Summary
LinkedIn posts with pure-AI images do worse than none; posts with striking text do better than none; I don’t want to hand-make them in Photoshop ... can we do better with DALL-E?
Usefulness: Practical/Productive
Update
Summer 2024: OpenAI released GPT4-o, which is specifically able to do excellent “text as images” — this experiment needs re-doing/updating.
Techniques
- Short prompt to generate variations, pre-sized to Linkedin’s required format.
- I’ll ask for different stylistic variations to see how good/bad the options are – but expectation is that I’ll only use the ‘pure text’ ones
- Prompt:
- “Generate an image sized 1200×644 pixels (suitable for posting to LinkedIn) that contains two pieces of text, a title and a body.
- The title is: “[my 1-phrase summary of the category of post]”
- The body is “[my 1-line summary of my post]”
- The title and text should use contrasting colors, in a professional looking font. Generate three variations:
- 1. Colored text and colored background with no photography and no illustration
- 2. Stylized text and colored background with no photography and subtle illustration, the illustration should be low contrast and/or high transparency so that it doesn’t distract from the text
- 3. Text with relevant illustration or photography
Discoveries
- ChatGPT-4 in spring 2024 refuses: “For the requirements you’ve described, my capability is to create one image at a time. However, I can generate one of the variations you requested. Which one of the three variations would you like me to create first?”
- FAILED: even the requests for plain text generate highly illustrated images
- MODIFICATION: I took the best outputs, loaded them into Affinity (the low-cost, non-Adobe, alternative to Photoshop) and cleaned + edited them to fit my goals.
- NOTE: creatively DALL-E 3 is still quite weak compared to doing photographic/stylistic work with MidJourney – and it’s vastly worse than working with a human designer/illustrator. I could literally have done better in only a little more than that time it took me to write the prompt and wait for the responses, and edit in Affinity/PS.
- NOTE: DALL-E completely ignores 1-line, 1-sentence, basic instructions, while stating explicitly that it’s following them.