AI Image Generators - Dall E Problems and Improvements
I've decided to write a further update on the Dall E image creator. I have attained my favourite image so far ( see top), but at great difficulty. The sheer amount of images that I have had to trawl through to get this image…well let's just say it was a lot. The problem lies in a number of areas. Firstly the software does not seem to produce great composition on a regular basis. Composition, of course, is integral to an image capturing the eye. Secondly, the formation of hands, legs, and eyes seems to be a problem. Eye's in particular can be very frustrating, as the rest of the image can be perfect. Also, just the general formation of the body with relation to bodily positions specified can be a problem. Images with malformations can be very distracting. You can attempt to fix this by using the edit function, but I have found this to be currently poor with relation to realistic faces with eyes that need to be fixed. Thirdly, the randomness factor. Each variable you supply to the algorithm seems to have a randomness factor associated with it, so that there is often deviation from what you have specified. When you have lots of elements or variables, then attaining what you intended can prove to be very difficult indeed. Perhaps the algorithm creators could allow the user to specify the randomness factor for each image. Of course,it could also be a text processing issue. For instance, I suspect that certain words or descriptions associate with specific styles, so that even if the style isn't specified,that's what you get. Also, certain words or descriptions seem to be able to clash, in the sense that the algorithm has difficulty fulfilling them all for one image. Also, sometimes your prompt will produce a style that you really like, but then you are unable to reproduce this style in further iterations. This can be very frustrating, and I would suggest that perhaps having some metadata associated with each image generation could be useful for image creators. Of course, this could prove troublesome for the AI image generator companies if they are using specific people's styles in order to generate certain images. However, we have been told that this is not how the algorithm works unless specified. Personally speaking, I no longer specify a specific person's style, as I don't want to affect an image's usability in the future. All in all, I'm enjoying using image processors again. The "sickly" feeling I was experiencing appears to have disappeared, although I still worry about the speed at which this may replace normal artists. I say "may", as I'm not sure whether AI is capable of creating something new yet, a new style for instance, or even how this would relate to the prompt. There would need to be a much finer grain on the prompt, perhaps pages and pages describing each minute detail of the image, and probably a series of iterations over the same image, as the artist altered slightly from image to image. So that, I think, is still someway off. This leads to another problem. Although you can create a lot with the current prompt description, the character limit is now being reached. For the above image, the character limit was more or less met, and I could really do with more characters to be even more specific. You can find this print and more available at the below link: AlmosLataanArt at Etsy