Anima Preview 3 - Text to Image

Anime and illustration generation with Anima Preview 3, a 2 billion parameter model trained on millions of anime images and hundreds of thousands of artistic works.

Write a prompt describing what you want, and Anima turns it into an illustration. The model understands Danbooru-style tags, natural language captions, or a mix of both. It handles characters, scenes, and artistic styles well, but it is not built for photorealism.

Best at 1MP resolution (1024x1024 or similar). Runs in 30 steps by default.

How do you prompt Anima Preview 3 for text to image?

Write at least two sentences. Use Danbooru tags, natural language descriptions, or combine both. Include quality tags in your positive prompt, and add safety and quality exclusions to your negative prompt. Longer, more detailed prompts give you more control over the output.

Positive Prompt Describe the scene, character, and style you want. You can use natural language ("A teenage wizard floating above a moonlit library") or Danbooru-style tags ("1girl, long hair, school uniform, cherry blossoms"). Mixing both works too. Short, vague prompts tend to give unpredictable results. Add detail to steer the output.

Negative Prompt Default is set to filter out low-quality outputs: "worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts." Add anything specific you want to avoid.

CFG (Guidance Scale) Default is 4. Want the image to follow your prompt more closely? Go higher, up to about 6 or 7. Going too high can burn the image and introduce artifacts, especially with certain samplers. Start at 4 and nudge up if the output feels too loose.

Steps Default is 30. That gives solid results for most prompts. Going lower speeds things up but can lose fine detail. Going above 30 has diminishing returns for this model.

Sampler Default is er_sde. This gives a neutral style with flat colors and sharp lines. Want softer, thinner linework? Try euler_a. Want more variety and creative interpretations? Try dpmpp_2m_sde_gpu, though it can get unpredictable with some prompts.

Shift Default is 3. This controls the noise schedule through ModelSamplingAuraFlow. Most users can leave it at the default. Lower values can change how the model handles detail at different stages of generation.

Seed Lock the seed to reproduce a specific result. Set it to random to explore variations.

What is Anima Preview 3 good for?

Anima Preview 3 is built for anime illustrations, character art, and artistic imagery. It handles a wide range of anime styles and non-photorealistic content at 2 billion parameters, making it fast for its quality level. It is not designed for realism or complex text rendering.

Character design and concept art are where this model shines. It knows a large vocabulary of anime characters, series, and artist styles, trained on data up to September 2025. Multi-character scenes work better when you describe each character's appearance instead of relying on names alone.

It also handles non-anime artistic styles. Painterly looks, illustrations, and stylized scenes all work. For photorealism or detailed text in images, use a different model.

This is a preview checkpoint. The model is still in training. Fine details and overall aesthetics will improve in the final release. High-resolution generation above about 2MP can start breaking down.

Anima is developed by CircleStone Labs in collaboration with Comfy Org.

FAQ

What resolution should I use with Anima Preview 3? Stick to about 1MP total. That means 1024x1024 for square, or 896x1152 and 1152x896 for portrait and landscape. Going above 2MP can cause the image to break down. High-resolution training is still in progress for the final release.

Does Anima Preview 3 support LoRAs? Yes. Community LoRAs trained on Anima are already available on CivitAI and HuggingFace. The model's architecture supports LoRA fine-tuning for custom characters, styles, and concepts.

What sampler works best with Anima Preview 3? er_sde is a good default for clean, sharp anime lines. euler_a gives softer results and handles higher CFG without burning. dpmpp_2m_sde_gpu adds more variety but can get unpredictable. Try a few to find what fits your style.

Can Anima Preview 3 generate realistic images? No. The model is trained on anime and artistic images. It does not produce photorealistic output, and that is by design. For realism, use a different model like Flux or Stable Diffusion 3.5.

How to run Anima Preview 3 online? You can run Anima Preview 3 online through Floyo. No installation, no setup. Open the workflow in your browser, write your prompt, and hit run. Free to try.