Anima-3.8B: 52-Block Anime Model With Qwen3.5 4B and ComfyUI Nodes

Anima-3.8B ("Pro52 / Qwen3.5 Edition") is a community release by lylogummy that rebuilds the Anima anime model around a 52-block DiT and a Qwen3.5 4B text encoder, giving the Anima visual style stronger natural-language understanding (Hugging Face).

Multi-character comparison from the Anima-3.8B release

What's in the release

The repository ships three model files plus a ready-made two-pass workflow:

• Anima-3.8B.safetensors — the 52-block diffusion model, placed in ComfyUI/models/diffusion_models/
• qwen35_4b.safetensors — the Qwen3.5 4B text encoder, placed in ComfyUI/models/text_encoders/
• Anima-3.8B-expanded_adapter.safetensors — the progressive-cross adapter that blends the Qwen3.5 conditioning with the native Anima conditioning
• workflow_2_pass.json — a two-pass workflow that uses the native CLIP conditioning for the high-resolution pass to avoid artifacts

ComfyUI usage

Anima-3.8B is used through the dedicated comfyui-anima-3-8B custom node (a forge-anima-3.8B port is also available):

1. Clone comfyui-anima-3-8B into ComfyUI/custom_nodes/ and install its requirements.
2. Use AnimaQwen35Loader to load qwen35_4b.safetensors, and AnimaQwen35UnifiedPrompt to build conditioning from the Qwen3.5 model, the native Anima CLIP, and the adapter.
3. The recommended starting point is 832x1216 (or any other 1MP resolution), adapter strength 1.0, CFG 7-8, and 28-50 steps with a Beta scheduler.

The prompt format mixes Danbooru-style tags with a Description: natural-language section, and for multi-character scenes each subject should get its own sentence to keep them distinct. Since the model was trained at low resolution, the native CLIP conditioning is recommended for the second high-resolution sampler pass.