Would DALL-E Mini be appropriate for fine-tuning for human face-based model?

[NomeQ]

Hello! I am currently researching options for a pre-trained model. I'm working on a personal project to tune a model to do text-to-image just for D&D character portraits, and curious if you've had any feedback or done any experiments on fine-tuning DALL-E Mini for a niche task.

• I noticed, both in using DALL-E Mini and in your report, that the current model doesn't do particularly well with human faces. Are you aware of anyone doing experiments to fine-tune it to be better at generating faces?
• Overall, do you have any data or observations from fine-tuning DALL-E Mini on small datasets or few-shot learning? That would be an awesome future report, if not. I've been amazed by how NLP models, like GPT-J, have shown such good tuning results on datasets with 20-100 examples. I wouldn't expect the same from image models, but I'm curious what the threshold is.

[borisdayma]

For faces, I think a few other models such as GAN's may be performing better for now