Stable Diffusion 3 HANDS ON! How Good Is It Really?
VloĆŸit
- Äas pĆidĂĄn 16. 05. 2024
- đš Try Stable Diffusion 3: www.pixeldojo.ai/?...
âš Support my work on Patreon: / allyourtech
đ»My Stable Diffusion PC: kit.co/AllYourTech/stable-dif...
Stability AI released Stable Diffusion 3 and Stable Diffusion 3 Turbo via API, so I got to work building out a user interface over at PixelDojo.AI where you can use both models to generate your own images. How good are these models? Were the press photos all cherry-picked, or is the text adherence and prompt adherence as good as they claim? Check out the hands-on review and judge for yourself! - VÄda a technologie
đš Try Stable Diffusion 3: www.pixeldojo.ai/?
Ideogram is way better, and you get way more images for your buck, and it's way more coherent with prompts since the 1.0 update. I was never able to really use sd for anything good before for free. And I don't think I will be paying for this any time in the future. They need to stick with free. Still say open source, but it's not anymore if they start charging for it... Just my two cents.
Two cents*
@@archael18âïž
My first couple generations were seriously blurred as if they were being censored, and I didn't ask for anything explicit or naked. I didn't play around with it much after that. Kinda sad. Dreamstudio started doing this a few months ago with XL, and I'm not a fan. Thanks for adding 3.0 though. Great for the Dojo and thanks for not censoring it on your end.
Once the full models are out hopefully i'll have the option to bypass whatever safety checks they have. I hate it when they are overly cautious in these API's.
I just wondering after we started to use sd 3 checkpoints, is it possible to generate images in 512x512? Or minimum in 1024x1024 like in the sdxl crap?
Min is 1024 I believe
@@allyourtechai that's shame then
@@gokhandanismend why?
@@imresomodi4961 sd 3 will surely be as good as mj, dall-e in prompt understanding and all but if results are generated as 1024x1024, then it means it won't be suitable for 8 gb vram machines and it's sucks but some people says that sd3 checkpoints are may be published with optimization for working properly in 8 gb vram machines and i hope it happens because i have a 8 gb vram
â@@imresomodi4961 sd 3 will surely be as good as mj, dall-e in prompt understanding and all but if results are generated as 1024x1024, then it means it won't be suitable for 8 gb vram machines and it's sucks but some people says that sd3 checkpoints are may be published with optimization for working properly in 8 gb vram machines and i hope it happens because i have a 8 gb vram
So nice of u to let us try
How well does it handle anime? are fine-tuned like NovelAI/animagine still king or is SD3 doing something special for anime-style prompts?
I haven't tested it thoroughly on anime. It's capable of producing anime for sure, but i'm betting it will be much better at it once we have open models that we can fine tune like sdxl.
Bruhhh those testimonials on the website. Really sick đđđ Loved it.
đ€Ł I always wonder if anyone makes it that far on the page
@@allyourtechai đđ
when local install?
No idea, they donât give any indication of when that might be
I'm a big fan of the DOJO :)
My God. I use it so often, I don't think you make money out of me..
I'm sorry (Canadian)
Haha! Love it!
@@allyourtechai I feel like I confess myself now :)
Thank again :)
is it wll be localy or not ?
They say they will eventually release the models, but for now itâs only accessible by api
It looks like T5 is disabled there. The SD3 paper gave examples of what would happen if the T5 model was disabled. The quality of text consistency was deteriorating there. I hope the T5 can be turned on locally on the second graphics card.
Fingers crossed. This is hitting the APIâs from stability ai directly
ХпаŃĐžĐ±ĐŸ ĐŽŃŃĐł. Thank you
Thank you for watching!
you mentioned the pro plan is $9.95 a month in this video? Says $25 a month on your website unless I am mistaken.
Correct, Iâve added 6 new features since then, and raised the price to help cover GPU costs. Still the least expensive on the market
@@allyourtechai Okay, i understand. Thank you for your reply!
What's the point of sd3 if it as censored as mid journey and dalle 3, atleast Mj and dalle 3 has better image quality than sd3, i hope that a less censored version of sd3 will be available in the future.
So do I. Once they release the weights and files we should be uncensored.
would have been nice to have tested what hands are like
Good call, I was just using press sample prompts, but i'll keep that in mind.
Yeah. You have to state outright if it's censored. We could guess but if you're selling access to it you should say. It's fucking ridiculous what they censor.
I had this up within 3 hours of launch. I donât think anyone knew how censored or uncensored the model was.
And I agree. They go way overboard
@@allyourtechai Gotcha. Fair enough. I may try it out just to see what it can do.
I use stable diffusion but its censorship is ridiculously exaggerated
Yes it is!
Bring back Emad.
how you can do a video like this and not try a human face is beyond my comprehension
I just focused on the images and prompts from the press releases. Feel free to generate some images on pixeldojo