So many great models and merges coming out now but without a doubt this is my number one favorite currently. Loads of fun. Thanks for putting it out here for everyone.

I just tried this for only like 2 hours, but i already have to write a 50 page review, and ask: Why is NOBODY talking about this after almost a month? Why are there almost no comments here? And why is it mogging even flux this hard?

You know when something immediately looks like "Well, this seems too good to be real, has to be a scam, they surely photoshopped all those images etc." Thats kinda what i thought. But because it (somehow?) still was under 7GB like all other models ive seen so far, luckily there was nothing stopping me from trying.

So i did. And i was pretty damn surprised. I never used Flux, Midjourney or Dall-E before, only a bit of bings copilot, but i already can say, its closer to that one then to literally any SD model ive ever seen. It looks ridiculously realistic. On the recommended settings, with a naturally written prompt. I didnt see a VAE mentioned, so i tried with the standard SDXL VAE and without. Actually it didnt seem to be much of a difference, but i think using it might improve things a bit. Its hard to tell tho cause it already looks so good, and i might be tripping and maybe it does literally nothing.

Btw, for a pic at 1080x1520 (i like this because 1920 stretches too much), i need around 45s... on AMD. Which is still impressive because it will always be at least 20x the time you need with any nvidia gpu. At least under windows.

To test i used a short prompt (longer ones are prob better and give more details ofc), with just something "simple" i could never get ANY other SD model to do (except flux, couldnt try that yet), just to try and already push this model to its limits: A bearded dragon. You know what every other model gives you instead of the lizard? Thats right. A dragon. Sometimes with a beard.

This one made a correct bearded dragon. In realistic detail. With pretty good eyes. On the first try. It even gets the claws right! Its crazy. That alone proved to me, that his model is different. It did sometimes have slight characteristics of other lizards, but this model still has a "limited" data set, and the tags for bearded dragons are probably too intertwined with general lizards to separate them completely, so thats fine.

But of course this still is SD. While the detail was actually very realistic, sometimes it still had 2 tails, an extra limb, a missing finger, however, that was basically it. Otherwise the anatomy was often shockingly good. And with something as complex as i tried, these mistakes a very forgivable and fixable. Also i couldnt get the lizard to do the most complicated things like human poses (yet), but when not specifying too much, it looks actually pretty natural. I even got it do stand up, with just a prompt.

The backgrounds are very realistic and natural too, but seem to love blur, however its the right kind of blur, and makes it look more like a realistic photo. I probably could get rid of that if i tried, but thats one of the 500 things i have to try later, including any different style then realistic. And while backgrounds can sometimes still have some weird parts that tell you its AI, its not a lot, not very noticeable, and often hidden enough by mentioned blur.

I need to try a lot more. and while i cant compare it directly to flux yet, i think from what ive seen i can say: The realistic detail seems to be mostly on par with flux and things like copilot, sometimes even better imo. But flux has likely even better prompt understanding, however this seems already very close. Lastly, i tried making the flux trademark, a sign with words on it. It didnt quite work, BUT it actually got somewhat close! I tried just "Hello", and sometimes it actually got it in very good detail, but very often with other words, most times multiple of the same, or written wrong. But to my surprise, even when written wrong, it almost always was recognizable as the word and the letters looked like actual letters. So it would often be "Heelo" or something, which might not be too useful to actually use this model by itself to write things, but to me its impressive non the less.

What i would say is, it might in some key ways not be on flux's level yet, but the details and realism are, and its like "The pizza costs half of the even more delicious one, but tastes close enough, so its worth getting if you cant afford the other."

Btw im going for the longest comment written on this entire site did you notice? /s

Ok seriously, sorry for the wall of text, just give this a try already! Once i tested enough and got a pic im very happy with ill post something too. c:

Superb model, nice prompt coherence, super fast and high quality🎖️🎖️🎖️

with some prompts it seems like magic. but as soon as you try to do something a bit more strange/unusual or add a lora (or negatives, since you can't), everything goes off. but the speed is quite amazing