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 from 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
The photo as described by Chat GPT 4 in 150 words
This image shows an old, weathered ice cream truck. The vehicle is a Bedford, indicated by the faded emblem on the front. It is painted with a faded pink and black color scheme, and shows signs of significant wear and rust, suggesting long-term exposure to the elements. The windshield is dirty and covered with various etchings and words, likely made by passersby. On the side of the truck, there’s a sign that humorously reads “CUSTOMER SERVICE DESK CLOSED 24 H,” which indicates that the truck is not in service. The background features a partial view of a Ferris wheel, implying that the truck is parked in an amusement park or a similar festive location. However, the state of the truck and the tarp covering part of it suggest that both the vehicle and the location have seen better days and are possibly abandoned or out of use. The contrast between the ice cream truck, a symbol of joy and youth, and its current derelict state evokes a sense of nostalgia and loss.
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:
Here’s a comparison of the two images using a numbered list:
- Condition: The original image shows an ice cream truck that appears genuinely weathered and abandoned, with natural rust and wear, while the generated image depicts a truck with an artistic interpretation of wear and rust that looks less authentic.
- Color Scheme: The original truck has a muted, faded pink and black color scheme, with the pink being very pale and worn. The generated image shows a darker, more saturated pink and black, giving a less faded appearance.
- Details and Clarity: The etchings and words on the windshield of the original are randomly placed and seem haphazard, which is consistent with passersby markings. The generated image has etchings that appear more deliberate and less random.
- Signage: The sign on the original truck stating “CUSTOMER SERVICE DESK CLOSED 24 H” is clear and prominently displayed. In the generated image, the sign is present with the same text but stylized differently and blends more with the truck’s body.
- Background: The original image has a clear view of a Ferris wheel in the background, suggesting an amusement park setting. The generated image also includes a Ferris wheel, but it’s more obscured and less central to the composition.
- Realism: The original photo has a realistic quality with natural lighting and shadows. The generated image, while detailed, has a stylized look that might be perceived as a digital or artistic rendering rather than a photograph.
- State of Disrepair: The tarp seen on the original truck indicates a more pronounced state of abandonment, while the generated image, though it shows a tarp, it does not convey the same level of disuse.
Each image has its unique qualities, with the original providing a snapshot of real-life decay and the generated image offering an artistic rendition of a similar scene.
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.
