Have AI neural networks learned to make anime or does it still look off?

AI neural networks are already generating anime, but the quality really depends on the model and settings. There are moments that look pretty decent - the characters are recognizable, emotions show on their faces, the color palette is bright. But if you look closer, you often catch yourself noticing weird proportions in the limbs, unnatural bends in the torso, and sometimes the hands just come out wrong. Hair and fabric still often look like smeared mush instead of individual strands.

The main problem is that neural networks learn from existing frames, but anime isn't just a pretty picture - it's also a sequence of frames with proper animation. Even if a single static frame looks good, movement is where things get tricky: characters jitter, lose consistency, facial features change from frame to frame. Real anime requires a ton of manual detail work.

For quick sketches or concept art, neural networks already help pretty well. But full-fledged anime that you can watch episode after episode without getting annoyed - a generator can't make that yet without serious post-production. Most likely, in the future they'll be an assistant to animators rather than a complete replacement - you'll still have to pay for quality one way or another, either with time or money.

When I tried generating anime scenes in different AI models, I noticed one major issue - the characters' hands often look weird, with fingers not being counted correctly. This affects almost all models, even when the face and clothing turn out pretty decent. In other words, neural networks handle overall composition and facial details fine, but they still stumble when it comes to anatomy in motion.

I've seen examples that guys posted online recently - in some cases they actually turn out pretty decent. The faces, hair, overall composition can be quite passable. But the longer you look, the more glitches you spot. Apart from hands, which people already mentioned, anatomy often falls apart - body proportions are off, the legs look weird. And if there are multiple characters in a scene, they can overlap or stand in awkward positions.

I tried generating a simple scene once - a girl with a cat in a room. The cat came out kind of abstract, the head doesn't match the body, and the girl turned out weirdly shaped, like she's bent wrong. So the neural network understands that there should be characters there, it conveys emotion, but the details don't mesh. For static portraits it might be okay, but for a full anime episode with movement and character interaction, it's still way off from a finished product.

I think the thing is that anime is an art form with its own rules and style, not just a photographic reproduction of reality. Neural networks haven't really grasped this specific language yet.

The problem isn't just the hands, but also the fact that neural networks struggle with movement and dynamics. They can create a static, beautiful scene just fine, but when you need a sequence of frames, you get desynchronization - characters seem to teleport between frames or change poses unnaturally. For actual anime, that's critical because the whole point is smooth, expressive movement, not just pretty faces.