This is the latest entry in our ‘State of the art-ful AI’ series, where we upload a photo to Chat GPT 4 and ask it for a photo description, then ask DALL-E 3 to generate an image using the description. How similar will they be?
Why this workflow?
Less than a year after its launch, ChatGPT and DALL-E have already seen multiple iterations. In October 2023, ChatGPT 4 gained the capability to describe uploaded photos. To track the progress of generative AI, we uploaded these AI descriptions to the Dall-E 3 image generation model to see how faithful its output is to the original photo. A variety of photos have been selected which we’ll bring you throughout November and December.
Our original photo
This image is of the base of the circular ‘Wall of Death’ construction that tours the UK.
The photo as described by Chat GPT 4 in 150 words
This image captures a vibrant and dynamic scene inside what appears to be a motordrome, also known as the “Wall of Death.” The circular wooden arena has a striking red and yellow striped pattern, reminiscent of a circus tent or carnival design, enhancing the spectacle’s visual allure. The floor of the arena is littered with various objects including tires, a red box possibly for equipment storage, and an intriguing geometric sculpture.
Several vintage motorcycles and a single go-kart are strategically placed within the enclosure. The motorcycles, with their gleaming chrome and leather seats, are classic in design, suggesting a retro or perhaps a custom aesthetic. The go-kart, with its exposed metal frame and black tires, stands out due to its more utilitarian design compared to the motorcycles’ sleek lines.
The image is shot from an elevated angle, giving a full view of the arena and creating a sense of anticipation, as if performers are about to ride the walls at any moment. The worn patches on the wood and scattered debris hint at frequent use and a raw, adrenaline-fueled environment.
Photos generated by Dall-E 3 using the AI description
Spot the difference
Using a separate chat, we asked Chat GPT what differences it finds between the original photo and the FIRST image generated above:
Comparing the two images provided, here are the main differences in a numbered list:
- The perspective of the photos is different. The first image is taken from a higher angle, looking almost directly down into the well of death, whereas the second image seems to be taken from a lower angle, more towards the level of the riders.
- The color scheme in the first image shows a consistent pattern of red and yellow stripes on the floor, while the second image has a more varied color pattern with red, yellow, and natural wood tones.
- The first image shows motorcycles and a go-kart with no people present, while the second image includes miniature figures of people, motorcycles, and cars in motion, indicating activity.
- In the first image, there’s a red box in the center with various objects on top of it, such as a helmet and a spare tire. In the second image, the central red box appears to be closed with no objects on top.
- The overall setting in the first image appears to be less cluttered, with fewer objects scattered around, compared to the second image, which includes additional items like tires, balls, and vehicle parts around the scene.
- The first image has a single central pole, which appears to be supporting the structure, while the second image does not show such a pole in the center.
- The second image has added elements of artistic flair, such as exaggerated shadows and highlights, giving it a more stylized look compared to the more realistic appearance of the first image.
These differences indicate that the first image might be a real photograph of a well of death, while the second image appears to be a stylized, possibly digitally altered or illustrated representation.
Example workflow video
Here’s our workflow in action, using an ice cream truck image. This is not an explainer video – It shows the main steps of the above ChatGPT 4 workflow in November 2023.
