Also the potential outcome of hundreds of generations in terms of finding gems is especially interesting. Like finding one seed within something like 400 generations that is very special compared to the others. As well regarding learning how to prompt with that specific AI a greater number of generations makes understanding/exploring easier. The development of an idea is something I´m doing via refining the prompt itself, at least narrowing down the range in that the generation most likely has a favourable outcome. Afterwards I can simply regenerate the identical prompt for basically an arbitrary number of times, delivering varying satisfying images all over the place.
Naturally combining both approaches is favourable after all and something I´m looking forward to. Saying both ways of working have their advantages and disadvantages that can be maximised/minimised when being united.
Anyway, would be interested in your seeding refinement results if you feel like sharing
And yes, rendertimes were a little bit different back then
crimsonwarlock wrote: ↑14 Sep 2022Running locally indeed gives more options. As I can use the same seed to re-render the image with little tweaks, generating hundreds of images a day is no real use to me. I like the option to use this to actually develop an idea instead of just randomly generate stuff. For what I (try to) do with it, 12 minutes for a render is acceptable. I used to run DKB-trace on MS-DOS back then (yep, I'm old) and rendering a 320x200 image (the ones with all the shiny spheres) took about 72 hoursmoofi wrote: ↑14 Sep 2022Yeah, well, currently I´m creating up to 800 images a day, saying 12 minutes rendertime would be a little slow here
With an RTX 3080 you can get one image in around 20 seconds. Even then, though quite acceptable, with wombo I can create a preview grid of 4 pictures within 13-20 secs and only choose the ones I like to render big while the upscaling from a chosen one takes like 1-2 seconds. Then wombo isn´t aiming for seeding, saying it will always be a new outcome generated from the prompt. The seeding plus using varying sampler models is something I´m going for locally eventually when I can afford a new videocard.