@Cyberdelia incremental improvement is improvement. Sure, sometimes it's a quantum leap, but we've all come to trust and accept that what you perceive as being release-worthy is very likely to be release-worthy. It's always exciting to see what direction you take things. It's my hope that our enthusiasm is contagious and you keep at it. Just don't burn out. I work in fits and starts too. A week or two of long training runs and sleepless nights, then two weeks of not even clicking on civitai.

@Cyberdelia Yes, absolutely release them, it's appreciated. Seems to improve with every update. Thanks!

It's weird they'd single out any particular Pony model TBH, 100% of them are capable of creating every possible type of TOS breaking content simply because the base model was to begin with lol.

@Cyberdelia It's not for me. I still get the error message.

Works!

I'm still receiving the unavailable message. For some reason, it seems to be in some limbo state where some people can use it and some can't.

Is the on-site generator disabled for this model possibly because it was trained on images of real people?

According to the devs on discord help it's fixed and taking a bit to reindex but that was three hours ago. Feel free to ask again. Or I will in a bit since I really like this model.

@Cyberdelia It's still showing 6.2 and will not generate. The model picture is the one from 6.3. Another user posted this https://discord.com/channels/1037799583784370196/1130679274999910490/1279165228470501396 but never got an answer.

@Cyberdelia Actually Maxfield just said it was still likely cached in the browser and refresh fixed it on brave. Oddly not on chrome but at least it's available now.

@Cyberdelia Indeed it does, keep on keeping on, absolute legend.

I found the same thing happeining.

Update In my initial an typical testing this still held, in extended testing I would say 6.2 really can do the job, it has much better prompt adherence (I ran a prompt for hand placement and 9 out 10 6.2 was on the money 6.0 close but off, ironically as well 6.2 that run had better hands and finger tips on that run. I did a lot of waving tests and in general my initial findings held of course as we all know other than FLUX (which has its moments) hands are never an easy thing with SD.

I highly recommend 6.2. Like all it will throw some bad so seed up and try again but skin, hair, teeth, eyes much better. :)

Further update - I wanted to test against generations without people (no humans) to test general subjects or scenery, I had perplexity spit me out a bunch of random tests and 6.2 scored best in all.

I tried 6 and 6.2 and then 2 other realistic pony models, set each to have the sampler they work best with, fixed seed and ran a bunch. 6.2 gave pleasing results, the other 2 none cyber models would keep occasionally adding people. Abandoned library, old ruins, etc but (at least in my testing) the sybers stayed strong with 6.2 sticking closest to prompt

1 prompt they all failed at, was a sci-fi future lab (I could have played with seed if I was really pushing for a result) but as a sample they all added people despite many negatives to dissuade it. the others were worse adding bikini models in the lab. I guess training set for either sci-fi or labs usually has people in :) , 6.2 always gave me at least 1 humanless output so again 6.2 for a win here.

Other 2 models will get archived and 6.2 will stay.

You mean to use this checkpoint as a base model to train lora, right?

me too! can't wait

Keep CFG Scale to a low number. I also find Hiresfix makes the colours less saturated. Use image software to desaturate if necessary.