As a new member of the EDLI team, I was eager to start practicing with ChatGPT and testing its boundaries. I thought I would start by trying to get ChatGPT 4 to generate a fire-breathing camel dragon. In case you are unfamiliar, here is a depiction of the fearsome beast in question:
This sketch was first discovered in an undergraduate’s notes by Imari Tetu during a study investigating ChatGPT and technical writing. After sharing the sketch with Scott Schopieray, the two challenged themselves to recreate the image using ChatGPT – a challenge that quickly swept through the rest of the EDLI team. This fantastical creature proved to be a great icebreaker for getting acquainted with ChatGPT. To start with a good baseline, my first prompt simply asked to “make a picture based on the following description: A fire-breathing camel dragon.” ChatGPT returned the following image:
Not bad at all! I noticed a few troublesome details, such as the tail that appears to be its own fire-breathing dragon, and that I hadn’t included stegosaurus spikes in my prompt. Nonetheless, I assumed that I was only a prompt or two away from the perfect camel dragon. With my next few attempts, however, I found that this is about as close as I would come to making one student’s doodle a reality. I provided a polite prompt, asking to “remove the wings and add spikes to the neck of the animal,” and to “make the tail short and bushy.” At this point, I ran into two key limitations for generating camel dragons. First, ChatGPT frequently generates an entirely new image when revisions are requested. Second, ChatGPT does not appear to be able to identify anatomical aspects of certain images that it generates.
This isn’t exactly the revised image I requested, but I appreciated that the beast now included spikes and a tail that, while not exactly short and bushy, was at least a tail and not a hydra-like creature itself. However, I realized that this camel dragon had spikes on its back, despite my request that they be added to the neck. After a plethora of attempts, each phrased differently than the last, I simply could not relocate the spikes. It appears this could be the result of inherent limitations in the image generation process. For example, ChatGPT must rephrase image descriptions into its own prompt, which is then used to generate an image. This can result in some slight inconsistencies and inaccuracies. Additionally, when generating images, ChatGPT does not seem particularly effective with prompts asking to negate aspects of an image, or to restrict aspects of an image (e.g., “neck only”). Here are a few of the resulting images, after having prompted ChatGPT to do things like “move the spikes forward… away from the tail,” “remove spikes from the widest part of the body,” and “place spikes along that long and curvy portion of the animal.”
Interestingly, the last image even included some anatomical labels, though they appear to be approximations of real words (i.e., “hoad” instead of “head”). This piqued my curiosity, and I asked ChatGPT to draw a camel and label the neck. The new image included an arrow pointing directly to the camel’s humps.
At this point, I concluded that ChatGPT would not be able to understand anatomical prompts, but I wasn’t ready to give up on my quest just yet. I requested that the arrow be moved forward on the animal, and it did so successfully:
Referencing these two images, and with unjustifiably high confidence, I then prompted ChatGPT to “place spikes similar to those of a stegosaurus between those two arrows.” Though the image speaks for itself, let’s discuss:
This camel has perfect stegosaurus spikes on just about every other part of its body and, I don’t want to get too off track, but why did it turn into a skeleton camel? Or are we seeing through the camel’s body like an x-ray? Out of all the fantastical prompts and pictures, this was just supposed to be a regular ol’ camel, after all. Perhaps ChatGPT is attempting to create an image most likely to correspond to other images of stegosauruses, which are often depicted in a skeletal or fossilized form. My final attempt was to ask ChatGPT to “remove the spikes on the back three quarters of the animal.” Here is the image that was generated:
Between the inaccurate spike placement, the inclusion of arrows (which seem to be taunting me with their sudden, unprompted appearance and inaccuracy), and the even more skeletonized camel, I decided that this journey had come to an unexpected end and that my first camel dragon would remain my finest.
Despite my tribulations, trying to generate the fire-breathing camel dragon was a worthwhile experience and a great way to get familiar with ChatGPT. On the positive side, I found that ChatGPT was usually quite responsive to my prompts and able to come admirably close to illustrating small parts of our imagination. Furthermore, it seemed able to keep track of which image I was referencing when requesting revisions, and even made attempts to adhere to prompts which referenced multiple images (i.e., placing spikes between the arrows on two images). On the downside, however, I found that requesting even minor revisions to an image frequently led to the generation of entirely new scenes which often included new inaccuracies. This makes any image difficult to refine and suggests that those using ChatGPT to generate novel images my have to accept slight deviations from prompts. Finally, ChatGPT seemed to struggle with references to anatomical aspects of images, even those of normal camels. Prompting ChatGPT to include arrows to certain parts of the camel (e.g., the neck) highlighted this limitation, though even asking the program to place spikes on specific, highlighted areas yielded similar disappointing results. Overall, I walked away from this silly experience with an appreciation for how generative AI programs can translate creativity and imagination into a concrete image and, despite the spikes, was still quite impressed with the results. At the end of the day, it’s pretty neat that I can ask for “a fire breathing camel dragon” and get a close resemblance in seconds.