About Anima

Anima is trained on Danbooru and Gelbooru and understands a mix of tags and plain English.
Use tags for quality, artists, background and scene. Use a sentence or two for the characters, their pose and how they interact.

The e6 dataset was not included in its training, so it is not very good at furry stuff.

Do not put a weight on artist tags, as in (@artist:1.2). Anima recognises an artist by the @ directly in front of the name. When a tag is put in brackets to weight it, the webui cuts it out of the prompt and reads it on its own, and the model no longer sees it the same way. The style then shifts even at a weight of 1.0, and from about 1.2 upwards bodies can start to break. Write artist tags plain: @artist.

Anima models come in three parameter sizes:
base (28 layers)
2.9B (40 layers)
3.8B (52 layers)

Each model's table below lists its parameter size, check which variant a LoRA was made for before you download it.

• A LoRA made for a smaller variant also works on a larger model, for example a base LoRA on a 2.9B model.
• A LoRA made for a larger variant has no effect on a smaller model, and the image comes out as if no LoRA was used.

Turbo gives more consistent and more refined results than the base model: fewer broken bodies and cleaner details. Most merges do the same, Turbo or not. That carries over to LoRAs, which also behave better on these models than on base. Turbo requires CFG to be set to 1.0, which disables the negative prompt.

Hires fix can cause trouble on Anima. In our tests the result turned soft and blotchy from about 1.5x upscale, whatever upscaler or settings we tried. Generating directly at a larger size worked better: 1536x1920 and 1792x2240 came out sharp on the merges, while the base model went soft at 1792x2240.

Models built on Anima 3.8B can use a bigger text encoder, Qwen3.5 4B, alongside the standard qwen_3_06b_base. It needs an extension: forge-anima-3.8B in Forge Neo, the custom node comfyui-anima-3-8B in ComfyUI.
The text encoder is the part that reads your prompt. The bigger one is supposed to follow long, complicated prompts better: several characters each doing something different, or "this on the left, that on the right". It takes more VRAM and is slower.

This does not apply to the models listed here as none of them support it.

All images on this page were generated with the text encoder qwen_3_06b_base and the VAE qwen_image_vae.