3 ms·

It doesn't guide you but DALL-E 3 via ChatGPT essentially mimics an InstructPix2Pix workflow so you can continue to iterate on the same image requesting changes

It doesn't guide you but DALL-E 3 via ChatGPT essentially mimics an InstructPix2Pix workflow so you can continue to iterate on the same image requesting changes to the picture using simple English, i.e.: Generate a picture of a house house.jpg Let's change it to colonial style House2.jpg I don't like the color let's change it to red. House3.jpg Redo it using acrylic oil painting. Etc. Since we're focusing on images right now, you can't just "build one prompt" to rule them all. Effective prompts are highly dependent on the checkpoint, sampler, Loras, embeddings, etc. Incidentally, I believe that Foocus tries to use a smaller LLM To try to expand a prompt.

hanniabu 3y ago
The problem is not knowing all the keywords, otherwise I'd just use them all at the start. That's the point of having something that guides you.
- vunderba 3y ago
  Keywords is a little too handwavy for my taste - there's no universal definition of what this means. For example, a lot of the stable diffusion 1.5 checkpoints can generate better looking imagery if you add in essentially keywords such as "trending, masterpiece, award-winning, etc.", but they aren't nearly as important in certain SDXL checkpoints. To me it sounds like you want an interactive wizard / tutor which could be useful, but my point is you'd have to customize or fine-tune it at a very deep level per model (in the case of images generation). It would definitely be a non-trivial amount of work. That's why ChatGPT DALL-E 3 is so nice, it's end-to-end in terms of allowing simple expression of what you want to see and an iterative workflow. My younger siblings are in elementary school and are able to get great results from it. Midjourney also lets you upload an unrelated picture as a stylistic reference, so you can find a picture that has the aesthetic you want (ink, watercolor, matte, etc.)

hanniabu 3y ago

The problem is not knowing all the keywords, otherwise I'd just use them all at the start. That's the point of having something that guides you.

vunderba 3y ago
Keywords is a little too handwavy for my taste - there's no universal definition of what this means. For example, a lot of the stable diffusion 1.5 checkpoints can generate better looking imagery if you add in essentially keywords such as "trending, masterpiece, award-winning, etc.", but they aren't nearly as important in certain SDXL checkpoints. To me it sounds like you want an interactive wizard / tutor which could be useful, but my point is you'd have to customize or fine-tune it at a very deep level per model (in the case of images generation). It would definitely be a non-trivial amount of work. That's why ChatGPT DALL-E 3 is so nice, it's end-to-end in terms of allowing simple expression of what you want to see and an iterative workflow. My younger siblings are in elementary school and are able to get great results from it. Midjourney also lets you upload an unrelated picture as a stylistic reference, so you can find a picture that has the aesthetic you want (ink, watercolor, matte, etc.)