Anima

Highlights

• 2B parameters, anime and illustration oriented; trained on several million anime images plus roughly 800K non-anime art images, with no synthetic data
• CFG 1 and 8–12 steps are enough to produce an image; the official recommendation is to start from the Turbo version
• The text encoder is only Qwen3-0.6B and the VAE reuses Qwen-Image VAE, so VRAM use is very low
• Native ComfyUI support, with official Base v1 and Preview templates

Scope

• Good for: anime characters, fan art, illustration styles
• Good for: low-VRAM setups and cases that stack many LoRA
• Not for: photorealistic photos (the official notes state it does this poorly)

Tips

• The anime LoRA ecosystem is very large; filter by the Anima base model when hunting for styles
• Keep CFG at 1; on distilled and fine-tuned models like this, adjusting CFG gains very little
• Text rendering is weak; add text in post

Cautions

• Non-commercial license; commercial use requires separate authorization
• 1.0, v1.0b, Aesthetic and other versions differ noticeably; check the suffix when downloading

Parameters

Steps

8–12

CFG

1

Text encoder

Qwen3-0.6B (qwen_3_06b_base)

VAE

Qwen-Image VAE

Weights

anima-base-v1.0.safetensors (several fine-tuned variants also available)

Deployment

16GB verdict

No pressure at all: 2B weights plus a 0.6B encoder, so 16GB can host multiple LoRA and upscaling at the same time.

VRAM notes

Coexists well with other models; generation speed is among the fastest on this list.

Suggested starting point

Start with the official Base v1 template: CFG 1, 10 steps, 1024².

Comparison

A 2B anime-specific base model that nonetheless drew 1.47M downloads in the past 30 days — higher than any general-purpose model. The reason is that it makes anime good enough and cheap: CFG 1, 8–12 steps, only a 0.6B text encoder, and room for a stack of LoRA on 16GB. For photorealism, rule it out entirely.