Conversation

Notices

1. LS (lain@lain.com)'s status on Wednesday, 09-Aug-2023 04:38:42 JST LS
   XL is interesting, but not perfect yet
   • Token likes this.
   • Token (coin@asimon.org)'s status on Wednesday, 09-Aug-2023 04:46:34 JST Token

     in reply to
     @lain It's impressive nonetheless. Thinking how we didn't even have this possibility at all just 2 years ago
     LS likes this.
   • LS (lain@lain.com)'s status on Wednesday, 09-Aug-2023 04:48:01 JST LS

     in reply to

     • Token
     @coin yeah, and this is pretty much standard XL, if you see how far SD 1.5 has come from SD 1.5 base to current finetunes, i'm really expecting a lot SDXL in the future
     Token likes this.
   • waltercool (waltercool@pl.slash.cl)'s status on Wednesday, 09-Aug-2023 04:51:07 JST waltercool

     in reply to
     @lain Try some SDXL models. The base model always scks
     
     https://civitai.com/models/118114/sdvn6-realxl
   • LS (lain@lain.com)'s status on Wednesday, 09-Aug-2023 04:51:07 JST LS

     in reply to

     • waltercool
     @waltercool this one's axtually dreamshaper xl, so it's already a bit better, but the one you linked doesn't seem to suffer from the softness that much, i'll try it out
   • LS (lain@lain.com)'s status on Wednesday, 09-Aug-2023 05:08:22 JST LS

     in reply to

     • Piggo :verified_horse:
     @piggo it's trained on 1024x1024, so you essentially get high-res images without upscaling. it's also has a two-tiered setup with a base model and a refiner model that adds details. Overall i wouldn't say it's better than a tuned 1.5 yet, but if you've ever used 1.5 base you know how bad it is, and XL base gives pretty good results already
   • Piggo :verified_horse: (piggo@piggo.space)'s status on Wednesday, 09-Aug-2023 05:08:24 JST Piggo :verified_horse:

     in reply to
     @lain how's this different from just standard SD?
   • Piggo :verified_horse: (piggo@piggo.space)'s status on Wednesday, 09-Aug-2023 05:21:54 JST Piggo :verified_horse:

     in reply to
     @lain 1024 sounds exciting. So basically I don't have to buy a new GPU just yet, but the time is coming. Vanilla SD is indeed crap, but some models make really good results even with like 8 inference steps
     
     
     
     
     
     ...
     although you can jack off to 512px just fine also
     LS likes this.
   • LS (lain@lain.com)'s status on Wednesday, 09-Aug-2023 05:22:51 JST LS

     in reply to

     • Piggo :verified_horse:
     @piggo i think 1.5 is still more versatile and can get you to the high resolutions with more advanced techniques (controlnet, upscalers, etc), but it's a great start.
   • LS (lain@lain.com)'s status on Wednesday, 09-Aug-2023 05:30:17 JST LS

     in reply to

     • Piggo :verified_horse:
     @piggo these ones are upscaled / hires fixed
   • Piggo :verified_horse: (piggo@piggo.space)'s status on Wednesday, 09-Aug-2023 05:30:18 JST Piggo :verified_horse:

     in reply to
     @lain I dunno if I did it wrong, but I tried upscaling and it was mostly quite mediocre and fuzzy, not worth the time. This is with "stable diffusion UI"
   • guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 05:58:05 JST guizzy (in exile)

     in reply to

     • Token
     @lain @coin Yep, we're only just starting to see some finetunes getting merged for SDXL. The best 1.5 checkpoints by now are a convoluted tree of finetunes and merges, so we're still in the very early stages of building that for SDXL. https://civitai.com/models/124419?modelVersionId=136753
     
     As of now, I find the most interesting thing to do is check how SDXL fares with things that SD 1.5 had a hard time doing. Like raccoons :blobraccoon:
     LS likes this.
   • LS (lain@lain.com)'s status on Wednesday, 09-Aug-2023 05:58:47 JST LS

     in reply to

     • Token
     • guizzy (in exile)
     @guizzy @coin i found that it is really good at situations where you want 2 people, something that 1.5 always struggled with
   • guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 06:07:03 JST guizzy (in exile)

     in reply to

     • Piggo :verified_horse:
     @piggo @lain I run it with same number of steps as original pass (set to 0 steps), and the choice of upscaler matters. Latent isn't good. I like 4x Valar and I know a lot of people use 4x UltraSharp.
   • LS (lain@lain.com)'s status on Wednesday, 09-Aug-2023 06:07:03 JST LS

     in reply to

     • Piggo :verified_horse:
     • guizzy (in exile)
     @guizzy @piggo latent can be good for things where there isn't one subject (like a forest) or if you turn the denoising way down, to like 0.3
   • LS (lain@lain.com)'s status on Wednesday, 09-Aug-2023 06:12:58 JST LS

     in reply to

     • waltercool
     • Token
     • guizzy (in exile)
     @waltercool @coin @guizzy yeah, i've done some advanced images with controlnets, tiled rendering and so on, it works but it really takes time, i worked for hours on some of these images.
   • waltercool (waltercool@pl.slash.cl)'s status on Wednesday, 09-Aug-2023 06:12:59 JST waltercool

     in reply to

     • Token
     • guizzy (in exile)
     @lain @coin @guizzy There are extensions for that.
     
     Not perfect, but better than nothing.
   • guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 06:23:14 JST guizzy (in exile)

     in reply to

     • waltercool
     • Token
     @waltercool @coin @lain Yeah, the Regional Prompter technically blows it all open, but it's tough to use.
     
     I think they recently added SDXL support, I gotta try that.
     LS likes this.
   • waltercool (waltercool@pl.slash.cl)'s status on Wednesday, 09-Aug-2023 06:47:46 JST waltercool

     in reply to

     • Token
     • guizzy (in exile)
     @guizzy @coin @lain I can't lol. Reaching anything like this is almost impossible without Lora
     LS likes this.
   • guizzy (in exile) (guizzy@shitposter.club)'s status on Wednesday, 09-Aug-2023 06:47:47 JST guizzy (in exile)

     in reply to

     • waltercool
     • Token
     @waltercool @coin @lain Yeah, but thankfully SDXL can give pretty good results without LORAs
   • waltercool (waltercool@pl.slash.cl)'s status on Wednesday, 09-Aug-2023 06:47:48 JST waltercool

     in reply to

     • Token
     • guizzy (in exile)
     @guizzy @coin @lain I don't wanna give bad news, but you may still face some Lora/Lycoris issue.
     
     Not a problem of SD, but the way Lora works.