Hello, Alchemists~

After some thorough investigations, we finally came to conclusion that we’re making several changes the guideline for Animagine-XL 4.0 a.k.a. Anim4gine. These changes are made because of our oversight on preparing and testing the former guideline, thus the problems arose. But, before getting into the guideline itself, we have a bit of announcement. Let’s just get straight to it, shall we? 👀

Anim4gine Model Update#

During the investigation, not only we looked for solving the issues, we also made a bit of fine-tune to Anim4gine. We have released two versions of Anim4gine: Opt and Zero. Opt (short for Optimized) was trained with an additional dataset and further refined to enhance its performance, making it more optimal for general use. Zero, on the other hand, is the pretrained base model, making it ideal for LoRA training and finetuning. You may notice in the following samples that this version has crisper line art and more solid background than the base’s overexposed lighting in the background.

Generated with Animagine XL 4.0-Opt

Generated with Animagine XL 4.0-Opt

You can access these models on our Huggingface repository right here:

Opt: https://huggingface.co/cagliostrolab/animagine-xl-4.0 (same as initial release)

Zero: https://huggingface.co/cagliostrolab/animagine-xl-4.0-zero

So, do you want to generate better result with Anim4gine? Without further ado, let’s deep dive in to this in-depth guideline! And don’t worry, the following guideline can also retroactively applied to the Animagine-XL 4.0~

Quality Tags Placement#

First things first, the order of tags is crucial in this model. This has been the case since versions 3.0 and 3.1, as we trained the model with tags like ‘1girl,’ character name, series name, and everything else in any order. In the 4.0 release, we overlooked this and thought that placing quality tags at the start of the prompt would work. While the results were good, the outputs turned out inaccurate at times because quality tags were never placed at the front during training. As a result, some concepts and art styles didn’t always appear correctly.

At this point, we strongly recommend following our suggested prompt guidelines by placing tags in the correct order.

Now, let’s dive into the in-depth guide and comparison!

The summary can be seen in the image for the prompt guideline.

First of all, we would like to recommend you to put enhancement or quality tags at the end of your prompts instead of front (like we suggested beforehand). By putting enhancement or quality tags in front of your prompts, the artist tags can’t be triggered. You can take a look at these examples below.

1girl, solo, long hair, grey hair, green eyes, white shirt, collared shirt, open mouth, v, red bowtie, field, wind, safe, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

masterpiece, high score, great score, absurdres, 1girl, solo, long hair, grey hair, green eyes, white shirt, collared shirt, open mouth, v, red bowtie, field, wind, safe

Generated with Animagine XL 4.0-Opt

1girl, solo, torino aqua, long hair, grey hair, green eyes, white shirt, collared shirt, open mouth, v, red bowtie, field, wind, safe, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0

masterpiece, high score, great score, absurdres, 1girl, solo, torino aqua, long hair, grey hair, green eyes, white shirt, collared shirt, open mouth, v, red bowtie, field, wind, safe

Generated with Animagine XL 4.0

1girl, solo, torino aqua, long hair, grey hair, green eyes, white shirt, collared shirt, open mouth, v, red bowtie, field, wind, safe, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

masterpiece, high score, great score, absurdres, 1girl, solo, torino aqua, long hair, grey hair, green eyes, white shirt, collared shirt, open mouth, v, red bowtie, field, wind, safe

Generated with Animagine XL 4.0-Opt

Rating Tags in Prompt#

By putting rating tags (safe, sensitive, nsfw, or explicit) inside your prompt list, you have bigger chance to make your generated images better suited to the rating you’re aiming for. Putting safe rating tag will make the model more likely to generate safe for work image, sensitive for more revealing clothings. NSFW tag, like the name suggest (”not safe for work”), can yield even more revealing clothes/nudity, while explicit tag can yield sexual act. Take a look at these examples below (and of course, for an obvious reason, we cannot provide examples for NSFW tag and Explicit tag.)

1girl, solo, blue archive, long hair, grey hair, red eyes, large breasts, white shirt, collared shirt, long sleeves, halo, v, skirt, outdoors, ruins, standing, sensitive, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0

1girl, solo, blue archive, long hair, grey hair, red eyes, large breasts, white shirt, collared shirt, long sleeves, halo, v, skirt, outdoors, ruins, standing, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0

1girl, solo, blue archive, long hair, grey hair, red eyes, large breasts, white shirt, collared shirt, long sleeves, halo, v, skirt, outdoors, ruins, standing, sensitive, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

1girl, solo, blue archive, long hair, grey hair, red eyes, large breasts, white shirt, collared shirt, long sleeves, halo, v, skirt, outdoors, ruins, standing, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

Don’t Forget: Put the Series Tag!#

So, you want to generate certain character(s) from games or animanga series? Don’t forget to put the series/copyright tag after the character tag! Our dataset is prepared with series tags alongside character tags. If you want to generate the character and want to pertain their appearance from the source, it is mandatory to put series tag inside your prompt list.

But, what if you want to generate your own original character? Well, you can skip this step and just go ahead with the general tags right after gender tag~

1girl, oyama mahiro, onii-chan wa oshimai!, long hair, pink hair, red eyes, white shirt, oversized clothes, open mouth, yawning, on bed, bedroom, sensitive, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

1girl, oyama mahiro, long hair, pink hair, red eyes, white shirt, oversized clothes, open mouth, yawning, on bed, bedroom, sensitive, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

Positioning Artist Tags#

The placement of artist tags in your prompt plays a crucial role. We don’t recommend placing artist tags at the end of the prompt as this will cause them to malfunction. Similarly, using artist tags for a character without including their series tags is not advised. As shown in the examples below, using artist tags without character’s corresponding series tags significantly weakens their effect.

1girl, shiroko \(blue archive\), blue archive, ciloranko, long hair, grey hair, blue eyes, animal ears, halo, scarf, white shirt, name tag, suit jacket, necktie, v, head tilt, field, wind, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

1girl, shiroko \(blue archive\), blue archive, long hair, grey hair, blue eyes, animal ears, halo, scarf, white shirt, name tag, suit jacket, necktie, v, head tilt, field, wind, ciloranko, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

1girl, shiroko \(blue archive\), ciloranko, long hair, grey hair, blue eyes, animal ears, halo, scarf, white shirt, name tag, suit jacket, necktie, v, head tilt, field, wind, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

1girl, shiroko \(blue archive\), long hair, grey hair, blue eyes, animal ears, halo, scarf, white shirt, name tag, suit jacket, necktie, v, head tilt, field, wind, ciloranko, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

Don’t fret on putting slightly higher or lower steps!#

While we recommend 5 CFG Scale & 28 steps for generation, some images may need slightly higher or lower steps to achieve optimal results. Feel free to adjust the generation steps up or down as needed!

1girl, mita \(miside\), cool mita \(miside\), miside, animal ear headwear, blue gloves, blue hat, blue skirt, cabbie hat, choker, collarbone, gloves, hat, looking at viewer, low ponytail, ponytail, purple eyes, purple hair, red choker, red sweater, skirt, smile, sweater, teardrop facial mark, v, sensitive, field, sensitive, masterpiece, high score, great score, absurdres

Generated with Animagine XL 4.0-Opt

Anim4gine v. Animagine-XL v3.1 Prompt Comparison#

Now that we know the dos and don’ts for the Anim4gine model, how about we take a closer look? We’ve mentioned in the Anim4gine release blog post that the prompt is similar to its predecessors, so the user will be easily familiarized and adapted to it. Well, let’s see if that’s how it works🧐. These examples below are using mostly same prompts; the differences are their respective quality tags (marked by bold texts on caption), default negative prompts, and recommended settings (CFG 5 & 28 Steps for Anim4gine; CFG 7 & 28 Steps for v3.1).

Positive Prompt: 1girl, souryuu asuka langley, neon genesis evangelion, sensitive, solo, eyepatch, red plugsuit, sitting on throne, crossed legs, head tilt, holding weapon, lance of longinus \(evangelion\), cowboy shot, depth of field, faux traditional media, painterly, impressionism, masterpiece, high score, great score, absurdres

Negative Prompt: lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurry

Generated with Animagine XL 4.0-Opt

Positive Prompt: 1girl, souryuu asuka langley, neon genesis evangelion, eyepatch, red plugsuit, sitting, on throne, crossed legs, head tilt, holding weapon, lance of longinus \(evangelion\), cowboy shot, depth of field, faux traditional media, painterly, impressionism, photo background, masterpiece, best quality, very aesthetic, absurdres

Negative Prompt: lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]

Generated using Animagine-XL v3.1

Positive Prompt: 1girl, hatsune miku, vocaloid, sensitive, blue eyes, blue hair, bowl, can, chopsticks, collared shirt, detached sleeves, eating, elbow rest, fish \(food\), food, holding, holding chopsticks, katsudon \(food\), long hair, long sleeves, looking at viewer, meal, nail polish, necktie, noodles, onigiri, solo, plate, ramen, sashimi, shirt, shrimp, shrimp tempura, sleeveless shirt, spring onion, tempura, twintails, masterpiece, high score, great score, absurdres

Negative Prompt: lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurry

Generated with Animagine XL 4.0-Opt

Positive Prompt: 1girl, hatsune miku, vocaloid, sensitive, blue eyes, blue hair, bowl, can, chopsticks, collared shirt, detached sleeves, eating, elbow rest, fish \(food\), food, holding, holding chopsticks, katsudon \(food\), long hair, long sleeves, looking at viewer, meal, nail polish, necktie, noodles, onigiri, solo, plate, ramen, sashimi, shirt, shrimp, shrimp tempura, sleeveless shirt, spring onion, tempura, twintails, masterpiece, best quality, very aesthetic, absurdres

Negative Prompt: lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]

Generated using Animagine-XL v3.1

As you can see, Animagine XL 3.1 appears noticeably rigid and leans towards a 2.5D aesthetic, which prompted our evaluation of the quality and aesthetic tags. The model struggles with achieving organic fluidity, resulting in compositions that feel somewhat stiff and less dynamic.

In contrast, Animagine XL 4.0 demonstrates more artistic approach, delivering compositions that feel more natural, visually rich, and conceptually diverse. The model exhibits greater adaptability to different art styles while staying more faithful to the prompt instructions. The improvements in lighting, texture detailing, and depth perception contribute to more immersive and polished outputs, making it the superior choice for generating high-quality AI imagery.

Overall, this comparison shows the significant amount of improvements from v3.1 to Anim4gine. Despite all that, we’ve tried to ensure that the user won’t have a hard time on following this model’s guideline; by making the prompting system similar to previous versions and keeping the changes as minimum as possible.

Fundraising Has New Methods!#

We’re excited to introduce new fundraising methods through GitHub Sponsors to support training, research, and model development. Your support helps us push the boundaries of what’s possible with AI.

You can help us with:

• Donate: Contribute via ETH, USDT, or USDC to the address below, or sponsor us on GitHub.

• Share: Spread the word about our models and share your creations!

• Feedback: Let us know how we can improve.

Donation Address:

ETH/USDT/USDC(e): 0xd8A1dA94BA7E6feCe8CfEacc1327f498fCcBFC0C

Github Sponsor: https://github.com/sponsors/cagliostrolab/

Join Our Discord Server#

Feel free to join our discord server

Afterwords#

Hope this guideline/”bug fix” and model update makes up for the inconvenience on this model’s initial release. We apologize for our oversight on making the former guideline so that our model didn’t bring the optimal result. We are truly grateful for all the feedbacks and supports given by the community. And also, please look forward for our next projects in line, Animaestro!

See you later, Alchemists~