Animagine XL 4.0: What’s New and How to Use It

Animagine XL 4.0 (also called Anim4gine) is the current main line of the Animagine XL series.

If you only remember one sentence, remember this:

4.0 is a new training run from SDXL 1.0, not just a small patch on 3.1 — and it is tuned for anime tags, quality control tags, and more natural light and texture.

This article explains those claims in plain language and shows a simple starting setup. It is based on public model cards and author posts — not private secrets.

---

What is 4.0?

| Fact | Public detail | |------|----------------| | Full name | Animagine XL 4.0 / Anim4gine | | Base | Stable Diffusion XL 1.0 (retrain) | | Team | Cagliostro Research Lab | | Training images | about 8.4 million anime-style images | | Knowledge cutoff | about 2025-01-07 | | Training compute | about 2650 GPU hours | | Public model hubs | Civitai model 1188071, plus public model cards on major hubs | | License (public pages) | CreativeML Open RAIL++-M |

Public release notes place the initial 4.0 drop around late January 2025 (weights/changelog and the announcement post are a few days apart).

---

What changed compared with 3.1?

Official comparison language is easy to understand:

1. More natural and artistic

4.0 is described as more artistic, more natural, and more flexible in style. 3.1 can look more stiff or “2.5D.”

2. Better prompt following

Public notes say 4.0 sticks closer to what you asked for. That is helpful when your tags are careful and ordered well.

3. Stronger light, texture, and depth

Author comparisons call out better lighting, surface texture, and depth of field. In plain words: scenes can feel less flat.

4. Built from the SDXL base again

This is the big technical point. 4.0 is not “only more training on top of 3.x.” The team went back to SDXL 1.0 and trained the anime model with a huge new data pass.

---

Meet Opt and Zero

In February 2025, public updates added two important siblings:

Opt (everyday generation)

Opt is refined with extra data for:

• more stable results
• better anatomy goals
• cleaner denoising
• nicer color

If you just want to generate images, Opt is usually the practical default.

Zero (for trainers)

Zero is a pretrain-style base. It is aimed at people who will add LoRAs or continue fine-tuning.

Rule of thumb:

• drawing pictures → try Opt
• training your own add-on → look at Zero

---

How to write prompts for 4.0

Recommended tag order

Public guidance uses a Danbooru-like order:

1. 1girl / 1boy / 1other
2. character name
3. series / copyright tag
4. rating tag (safe, sensitive, …)
5. other tags in any order
6. quality boost tags at the end

A common quality ending is:

masterpiece, high score, great score, absurdres

Do not put those quality words at the very beginning if you care about artist-style tags. Public notes say front-loading quality tags can interfere.

Character + series together

If you name a character, also name the series. Training prepared character and series as pairs. Leaving the series out can make identity less reliable.

Score tags matter

The score family (high score, great score, good score, average score, bad score, low score) is described as a strong quality control. Many people use high/great score together with masterpiece-style tags.

Year tags for era vibes

Tags like year 2010 or year 2024 can push older or newer anime looks. This is optional — use it when the time period is part of your idea.

---

Recommended settings (public 4.0 guidance)

These are starting points, not laws. Change one thing at a time.

| Setting | Suggested start | |---------|-----------------| | Sampler | Euler a | | Steps | 25–28 (many examples use 28) | | CFG | 4–7 (often 5) | | Resolution | normal SDXL sizes (for example around 1024-class widths/heights) |

Older announcement notes sometimes listed slightly different numbers (for example CFG 6 / 25 steps). Prefer the newer model card and optimization guide when they disagree.

---

When 4.0 may struggle

Public limitation notes include:

• prompts that are only long natural language
• very complex hands
• readable text drawn inside the picture
• characters newer than the knowledge cutoff
• crowded multi-character scenes
• extreme resolutions far outside normal SDXL practice

The training story also focuses more on who is in the picture than perfect long-term style matching. If you need one locked art style forever, you may still want LoRAs or careful prompting.

---

License reminder (short)

Public 4.0 pages list CreativeML Open RAIL++-M. That usually means you can use, change, and share the model under the license rules — including many commercial cases — while harmful and illegal uses are forbidden.

Read the full text before any business use. Community gallery sites may also show extra permission checkboxes for sample images; those are separate from the model license.

---

Try it and read more

• Features, history, and future of Animagine XL
• Playground — switch tabs between v4.0, v3.1, and v3.0

---

Quick recap

1. 4.0 is a full anime SDXL retrain with huge data and a 2025 knowledge window.
2. Opt is the friendly everyday variant; Zero is for further training.
3. Write tags in a clear order and put quality/score words last.
4. Start with Euler a, ~28 steps, CFG ~5, then adjust slowly.