After some further testing, it seems the same sort of issue occurs when using the TAME checkpoint (albeit less severely), which is unfortunate because I presume that's where YAPRM gets most of its skin texture realism from.

Maybe it's because that model has some lightning stuff merged into it, but we're using it like a normal checkpoint? But from playing around with it, TAME does not seem to follow prompts very well in general.

@DrA_263 You are likely correct that the issue is related to TAME. There are multiple SDXL models merged into TAME, including both Lightning and Hyper models. I've done my best to iron out the issues resulting from that, but it's not perfect. I haven't used YAPRM yet (although I look forward to testing it out when I get a chance), but I can tell you that TAME is notoriously picky about schedulers.

I imported one of your images that gave the artifacts into TAME and indeed there was some weirdness there too (not quite an artifact but weird looking clothing that didn't belong). This was instantly solved by switching from the Uniform to the SGM Uniform sampler (TAME works best with SGM Uniform and Simple unless you change your sigma min setting). Removing the NGMS (Negative Guidance Minimum Sigma) setting that you have on also solved it. And finally, removing the score tags also solved it, while improving the prompt following (however that also made the lighting less interesting, likely because there was nothing about it in the prompt).

Again, I haven't used YAPRM yet, so can't say whether it works the same as TAME, but maybe this can give you a starting point.

@zyxt99565 Hi! No idea how you saw this post, but thank you for looking into it 😄

Tested again with: 3261737187
I've just removed NGMS, swapped to SGM Uniform and removed the score tags, but unfortunately the same issue still occurs in YAPRM. Swapping some of those back in I sometimes get the artifacts to show up as clothing like I imagine you're seeing in TAME, but they never fully disappear.

As for on TAME itself, I still get artifacts on that seed regardless of the other settings until I change to 'SGM Uniform', but with other seeds the artifact clothing returns regardless of scheduler or anything else.

Example of a bugged one in TAME with all the suggested changes:

amateur photography, flat chested, topless
Negative prompt: worst quality, compression artifacts, young, child
Steps: 20, Sampler: DPM++ 3M SDE, Schedule type: SGM Uniform, CFG scale: 7, Seed: 886948088, Size: 832x1216, Model hash: ae50f0a320, Model: tamePonyThe_v25, VAE hash: 74bd301605, VAE: s.safetensors, Denoising strength: 0.2, Clip skip: 2, Token merging ratio: 0.2, Token merging ratio hr: 0.2, Hires upscale: 1.5, Hires steps: 15, Hires upscaler: 4x_NMKD-Siax_200k, Version: v1.10.1-4-g48239090

@DrA_263 I saw it by chance while browsing the latest Pony checkpoints/updates. Quite an intriguing bug you've discovered though. .It seems to be sampler specific, it occurs with DPM++ 3M SDE/ DPM++ 2M SDE and DPM++ 2M SDE Heun, but so far none of the others I tried. Usually Euler a is the most stable/obedient but for the latest version I've been recommending DPM++ 3M SDE because I found the results to be a bit nicer. But given this little bug, maybe I will have to go back to recommending Euler a! :-)

@zyxt99565 Yeah I generally stick to Euler a or DPM++ 2S a, but for realistic models I definitely feel like the SDE ones give better skin texture, which is what I'm after really.

@DrA_263 Well I didn't notice the bug with DPM++ SDE, at least in TAME.. so maybe that one would work?

Oh n/m it does occur at lower step counts, only not if you keep the steps at 20. 😄

@zyxt99565 I think the steps might be the thing.

I noticed that 'DPM++ SDE' takes me about twice as long at 20 steps as '2M SDE' and '3M SDE', so I figured maybe it's the same sort of deal as 'DPM++ 2M' where you need double the amount of steps to get the same quality of results as other samplers?
As such, I tried '3M SDE' with 40 steps and all the results have been artifact-free in both YAPRM and TAME, even with my other settings back in place except for the scheduler (which still bugs out as 'uniform').

Never mind, just got artifacting again with 40 steps DPM++ 3M SDE / SGM Uniform 🥱
Guess I just had a good roll of RNG for the 20 samples I'd done.

But yeah, DPM++ SDE seems fine, though subjectively not as nice for the details.

@DrA_263 Yes, DPM++ SDE is very slow because it does 2 passes on every step. But it works out about the same because it also gives good results with fewer steps. With TAME you can usually get decent results at 10-12 steps or even less, whereas most other samplers need a lot more.

@zyxt99565 Okay yeah, that makes sense.

Have to wonder how well TAME would do without the lightning/hyper stuff in it though, because this has turned into a lot of caveats to get good output 😅

@minghia No worries, I think we figured it out now 😄

Though why not share the mixture percentages? Is it a secret family recipe? 🤫

@DrA_263 Some caveats yes, but I didn't achieve that level of realism only through custom training. The Pony models in TAME, with the exception of Goddess of Realism, are mostly interchangeable without losing too much.. but the SDXL models, where most of the Lightning and Hyper stuff comes from, are quite important. It wouldn't be TAME without them. Also, it makes the model quite fast, which is important for people like me who have old video cards. ;-)