Well done, my man! I can't wait to try it out.

Now that's exciting to try out

很高兴，又见到你了，老朋友

The only question is how it fares against illustrious. Animagine 3.0 was overshadowed by Pony, wonder if it will suffice against another peer this time. Huge project. much deserved fame, let's see.

I like this version 4, but it doesn't feel/act like Animagine 3.1.
Some things are better (more user friendly, less tinkering, generally better anatomy, better instruction following, digital painterly styles), somethings less preferable (less expressive artist tags, flat color, two tone, simplified light physics, ai-generated doesn't work in negative prompt, etc.).

I'd say that After some testing and playing with artist tags and characters, this version absolutely can be compared to illustrious like it feels as one of illust models but it's SDXL Definitely a huge improvement from V3.1

This is probably the best Anime base model right now. One of the downsides of Illustrious and Noob is that they are based on a fine tune of SDXL that cooked the text encoder on danbooru tags. This model retains the knowledge from the SDXL base model so it is more flexible and creative.

For example, try this silly prompt and you will see what I mean: "anime screencap, year 2023, best quality, key anime shot, dynamic angles, toho animation, 1girl, hat inspired by lululemon, purple eyes, elf, holding a sword, battle pose, plugsuit designed by Gustave Eiffel, Grand Canyon mystical forest with purple leaves, depth of field, 300 movie inspired"

All the Illustrious and Noob finetunes will fail at this (they are still great models though). You can even use celebrities or internet culture references to make some cool stuffs since this knowledge isn't lost.

Great job and thanks for this release!

For my testing, the main issue is Anatomy, but it was already mentioned in the Limitations. Ones that is improved or Fixed, this could be really really good.

Also, I noticed BREAK has a lot of power, it could literally separate character and style completely. I will do more testing, so far, I'm enjoying it.
Will there be a 4.1 eventually?

Make it good or bad, this model is not very supportive of furry content.

In order to using SDXL LoRA, probabaly this is the best model in SDXL 1.0

Prompt Adherence: Ignore the haters - people are still telling me my Pony AND Illustrious ones blow at that.
Style: It works with loras.

This is really great! It's very good with prompt understanding and composition, it even knows the newest genshin character (Lan yan), which I was surprised as they just dropped not long ago at all. It even works with some lora trained on illustrious and noob. Only issue I have now is it seems to need highres fix to polish out any fine details/outlines otherwise it's a toss up whether or not you'll end up with defining lineart that isn't like muddled/squishy. kind of like what happens with an underbaked lora. Since it's random when this happens my guess is that some style and/or artists are undertrained? Without a prominent style and artist added into the prompt it can give you undertrained ones (or maybe that's just the artist's style?) which probably causes those issues. Since, as I put 'garouma' instant clarity. Though regardless thank you so much for this amazing model!

This model very powerful when you combine between character name and series. I can give you an example such as "Nakano Miku" from "Go-Toubun no Hanayome", if you prompt only character name model will doesn't know about this character (but seem it know well about rest of four from Quintuplets, poor Miku lol)

Any reason to use this instead of NoobAI-XL Epsilon (Illustrious 7.5m arts + NoobAI Epsilon 12m arts) or NoobAI-XL V-pred (Illustrious 7.5m arts + NoobAI Epsilon 12m arts + NoobAI-XL V-pred 16m arts)?

Gotta be honest: this has some issues. Artist styles are inconsistent, fingers are abysmal, and it seems to always default to super low saturation.

Quality: Animagine = Illustrious (up to model, lora) > Pony

Characters, Series: Animagine > Illustrious > Pony

Danbooru Tags: Illustrious > Animagine = Pony

Artists, Styles: Animagine > Illustrious > Pony

Fingers, Anatomy: Illustrious > Animagine = Pony

LoRa: Pony > Animagine > Illustrious

NSFW: Pony > Illustrious > Animagine

Overall: Illustrious > Animagine > Pony

I really like this model; I see strong competition with Illustrous. I’m hoping for more versions that try to mitigate what is obviously not easy to fix. Does anyone know of a list of the characters this model has been trained on?

SO GOOD

I've done a few tests and I have to say it's pretty disappointing in terms of quality.

I like this model, but I get inconsistent results. I hope this issue is resolved soon

No separate section? Whats goin on?:))

Fairly low quality outputs. Sloppy and uncoordinated structures in the images, as if someone threw paint all over the place.

Sasuga animagine, while it lacks the pose that IllustXL provide, the quality and clarity is crazy. If you guys having finger issue, use "five finger" tag. it'll reduce some of the problem.

10/10

falls apart when doing more than "1girl"'ing but finetunes should go hard. not as good as base illust but is a solid base for models to build on.

训练信息的图片是不是放错了？

放了推荐分辨率得图片了？

Trained a LoRA for Ani 4.0. Can balance brightness and enhance contrast. Also some hand improvments. https://civitai.com/models/971952?modelVersionId=1340810

This model is very good at generating anime art IF you know the correct prompt and settings for it.

Many people have already mentioned about the broken results. In my case, I get those results mostly because I use the prompt style and settings from another model.

I tried the guidelines in the description, and the results were better than what I usually had. So, the problem could be either the model itself or just a skill issue.

And yes, it still has some weaknesses. But I think this model has improved a lot since the previous version.

Unfortunately, despite the addition of a scoring mechanism similar to Pony, there are still issues that have persisted since version 3.1. The accuracy of character movements remains poor, with problematic hand and limb rendering, and the model seems even less sensitive to NSFW content. The prompt weighting has also become quite inconsistent compared to version 3.1.

On the positive side, the image quality is still excellent, and the lighting effects and background details have improved over version 3.1. However, the anatomy of the characters continues to be problematic. The body proportions often look unnatural, with characters appearing too thin and small, resembling a loli-like body shape. Additionally, without using artist tags or LoRAs, the characters' faces still default to the classic "AnimagineXL anime face."

Overall, while there are some enhancements, I don’t see this as a significant improvement over version 3.1. That said, I appreciate the effort and work put into this update.

https://civitai.com/articles/11092

Prompting Guide. Covers about 300 tags and 60 characters, many of which are not in the github list.

An excellent model; I particularly appreciate its color scheme and attention to detail. Its composition and background are truly commendable. However, the issues with the fingers are quite critical, showing poor anatomical accuracy. If this problem could be resolved, it would outshine models in the 'il' series

Really like this model.

It takes a bit more generating/inpainting compared to noobai but the aesthetics are way better.

What are diffusers and is it a must to have?

I think this version has lost some of its former strengths. The style has been compromised by fads and it doesn't look like the anime I was familiar with. The physique is too similar to a particular artist. However, I cannot recall their names.

This model looks super interesting!

a40, I call it that, It's definitely a great model.

animage31 used to be my favorite anime model.

Sorry for the possible offense, I made some suggestions and comments.

Characters, backgrounds, style, body structure.

a40 and noob e11 characters may be comparable, a40 characters are newer, but a little character features seem not fixed enough, probably because a40 starts from sdxl, while noob comes from illu, with more inheritance.

I think a40 may look better in the background，But the stability of the limbs may be inferior，and probably because of multiple views, it is easy to have one more person in the horizontal image..

But in terms of style, the artist's words are almost not obvious, and the default style and physique seem too distinct.

when can we expect to see the dit based animagine?

Can this recognize artists from Twitter?

100/10

Cool

How do I get the diffusers?

where to get character tag file?

100/100

lots of anime characters arent showing up. Is it just me?