Stage 2

The application makes two requests to GPT3.5 comparing to four requests in Stage 1 now. Instead of breaking down an object into a list of visual components, this algorithm asks GPT to describe an illustration directly. It is more of a GPT-to-GPT interface, in a contrast to Stage 1.

Despite these changes, the quality of the generated images remains stable and roughly comparable to the previous stage, with about a 20% success rate. Generated illustrations are more likely be complete objects without missing parts. On average, it takes approximately 1100 tokens to generate an image.

This is the first version that has been publish to Internet. User-generated images begin after "hacker news" images.

I am publishing 320 selected images along with their descriptions generated during this stage.

The first few images begin with "I cannot generate a visual content" preface. I modified the prompt shortly after to resolve the issue.

Key findings:

• I decided to allow GPT to generate images with empty prompts. Interestingly, GPT-3.5 tents to draw mountain landscapes and cityscapes more often. All these images have <empty> titles.
• GPT-3.5 models may occasionally be overloaded and decline requests with an error message. As of now, I have not implemented a conditional retry mechanism. However, I believe that less than 1% of all images are affected.
• GPT-3.5 may occasionally generate animated SVG images. It is hard to say whether it is a good thing or not, requires more experiments.
• GPT-4 significantly outperforms GPT-3.5 in terms of creating illustrations, but it is also 20x more expensive. I will be switching the application to GPT-4 once operating costs are acceptable.
• OpenAI's built-in moderation step has proven to be an invaluable tool in filtering out low-quality prompts.