When I received an early invite to OpenAI’s new Dalle model almost 10 months ago, I knew I had stumbled upon something very special. What can I say, it’s been a fascinating journey so far. I’ve spent much of my free time providing feedback and contributing to its development, and I’m impressed and proud of how far it’s come. Dalle2, in particular, was a significant improvement over its predecessor, of course, there have been some issues, but overall d2 did its job pretty well.
Recently, ChatGPT has been getting a lot of attention, and while it’s exciting to see new models emerge, I wished OpenAI didn’t forget about D2. Now, imagine my excitement when I was again invited as an alpha tester.
Since the model is experimental and still in development, I’m not going to write a heavy review, for now, I will stick with listing notable improvements; how prompting works with the new model, image comparisons, and re-generating old prompts.
Improvements
1. Textures(painting mediums), skin, shapes, faces, hair
2. Less cropped images
3. Improved performance when generating hybrids (Centaur, Minotaur, Elves, Cerberus, chimera)
4. Sharp photos
5. Humanoid characters
6. Incorporating more details from the prompt
Remaining Issues
1. Color allocation, “yo-yo” issue
2. gender bias when creating facial expressions
3. Safety filters false positives
Prompting and the importance of word order
When generating prompts, word order is important to keep in mind. While image-generating models can understand natural language, how you phrase your request still matters. Word order becomes especially significant for more complex prompts, which was an issue with the previous model. The more complex your prompt is, the more confusing it gets for the model to prioritize tasks, usually leading to wrong generations.
To examine the role of word order in prompting, I use the following approach: Using a complex prompt, we compare how changing the wording affects outcomes.
Prompt: A propaganda poster depicting a cat dressed as french emperor napoleon holding a piece of cheese
Let’s break down the prompt element by element:
A propaganda poster (determining the style of generation)
A cat dressed as French Emperor Napoleon (subject)
A piece of cheese object
A propaganda poster (determining the style of generation)
A cat dressed as French Emperor Napoleon (subject)
A piece of cheese object
The prompt specifies generating a propaganda poster featuring a cat dressed as Napoleon.