First of all, the AI image generation tool I used is DALL-E.
In the beginning, I choose to describe my appearance to the AI, including hair color, length, and even hairstyle. Then, I described my skin color, eye color, and face shape. I also mentioned that I wear glasses and described my glasses style. Secondly, I provided cultural background information. I mentioned that I am Chinese and a junior student at the University of Alberta. At the same time, I also typed that I like watching movies and listening to music. And I like to go to a cafe to do my assignments. More details also include my usual dress is casual, mainly sweatshirts and jeans. The colors of the clothes are also concentrated in black, white, and gray.
Then, I sat in front of the computer, waiting for the algorithm-driven AI to calculate the basic information I provided and countless data in the database, trying to generate a portrait of "me." In the first portrait, since I did not mention that I did not have bangs, it directly generated an image of me with bangs. And the first portrait was more cartoon-like. I immediately adjusted and added the information about my hairstyle in the portrait without bangs, and I wanted the portrait to be more realistic.
The next four portraits all had the same problem: the AI listened to my suggestions but ignored the details I mentioned at the beginning. For example, I added that I didn't have bangs, but the length of my hair became short. I wanted the image to be realistic, but the background information of the portrait was not that rich, just a gray background of the person. It was not until the sixth time that I integrated all my requirements again and generated a relatively satisfactory portrait.
So, I reflected that AI models usually generate images based on probability and pattern matching, and they may not fully understand the logical relationships in human language (such as the association between "no bangs" and "hair length"), resulting in lost or incorrect details. When generating images, AI models may tend to focus more on the subject (such as people) and ignore the richness of the background or environment. In addition, the model may lack a deep understanding of the cultural background and life scenes I mentioned, resulting in a lack of personalization in the images generated at the beginning. In short, although the current AI technology for generating images is powerful, it is still limited by the quality of training data and the complexity of the model and cannot fully meet users' high requirements for personalization and details. In addition, when processing multiple rounds of interactions, AI models may not be able to fully remember all historical information or lack coherence between different generation tasks. This reflects the limitations of current AI in long-term memory and contextual understanding.
Comments
Post a Comment