One of the potential applications of AI in the web3 ecosystem is the creation of generative art on decentralized platforms. This would allow unique and original artworks to be created on the blockchain, with ownership and provenance tracked transparently through smart contracts. However, as we have discussed, AI still struggles with certain art aspects, particularly accurately depicting hands. This presents a challenge for integrating AI-generated art into the web3 ecosystem, as the accuracy and realism of these artworks are essential to their value and authenticity on the blockchain.
Why is AI Struggling with Hands? Why are they like that? 👋🏻
We asked it to create a post-apocalyptic world, and it did. A giraffe that lived in medieval Europe? It created that too. A man holding a bright red apple? It also made that, but the hands… why are they like that? We are facing a strange situation where AI art can dress the Pope in Balenciaga with a single command but struggles to draw a woman holding with her hands an umbrella. This is not just a random mistake. AI's struggle with hands can teach us something about how generative AI works.
First Steps : 🐾
Let's imagine a human who wants to be a painter and has artificial intelligence. The human will first enroll in a professional school and examine hundreds of objects and paintings. Shapes, movements, textures, light, reflections… This first stage is inevitable for artificial intelligence as well. The internet, the blessing of the last century, offers endless opportunities to learn objects and textures, and artificial intelligence wipes out this source in a concise time. An apple… a red apple on a brown table. Or other red apples on a white cloth…
Difference in Learning : 🍪
The biggest difference comes after this point: when you want to understand an apple, you can turn it in your hand, but if AI wants to understand an apple, it will only find another picture of an apple on the internet. This is, of course, a significant disadvantage.
Drawing Techniques : 🦋
Artists who want to draw a complex object simplify it to basic forms. That's why, when we examine a hand sketch, we first see an unpleasant, large, and rough shape of the palm and fingers. However, the artist adds their desired style, texture, and details to the drawing later. AI works very differently. As I mentioned earlier, artificial intelligence only received training on how objects look but did not learn anything about how they work. Therefore, if we look at the hands, it draws - although the shapes are generally strange - it does a great job of capturing the light and texture without knowing how they move.
As a result of all we have said, we can explain why hands are confusing for AI for two main reasons: (And I think it didn't take enough advantage of Leonardo Da Vinci's expertise on hands)
Data Size and Quality The number of high-quality and clear visuals that artificial intelligence can find for our hands is much less than that for other body parts, especially the face, and head.
Learning only from pictures The reason current artificial intelligence depicts our hands strangely is because our hands move strangely. Even if we had a billion hand visuals in our pocket, the movement of our hand would probably be different in each one. Instead, we must explain how our hands move and function to AI.
Although AI has taken significant steps in producing realistic and impressive artworks, it still struggles to depict human hands accurately. However, as technology continues to develop and more advanced models emerge, these limitations will be overcome.
"Creative artificial intelligence models are still in their infancy, but their progress is breathtaking. I believe that one day these models will reach a level where they cannot be distinguished from art, music, and movies created by humans." Andrew Ng
Especially with Midjourney-AI's Version5, which they shared last month, it seems they are gradually overcoming this problem. It is an exciting time for AI art, and we are eagerly awaiting what the future holds for this rapidly evolving technology.
This presents a challenge for integrating AI-generated art into the web3 ecosystem, as the accuracy and realism of these artworks are essential to their value and authenticity on the blockchain.