Hmm, what do you mean? At least to me the detail greatly exceeds base Pony, quality is more difficult to gauge because the styles are so dissimilar.

@Sharlin What I mean is the the experience is less rich, there are few details, the model tends to fixate and might be overturned for certain features. For example, background details are often ignored, it's hard to change the shot distance, the model doesn't seem to like changing hairstyle, or ethnicity. Right now I'm trying to get a range of ethnicities with different hairstyles in a gen, but the difference between Chinese and Lithuanian is very light skin tone changes, the difference between blond in brunet is is a few strings of blond hair mixed in. Which is not to say that I don't like the model and do not continue to experiment with it.

Nice! Much better, for some reason I was getting awful outputs with suggested settings.

@SleezeBagDiffusion what's interesting to me is using the lightning model with my settings above really isn't much faster than just using the main model with DPM++ 2M SDE Karras at ~ 25 steps and hires.fix at 1.5x with 16 steps. I get higher quality images using the main model too so I'm likely just sticking with that one. I assume the lightning model is still good for people who are okay with using Euler A but it's a little too soft for my liking. Overall great work here either way and it'll open up a lot of possibilities for people on lower end hardware. Even with my 3060 I'm still perfectly happy with the speed I'm getting on the main model.

@ZyloO thank you, I tried that but still no luck. Reinstalled webui too. It was working great! Then suddenly this.

By "Noise" you mean a completely fractal Image, or just a too grainy Image ?

If it is the first one, remove any Lora´s you are using, and try again, If it is still there, have a look at Brackets in your Promt. These Applications don´t like missing Brackets ;-)

@Draftsman completely fractal. Without loras too, and with very basic prompts. Thing is, I had loras trained on a base pony model, and this model worked. And I loved it! Just, I really don't know what I could have done in the middle, one gen it suddenly just went all fractal.

Like, if I use it as a refiner model too it just starts "noising" the image. Really scratching my head on this.

@coffeebrown0383 There is also another Reason for this: Your Graphic-card runs out of Memory...besides all this...switch to ComfyUi ;-) ...I mean, you are already at the Point of (re-) installing everything, why not give it a try ?

This seems to be a problem with the Pony models. Some keywords interact poorly with each other and can cause your image to become pure noise or start bearing little relation to your keywords whatsoever. When this happens, start looking at what words you most recently added to the prompt and remove them, one by one. When you find the culprit, you might be able to bring it back with different phrasing or by identifying any other keywords that might be conflicting with it. It can be very frustrating, but once you know how to deal with the problem, you can quickly recover.

def agree man, this has become my favorite model!!

Oddly enough, even though the text warns against Euler A, I get the best results with this setting (using ComfyUI and Krita AI Diffusion local run) Steps 41 cfg 7.0

Also I highly recomend Everclear https://civitai.com/models/341433/everclear-pny-by-zovya

@ZyloO I'll play around with it a bit, see what works. Thanks for the answer :D

@depressed_llama Something to do with how many steps per preview pic being updated.

Bumping the amount of preview steps up reduced it for me and it happens way less frequently now.

Yes, since DPM++ 2M does not work, Euler A is just as fast and images look just as good as with the slower samplers.

@ZyloO I don't see a difference and renders twice as fast

In general it's very difficult to impossible to get anything else than "well-exposed pro studio lighting" which, while nice-looking, gets a bit repetitive. Not sure if there's any Pony model that can do eg. low-key or harsh direct lighting because of the mostly 2d source material.

@Sharlin Hopefully one day tools like IC-Light will be compatible with XL/pony, even so I'm sure that the lighting issue will be greatly improved in the near future or in SD3, pony is just getting started!

As per the name, it's a Pony model. Which is XL, but the compatibility with other XL stuff can be hit and miss; there's specifically Pony trained loras.

so do i,found the solution?

建议用 gigantic breasts，huge penis之类的

Are you using generation data from another image? check the Hires Fix upscaler, it may be trying to use one that you don't have installed.

I'm having the same issue but have nothing in the hires fix upscaler loaded or selected. Were you able to fix it?

@taterdots I found out that I didn't have the same HIRES upscaler as a friend's generation, I had to download the 4x Nomos8kHAT-L_otf upscaler from the OpenModelDB (dot info) . Same as Dreason had mentioned, I added in the required upscaler & it was fixed.

I also moved all my lora models in to one folder & then loaded in the few that I do plan on using this week.

I figured out that I can stop using the AutomaticCFG nodes to fix this issue, sorry for thinking it was a problem with this model.

@ZyloO Hmm, locally I've generally found DPM++ 2M to be unusable (any ancestral or SDE variant of DPM works though). And UniPC, DDIM, and Euler are similarly unusable. In general, I found that only ancestral/SDE samplers work. This was with either sgm_uniform or AYS schedulers.

It's not a huge problem, but it's odd. Any idea why this is the case?

@ZyloO Interesting! I would be quite interested in trying that version of the model.

There are some speed-optimization LoRAs (PCM, Hyper, etc.) that work best with deterministic samplers, so that's where I could see some demand for it. Also, with lower-end hardware in general I find that DPM++ 2M is the fastest sampler (requires the least steps and also has the fastest it/s) so that's also useful.

Same here. No matter what I try, it always comes out grainy.

try using a vae, i noticed the same problem too and it looked a lot better when using a vae

If you are generating onsite, switch your sampler to DPM2a. DPM++ is grainy for some reason.

change sampler

@monsieurvanilla thanks a bunch! also @kfcpri

How do you know it's a merge? I can't see any indication of that?

@RedDeltas under the "download" button there's "type: checkpoint merge" line. It's in the details tab

@coyotte Ah I missed that, thanks ☺️

same here. it generates strange noise and shade...especially for lighting versions

Same, for the lightning versions. Normal version generates normal (non-noisy) images. I did check and applied the recommended settings (with DPM SDE sampler)