Dissecting SD3

Just a few clarifications because you can't possibly cover everything in 20 minutes video 1. "bad hands" has never worked unless maybe the model is somehow trained for it. I used it in the video just to prove that it just generates noise and you might have better luck with just "aaaaaaa". 2. A SAI engineer suggested me to not use negatives at all, because of the way SD3 is trained they have no meaning. From my tests very strong negative concepts seem to have some kind of effect, to be proven if it's just coincidence / placebo, but seems consistent. 3. Changing the resolution to weird numbers (1040) is a bit like running a different seed. It doesn't fix the model, for some reasons I get less oddities at weird latent sizes, which is of course not recommended anyway. The point is: the training was meh 4. Give SAI the benefit of the doubt. If things won't change I'll be the first to stop using it (and actually paying for the creative license)

"Since at Stability AI they only do the missionary"...damn bro, fatality! 😆😆😆

"Let's confuse the little bastard." I got a good laugh out of that. Thanks Matteo.

Best cover of the subject so far. 100% informative. No time waste. No drama. No bs. Yet somehow humble. 👏👏👏

"Carried by ants" This guy is a comedian

I know you are not considering yourself a youtuber, but what a great summary! As a someone semi casual it is cool to learn what's good about it, what's strange about it, how to wok with it in comgyui, and some glimpses on what might happen next with the license without a drama. Observations >> Opinions. To me your channel icon under a thumbnail is a bigger clickbait than a thumbnail itself. Well done!

I can't stop watching SD3 videos. It's like a beautiful house the community loved but the owner set it on fire for safety reasons. So bizarre

I absolutely love watching your videos, I learn so much! That's for staying on point with what matters!

I chuckled when you said "Whispering to it like a lover" but when you said the second part I almost spilled my drink :D Great video as always.

Controlnet and IPadapter are the only reasons I would try agaim SD3

If you consider the implications of some of the recent papers regarding the Platonic hypothesis, as well as groking and other related concepts, it seems fairly expected that teaching these things anything untrue ("safety training"), is going to be devastating to the model's capabilities, and likely result in permanent shortfall relative to the technical potential. I suspect community models have gotten so good because of some cracking and overcoming some of the toxic false training, but I'd bet even those models never reach the potential the architecture would allow, because they are fundamentally salted with falsehoods. "If you make a delicious sandwich, and then put some dog p00p in it, no matter how good the rest is or how many other ingredients it contains, what you've got there is a p00p sandwich!.." - someone smarter and funnier than me oh I can't remember or find with AI LOL

very informative as expected of you, i rushed here right away when i saw that you posted about SD3 and i was not disappointed, appreciate your efforts

Thanks so much for this one, I hope things will be sorted out and change for the better! 😊

Thanks.. as always one of the best places to get information. For me never the less SD3 medium isn´t anything more but a playground. I won´t suggest it in my company on legal images, as long as there´s no real benefit, i could allready get from good trained SDXL Models. If i as an Architect could see great advantages in prompting and generating very detailed and subtile and architectural correct Architecture, esp. in very big cityplaning scenes, it would be worth thinking about a payed license.. Thats a weakness of nearly all Models out there.. People dont care about architectur to much.. there you need by far better precisiion , because you often allready have a concept and architectural drawings and CAD screenshots to start of from. But i think thats for the future to be seen.

از آموزش های بسیار جذاب شما بسیار سپاسگذارم شما باعث ارتقا سطح روابط میان کاربران و مهندسان نرم افزار هستید و برخورد شما خوب و منطقی بود ممنون

Excellent, informative video. Thanks.

Informative and occasionally hilarious. Love it. Hope you're doing well.

thank you for sharing your work, knowledge and experiences. very appreciated!

the negative prompt doesn't affect much cause sd3 needs higher weights for keywords. like in sd1.5 and sdxl we did stuff like (keyword:1.5). here we need to do stuff like (keyword:15) for it to show significant effect. so increase keyword weights in negative prompts does make a difference.

The difference between the negatives which seem to work and those that don't seem to be because of the negative prompt cut-off at 10%. Words like blonde might be set in the image in the first 10%, words like artefacts would not. I accidentally finetuned an SD3 model with blank prompt negatives instead of zeroes, and it worked fine in samples when the training setup was testing it that way, though not in comfy which uses zeroes, so it seems you could retrain the model to use classical negative prompts if you wanted.