A slight tonal shift in a prompt word produces a totally different looking DALL-E 2 image.

Judge as of 2022-07-04. True: demonstrations, tests or published analysis of DALL-E 2 show that changing a single word for a near-synonym with different connotation (for example a 'Russell conjugation' pair such as 'firm' vs 'obstinate') generally produces visibly and substantially different images. Mostly true: such word swaps produce noticeably different images, but 'totally different' is somewhat overstated (for example, the composition is similar while style or mood changes). Mixed: the effect appears for some word pairs but not others, or only inconsistently. Mostly false: tonal word changes produce only minor differences, so the claim overstates a small effect. False: evidence shows tonal synonym swaps produce essentially indistinguishable outputs.

Set before the research began.

1. E1 UPDATED! DALL-E 2: Some Results (and Some Thoughts) After Using OpenAI’s Revolutionary, Amazing Artificial Intelligence Tool Every Day for Two Weeks to Create Images Ryan Schultz (blog) · 4 July 2022

   “(interestingly, using the word “woodprint” gave me monochrome results, while just “print” threw up a few coloured prints!)”

2. E2 UPDATED! DALL-E 2: Some Results (and Some Thoughts) After Using OpenAI’s Revolutionary, Amazing Artificial Intelligence Tool Every Day for Two Weeks to Create Images Ryan Schultz (blog) · 4 July 2022

   “(Adding the words “digital art” and colorful” really makes a difference in the results!)”

3. E3 Design Guidelines for Prompt Engineering Text-to-Image Generative Models Vivian Liu and Lydia B. Chilton, Columbia University (CHI ’22) · April 2022

   “We synthesized the following guideline from this experiment: When picking the prompt, focus on subject and style keywords instead of connecting words.”

4. E4 DALL·E 2 Prompt Engineering Guide @rendo1 & @luc (Discord community)

   ““Detailed” ■ Adds more precise details to the output, instead of simple art, but can also make the art overwhelming/over the top in small details.”

5. E5 Design Guidelines for Prompt Engineering Text-to-Image Generative Models Vivian Liu and Lydia B. Chilton (Columbia University), CHI '22

   “We found no significant difference between the quality of the images that these nine prompt permutations generated, and therefore no significant difference between different prompt permutations.”

6. E6 DALL·E 2 Prompt Engineering Guide created by @rendo1 & @luc on Discord (fan-made document, hosted on gwern.net)

   “Do NOT use this if you want to create a prompt which looks like a real photo, as the term “photorealistic” is used to describe an artificial image that looks real, not a real photo.”

7. E7 DALL·E 2 Prompt Engineering Guide created by @rendo1 & @luc on Discord (fan-made document, hosted on gwern.net)

   “when asked to generate a mouse chasing a cat (in a digital art style), the AI model is unable to deliver, instead showcasing generations of an expected cat chasing a mouse”

8. E8 What DALL-E 2 can and cannot do — LessWrong LessWrong (post by Swimmer963, Miranda Dixon-Luinenburg) · 1 May 2022

   “DALL-E is often sensitive to exact wording, and in particular it's fascinating how "in the style of x" often gets very different results from "screenshot from an x movie".”

9. E9 What DALL-E 2 can and cannot do — LessWrong LessWrong (post by Swimmer963, Miranda Dixon-Luinenburg) · 1 May 2022

   “I have definitely found that saying "a painting by X" or a "a movie by X" gets results that a) I personally like much better, and b) are much more consistently and recognizably in the requested style!”

10. E10 What DALL-E 2 can and cannot do — LessWrong LessWrong (comment by gwern)

    “With earlier models, there didn't seem to be that much difference between 'by X' vs 'in the style of X'”

None of the evidence tests swapping a word for a near-synonym with a different connotation (a Russell conjugation pair such as 'firm' vs 'obstinate'), which is what the criteria ask about. E8 and E9 say DALL-E 2 is sensitive to exact wording, but their examples are style phrasings ('in the style of X' vs 'screenshot from an X movie'). E1 and E2 describe medium and style words such as 'woodprint' and 'digital art' changing results. E5 found no significant difference in image quality across prompt permutations, which does not address whether the images looked different, and E3 says to focus prompts on subject and style keywords rather than connecting words. Taken together, the evidence suggests DALL-E 2 reacts to wording in general but does not settle whether tonal synonym swaps produce totally different images.