Anima-2.9B: Layer-Expanded Anime Model for ComfyUI

Overview

Anima-2.9B by Gazingstars123 is built on top of the Anima base model and continues its anime, illustration, and non-photorealistic art focus. The current release is Anima-2.9B-preview-v1, an initial preview while pretraining on a broader 10M-sample general dataset is still in progress.

Architecture: Layer Expansion

The core change is a transformer depth expansion from 28 to 40 layers:

• Each new layer is created by deep-copying its neighboring layer's weights, inserted in an interleaved pattern with zeroed-out output projections.
• This makes the expanded model functionally identical to Anima base at initialization, so the fine-tune starts from a clean slate.
• The result grows the model to roughly 2.9B parameters.

Training used the Muon optimizer on an 8x RTX 5080 cluster, with roughly 70% of compute spent at 1024px resolution. Captioning mixes Danbooru-style tags with natural language, generated with Gemini 3.1 Flash-Lite, Gemini 3.5 Flash-Lite, and Claude Sonnet 5.

ComfyUI Support

Anima-2.9B ships with a dedicated ComfyUI-Anima-2.9B custom node. The node patches the ComfyUI model loader at startup so the expanded 40-layer checkpoint loads correctly, and it works with any Anima-based model:

cd ComfyUI/custom_nodes
git clone https://github.com/gazingstars123/ComfyUI-Anima-2.9B.git

No workflow nodes are required. Place Anima-2.9B-preview-v1.safetensors in ComfyUI/models/diffusion_models/ and load it like any other checkpoint.

Prompting Tips

Anima-2.9B follows the same prompting conventions as the base Anima model:

• Use Danbooru/Gelbooru-style tags in lowercase, with spaces instead of underscores.
• Include quality tags, year/period tags, artist tags, and character tags (1girl, 1boy).
• Combine tags with natural language descriptions for best results.

Links

• Model weights: Gazingstars123/Anima-2.9B on Hugging Face
• ComfyUI node: gazingstars123/ComfyUI-Anima-2.9B
• Base model: circlestone-labs/Anima