1) Not necesserily. After i found interesting image, i just turn up upscaler or ADetailer, or both in txt2img, then just regenerate image once more (i don't like switching tabs).
Upscaler allows you to enchance image, but not always work.
You can combine this tags or place them in negative prompt.
3) Illustrious and Pony both are based on SDXL1.0. Illustrious checkpoint should work with your UI, but if not i recommend reForgeUI. Later versions of Forge are not very stable, at least in my experience.
Yes, img2img, is a good start to modify a generated picture. But the inpaint tab, just next to it, give you more control to what you want to edit.
For perspective, use danbooru tag like facing viewer, cowboy shot, from below, from above, from side, wide shot, distant, and so on.
Here's a list of tags to use
You must be registered to see the links
You can also download this extension, which autocomplete tags
You must be registered to see the links
You should also be familiar with “weight”. Weight reinforce a prompt and make it stronger. For example, if you want to make a perspective stronger, or anything really, select a prompt, and then ctrl + arrow up or down to decide if you want to enhance or weaken a tag. Never go beyond a certain threshold, or it's going to fuck with the render. Just experiment with increment.
Pony and Illustrious are both based on SDXL, so you can use an illustrious model like WAI or HassakuXL with forge no problem.
And just in case you didn't know. Use specific resolution from which all those SDXL based model are trained.
My attemps to use Sora from OpenAI:
Firts image looks like fail. Second is much better, environment looks really natural. Sword is weird, but this my fault, i typed greatsword instead of just sword. You can Sora by yourself at sora.com (3 images per day for free users).
It's really interesting how much VRAM needed to run this generative AI, to better realize gap beetwen Sora and Flux (or similar models).
Sora supports other languages too! Most amazing that Sora can generate more then one character at image, 7-8 characters is possible. Architecture is also a lot better, then with SD AIs. Plus supports video generation, which is aslo great.
Don't know how to describe her folded ponytail to Sora)
Looks decent, but failed to cast shadow from dragon on village.
All I can offer now is a fantasy illustration with Sabia. For now I'm testing the models I downloaded on similar (fantasy) prompts.
I'll generate something sexier with Sabia later.
