|
|
|
|
|
|
by johndough
1263 days ago
|
|
|
You can ask ChatGPT to create SVGs and at some point in the past you could even trick it into embedding them as base64 images. Not sure if it still works since ChatGPT is unreachable for me currently. More details: https://www.reddit.com/r/ChatGPT/comments/zsnscy/i_asked_cha... Adding diagrams as inputs is probably as easy as feeding in an additional CLIP embeddings during training. The trick here will be how to get enough training data. Perhaps there are enough StackOverflow questions with images in the question. For output, you could also finetune some diffusion model on that data. |
|
|