Yes; the only variation in generation times were from the sampler used, steps chosen and resolution as neither are turbo models or use the LCM sampler. Wasn't really testing for generation time or paying much attention to it but if there were an outlier I likely would have noticed.
These ain't too bad, sucks that it's practically a paid service. I wonder how many more years it's gonna take to get fingers right though lol.
Impressively good at mixing concepts, works great without the need for any easynegative embeds and has a fairly wide mix of styles outside of anime despite what the images suggest. It even creates readable text sometimes, but not as well as Flux. I see no need to return to Pony, it's a direct upgrade.
Very interesting. I've been trying to branch out with LoRA training, so I've had my eyes set on working with Pony as that appears to be what most people are still using. Gonna consider doing it for Illustrious...
I’m really enjoying the Illustrious models. One thing they do really well out of the box is getting your subject to hold an item. With minimal generations, I was able to get some bad ass armor-clad knights to swing around huge war hammers.
Are you using the right settings? It only works right if you're using Euler as the sampler and the number of steps and CFG have to be in a pretty narrow range too.
Correct, and have tried the various schedulers as well. Just not a pleasing result to me. Do you like the end results you’re receiving?
After messing around a bit I see it too. It seems like it's trained so heavily and drawings it can't handle trying to do photorealism the way other models can, which is a shame but it still makes some pretty good results so long as you play to its strengths and don't look at it too closely. If it's still looking weird regardless of what the prompt is using the default SDXL VAE instead of the built-in one might help.
Have you looked into Lobe? I ran it alongside Forge for a while. It provided a decent alternative, but it seems to lack frequent updates. It compartmentalized a lot of crap into sidebars that you could swipe left or right from edges to bring forward, like checkpoints and LoRAs, etc., and it’s highly customizable. Downside, for me, was that it made my phone run extremely hot when using and it would frequently refresh (latest iPhone and iOS version). It would also drain battery quickly because of it which was enough for me to revert back to the ugly and cumbersome A1111/Forge mobile UI.
How the hell are you running Stable Diffusion on a phone?
You run it on a machine with a GPU. Have it listen on 0.0.0.0, and then connect to it on your network or over the Internet using a domain. Cloudflare tunnels make it super easy to do that.
Thanks both. I'm honestly a little disappointed that things haven't taken off more. I guess I'm mainly still just going back to SDXL, then. I'd kind of hoped we'd had another leap forward as Flux seemed pretty cool when it was released. Well, is pretty cool - but still lacking the sort of community support SDXL has. And SD3 doesn't appear to be taking off very well from what I can see.
