[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/3/ - 3DCG

Search:


View post   

>> No.983626 [View]
File: 108 KB, 340x444, 983.png [View same] [iqdb] [saucenao] [google]
983626

>> No.974810 [View]
File: 108 KB, 340x444, schizotime.png [View same] [iqdb] [saucenao] [google]
974810

>>974416
Goes from "Clean SDXL" model to "Poisoned SDXL" model.
Yeah nah. That ain't how it works.
You're trying to tell me that some randos re-trained base the SDXL checkpoint from scratch with "poisoned" images instead? Fuckin what?
At best, they can either train a lora based on the original SDXL checkpoint, or train a completely new checkpoint from scratch based off of SDXL principles. By no means can they re-train the already released SDXL checkpoint from scratch.

If it's the latter where they just trained their own checkpoint. Then their results are disingenuous if they're comparing it against the OG SDXL. It could be just as likely that they're fucking shit at training, or intentionally used shitty images to skew results. Doesn't mean that their "poison" worked, unless the results are repeatable from a 3rd party.

I mean for fuck's sake, all things being equal (seeds, cfg, prompt, etc) those poisoned images should still look pretty similar to the outputs of base SDXL, just extremely fucked up. Yet all of them are completely different (different car models, dog breeds, colors). Using the same seed is pretty fucking powerful if you have a similar prompt, results should not be that different, even in their own model. Even between 100 samples and 300 samples, there'd be signs from using the same seed where you can see that even though the results are pretty different there's a common thread between them.

OP image looks like purposeful cherrypicking and not a good faith comparison.
Anyway, postin in a schizo thread. Let's fuckin go.

Navigation
View posts[+24][+48][+96]