CircleStone Labs' Non-Commercial License; verify compatibility with your intended use before deploying.
