Earlier this year, AI advancements were made as DALLE was revealed to the public, making compelling images. It allows users to generate images from a text prompt. However, the latest version is not open to the public.That all changed this month.
Earlier this month, NovelAI released their image generator that uses similar methods as DALLE to generate anime-themed images. While it’s impressive how decent the images are, they caused much controversy in the anime fanart and V-Tuber community. So, you might wonder what my opinion is on this whole thing. I decided to try the image generator on NovelAI, and no, it will not replace fan art and anime art commissions for various reasons, which I will go into.
How are Images Generated?
AI image generators like DALLE, Midjourney, Stability Diffusion, NovelAI, and Waifu Diffusion work similarly. They use a method called Stable Diffusion. They are trained on images through machine learning, using a set of algorithms for the program to study the data. This is training data, which doesn’t contain the actual images. Instead, it uses metadata like tags and captions that also describe the image as part of the training data. These images come from the internet through an automated program. Powerful GPUs are typically used to train a model for CLIP and diffusion.
To generate an image, the user types a text prompt. Stable Diffusion uses CLIP to decode the text to match what it knows from the training data, then it uses that to come up with an image that matches the caption or tag. The algorithm also uses diffusion, which corrupts the image. It then reverses it until it comes up with a clear and noiseless image that looks good.
Why it’s so controversial?
It’s easy to see why AI image generators are that it is threatening to kill artists’ livelihoods. After all, people who work with digital art for a living need commissioned or professional work. AI image generators lower the barrier for users who aren’t artistic enough to make something. Making art takes a lot of effort to perfect the craft. If someone can generate an image with a text prompt, you can see why they are up in arms.
The second reason is the data that AI is trained on. NovelAI uses images from a particular image board that contains anime images. After all, AI relies on existing data to train and can’t generate ideas without training data. While you can’t copyright an art style, you can with an image, and if they aren’t credited, you can see the problem there. It brings up a lot of ethical issues surrounding the use of copyrighted images. The thing is, copyright laws have not caught up to technological advancements.
With that comes people trying to pass off AI-generated works as their own. So you can see why platforms have a problem with users monetizing it. Several Japanese services block the sale of these images. Also, several VTubers don’t like how their art tags are flooded with fanart generated with AI that actual fanart gets drowned out. Not only that, it’s easy to see why since these image generators are trained on copyrighted images.
The last issue with images generated with AI that kills any commercial viability is the US copyright law. The US recently ruled that AI-generated images can’t receive copyright despite the service granting users full rights to the image. The copyright law only gives copyrights to works created by humans, not machines. This is important if you want to sell an image or use it as your brand. Who wants to use an AI image if it can’t receive copyright to protect it. At best, artwork that AI generates is basically public domain, no matter if a user makes an NFT of it or tries to claim it.
What are my thoughts on AI-generated art?
I am neutral on the issue. I agree that using existing artwork without consent and credit is not good. At best, it’s a grey area. But, at the same time, AI-generated anime artwork is a good thing to help with brainstorming. Even a regular user can come up with a concept and then commission artwork for something better and be able to own it. After all, you can’t claim ownership of AI artwork, but you can if you pay someone to make it.
The client can give some AI images as examples of how the character should look and additional details of how he or she wants the final image to look. With that, it should reduce the rework needed resulting from a lack of clarity.
With my experiences with NovelAI, the images look impressive at a glance. After all, I am working on a project to recreate WaniKani Kanji mnemonics as anime images as closely as possible, all images are copyright free, since you can’t copyright AI-generated images.
For simple poses and stuff, this is where NovelAI excels. However, when it involves positioning stuff in a text prompt, having the person do something like eating ramen, or anything involving hands and feet, this is where things break down. You have images with extra limbs, hands looking unnatural, or strange-looking people and objects.
Sometimes, part of the prompt gets ignored. Therefore, I must keep generating images until I receive a suitable one. Sometimes, I need to Photoshop two AI-generated images to get the desired result. Ramen-eating memes aside, there are a lot of limitations. Moreover, the images lack the human touch as most AI images are stiff despite how good they look. However, this tool helps finalize my original character concept for anime art I want to commission for this blog in early 2023.
Final Thoughts
Regardless people are for or against AI art, Pandora’s Box is already open. Like with any tool, there are pros and cons. Still, I think copyright law needs to get updated to that artists and creators are appropriately credited. Also, I think there need to be new regulations that protect artists and creators. First, creators should be able to opt-in for images used for training models before others can use them as training data. In other words, models can’t use the image without the creators opting in first.
Additionally, AI-generated creations based on copyrighted works should not be allowed for commercial use, nor should they be able to receive copyright. Only the creator of the work should be able to receive copyright for AI creations if they use their own works for training. In other words, this can limit the impact it has on an artist while remaining a helpful tool for educational use or gathering ideas for commissions or own creations.
This is probably the first and last time I post AI images on my blog. I rather share anime fanart with attribution/proper credit to help promote actual anime artists than use AI image generators.
With that, how do you feel about AI anime-style art? Do you think this hurts artists? Feel free to share your thoughts in the comments.
Leave a Reply