Reflection
Was the result relatively accurate in your estimation?
Arriving at this task, I have never used Craiyon before. I found the site easy to use but very slow on my computer which rarely ever happens. The generated images, although original. Seemed to lack a lot of detail and the prompting never really seemed to help the quality improve. The AI struggled with the details of the image and wasn’t able to match my instruction, such as accurately creating a dog wearing a red cape or creating a realistic human face. This does seem to be a common limitation of AI image generators that I have used. The AI always seems to produce variations that are close but not exactly what I envisioned.
Were those images what you had in mind when you gave the AI the prompts?
The images definitely did not meet my expectations. The AI’s interpretation of “a dog wearing a red cape” resulted in various red clothing items instead of a consistent cape. Similarly, the human face distortion indicates that the AI might not be adept at generating realistic human features, especially in dynamic poses like jumping when I entered the prompt to “create a man who has just graduated from his masters program at the university of british columbia jumping up in the air on campus, excited to be all done.”
What differed?
The main differences were:
- Inconsistent Details: The AI generated different red clothing items instead of a specific red cape. I assumed that a cape would be a specific enough type of garment that the AI would be able to produce.
- Distorted Human Features: The AI struggled with creating a realistic human face, particularly in an action pose. The images were mostly headshots of men smiling or graduation related images that didn’t fit my request.
What can you infer about the model or the training data based on the results?
These inconsistencies may suggest that the AI model might not have been trained on enough diverse and high-quality images of the specific scenarios that I requested. Perhaps this AI tool, Craiyon, relies on patterns and previously requested images to learn and improve. If Craiyon hasn’t seen requests like mine such as dogs in capes or people jumping in specific contexts, perhaps Craiyon struggled to produce original accurate pictures. The article from Vox highlights similar limitations of AI-generated images. It mentions that while AI models like DALL-E can create impressive visuals, they often fall short in generating detailed and accurate representations, especially for complex or specific prompts (Heilweil, 2022). This aligns with my experience, where the AI’s output did not fully match my expectations. While I would conclude that AI image generators have made positive strides, they definitely still have limitations, especially with detailed and specific prompts. While the technology continually improves, it may require multiple banks of information and refinements to get closer to a desired outcome.
References
Heilweil, R. (2022, December 7). AI is finally good at stuff. Now
what? Vox. https://www.vox.com/recode/2022/12/7/23498694/ai
artificial intelligence-chat-gpt-openaiLinks to an external site.
One reply on “Task 11: Text-to-Image”
Hi Carlo,
I enjoyed reading your post because I was curious what other prompts people used and the kinds of results they had on Craiyon. One difference I immediately noticed was that your prompts were very specific and detailed, whereas I had very general prompts like “boy with sports car” or “family cooking together”. I thought my prompts were not specific enough and that was why the images generated weren’t exactly what I was looking for and did not do a good job representing different races and genders. However, after seeing your results I now realize that AI is just not very good at generating images still due to a lack of input data. It would be interesting to look into how old Craiyon is, and how it would evolve in a few years’ time to be more accurate in its representations. Conversely, are there other AI image generators that are a lot more accurate and is there a reason for that?