I'm changing my training and judging process this time. Improving the image tagging, adding DALL-E samples to the training data, using multiple checkpoints to generate more training data in later-stage models, etc.

For judging, GPT-4 with DALL-E will be asked to generate 10 prompts(since it has been trained to write prompts). For each prompt I will give GPT-4V this market description; and a single image canvas containing 1. a reference image of Satania, 2. an output from an existing model(probably NovelAI or the most popular SD 1.5 lora on Civitai), and 3. an output from my model. The 3 images will be labeled "reference", "A", and "B" with the order of outputs randomized.

Then, it will use this grading rubric it wrote:

If 7/10 prompts it deems my model "higher quality", then this market resolves YES. Otherwise, resolves NO.

If I don't finish my Satania model on time, resolves NO.

If GPT-4V is not available to me, I will wait up to 2 weeks and then resolve N/A.

I will buy and sell YES on this market but will not take a net-NO position(or will immediately sell if I accidentally do).