It feels like SDXL is still holding on as the basis for most of the checkpoints/Loras/etc. that people are generating. I remember being pretty excited about SD3 when that came out, but it seemed to kind of fizzle out. I wonder if flux.1 is going to actually move up and become a well-supported base with lots of checkpoints and LoRas like SDXL did or if people will just kind of wait for something better to come along.
Flux is way more powerful than sdxl at following prompts, and it has a lot more parameters. I haven't done any training yet, but I hope we'll be able to generate sex scenes with two or three accurately represented characters, as well as a surrounding environment that makes sense.
Now, on my gtx 1660 ti that has 6gb of vram, I can generate flux images in 1024px that includes a lora. While it takes a fucking long time to generate, it still works and thus the requirement is quite low, making it accessible.
Additionally, Civitai now has flux training available. So we can test generating models to see what we can achieve.
But as I said earlier, I don't have a dataset. Well, I have the datasets Collyn made for each characters, but the images aren't captioned correctly for flux.
What I suggest is we set up a dataset in a huggingface repository where people will be able to upload pictures and complete the captioning.