Anima-2.9B: Anima 2.9B: 40 Layers, Trained on 1.7 Million Images

Anima 2.9B is a text-to-image model specialized for generating animation, illustration, and non-photorealistic art. It expands the Transformer layers from 28 to 40 compared to the original Anima-based model, increasing the parameter count to approximately 2.9B. The added layers copy the weights of adjacent layers and initialize the output projection to zero, ensuring functional equivalence with the original Anima at initialization.

The model was further trained on 1.7 million additional animation and illustration samples, with a knowledge cutoff for the training data of July 2026. Training was conducted on a cluster of 8x 5080s using the Muon optimizer. Captions combined tags and natural language descriptions, generated using a combination of Gemini 3.1 Flash-Lite, Gemini 3.5 Flash-Lite, and Claude Sonnet 5.

Officially supported in ComfyUI and Forge-Neo, it can be used directly without installing custom nodes in the latest versions of ComfyUI (0.33.1 or higher). LoRA training is also supported. When writing prompts, appending series or copyright tags after character names helps reduce confusion. More detailed prompts improve background and detail rendering.

Recommended generation settings are the Euler or Res-multistep sampler combined with the sgm-uniform scheduler. Supported resolutions are 812x1216, 1152x1536, and 1536x1536, with 28–50 steps and a CFG scale of 3.5–5 being appropriate. Model weights are distributed under the CircleStone Labs Non-Commercial License and are available for non-commercial use only.