For some reason it really like to use blue and red together. With Raden I only wanted blue and commonly got both, and now with Bae it's the other way around.
>>67016708>>67015678Actually, correction. It's been a consistent 109.36 seconds for the past ten gens. And that's also for a batch size of two. So overall it actually is almost twice as fast as before.
Also, it seems for scenes where there's "more" going on, it gets more unstable, especially during upscaling. Maybe you can combine it with normal PromptEncodes to get more "faithful" upscales? Or maybe my process is just wrong for it. I'm still using my previous workflow, just with the prompt nodes replaced.
It's probably just that it requires an entirely different way of prompting than before, and I still need to learn it.
>>67017347It should for me as well, I have 16GB VRAM with my 6800XT. Who knows.
Also yeah, I'm not so sure about it being necessarily better than with the normal prompts. It's definitely different. More "noisy" but less "wild" I feel like so far. if that makes sense. If you have seen them, these are almost the same prompts as the scary baes I posted yesterday.
If I ever make a groundbreaking discovery I'll try to pass it along to the wikianon. They might be interested.
>>67017254Your gens are super cute! But yeah, SDXL is tough to run. Auto1111 just straight up locks my PC if I try to run any XL model on it. I straight up have to reboot through SysReq since nothing else works anymore.
However all that talk about the nodes isn't actually required. SDXL works just fine with the normal prompt nodes.
