This model is also uncensored, and it is absolutely nailing anatomy, building geometry, and object solidity. I'm about to try physics, but this is jaw-dropping.
@@1littlecoder It nailed "A hand holding a pair of scissors casting a shadow on a wall" on the first try. This is the first model ever to understand shadows. That's genuine emergent intelligence, not something you can add to the dataset! :-O
@@jonmichaelgalindoit Is a data set thing actually. Tagging of data set with and without makes it understand shadows. A computer simulation of the sane photos with different angle shadows and no shadows could let back propagation teach itself
@@ickorling7328 I'm not going to argue with you, but I promise modeling physics is fundamentally different from memorizing shapes. A "shadow" doesn't have a shape. It can't pull an image of a "shadow" from its memory. The same is true for counting. There isn't an image of "four" it can pull, counting is a mathematical abstraction. Same is true for reflections, transparency, perspective and scale, and plenty of other things. A "cat" has a defined space of 3d image features that are inferred and memorized like a lookup database. Physics is fundamentally different. It's a set of models that have to be computed like logic puzzles. Only specific architectures can do that. The wrong architecture never can, no matter what data you give it. (But if you still disagree, you're entitled to your wrong opinion. Don't worry about it.)
❤lets go we want more competitiveness! i love it! and openai would get tempted or itchy to drop something new and exciting if they have something better under the hood!
Tried it on pixel dojo and it seems to be partially censored. Unless pixel dojo has a back-end site censor, which I am 99% sure he doesn't, I had a hard time getting the naked bottom half to show on prompt.
But same old 1024x1024 resolution. Is there no way to generate atleast double that amount without local install? I do not like upscaling at all, it completely ruins landscapes.
it's pretty good, does a lot of things well, but it does not work for me. I've tried around 20 prompts and refuses to follow basic instructions about paintings, poses, effects, etc. It is good but not that good.
@@1littlecoder I mainly use Dalle3, SD and Pixart (I like experimenting). I tend to get results that I like, even if they dont follow the prompt 100%, but Flex seems to give very "instagramy" or "corporate" results. For me, those are really boring. I like dynamic poses, effects, paint strokes, etc and Flex returns 0% of that. It's too early, and I have to experiment more, but it seems this model is not for me.
@@1littlecoder a simple prompt like "watercolor, big hand on a small mouse" does not work at all. it does not get "watercolor", "big hand" and "small mouse". I tried 3 times, and I only get a hand on a mouse (well done, but not what I wanted.
I honestly don't get why in the world people even want all that disgusting stuff. I'm obviously commenting so find out if there is actually a lesser known good use case for it and not something disgusting.
That's a huge category, but if you mean just full anatomy, it 100% does in my very tame tests. (Museum art pieces: Michelangelo's David and the Venus de Milo.)
Run Schnell it on Google Colab - ua-cam.com/video/rsSQJ-ACPaM/v-deo.html
This model is also uncensored, and it is absolutely nailing anatomy, building geometry, and object solidity. I'm about to try physics, but this is jaw-dropping.
@@jonmichaelgalindo yeah literally the first release
@@1littlecoder It nailed "A hand holding a pair of scissors casting a shadow on a wall" on the first try. This is the first model ever to understand shadows. That's genuine emergent intelligence, not something you can add to the dataset! :-O
@@jonmichaelgalindo you seem to be in love with this model
@@jonmichaelgalindoit Is a data set thing actually. Tagging of data set with and without makes it understand shadows. A computer simulation of the sane photos with different angle shadows and no shadows could let back propagation teach itself
@@ickorling7328 I'm not going to argue with you, but I promise modeling physics is fundamentally different from memorizing shapes. A "shadow" doesn't have a shape. It can't pull an image of a "shadow" from its memory. The same is true for counting. There isn't an image of "four" it can pull, counting is a mathematical abstraction. Same is true for reflections, transparency, perspective and scale, and plenty of other things. A "cat" has a defined space of 3d image features that are inferred and memorized like a lookup database. Physics is fundamentally different. It's a set of models that have to be computed like logic puzzles. Only specific architectures can do that. The wrong architecture never can, no matter what data you give it. (But if you still disagree, you're entitled to your wrong opinion. Don't worry about it.)
Tried it and its Pro prompt adherence was absolutely crazy! Crazy good and really surpassed every other image generator. The don't freaking lie.
@@Cingku did you try on replicate?
@@1littlecoder yep. But now wants to try locally in my PC too, dev and Schnell one. So happy with this news dropped.
Replicate is mostly unresponsive in my case.
The real question is that is it uncensored? 😈
❤lets go we want more competitiveness! i love it! and openai would get tempted or itchy to drop something new and exciting if they have something better under the hood!
Tried it on pixel dojo and it seems to be partially censored. Unless pixel dojo has a back-end site censor, which I am 99% sure he doesn't, I had a hard time getting the naked bottom half to show on prompt.
But same old 1024x1024 resolution. Is there no way to generate atleast double that amount without local install? I do not like upscaling at all, it completely ruins landscapes.
i hope the slim it down for GPU poor like me my 4GB 1050ti only able to handle SD 1.5
What about consistent styles?
it's pretty good, does a lot of things well, but it does not work for me. I've tried around 20 prompts and refuses to follow basic instructions about paintings, poses, effects, etc. It is good but not that good.
@@redanilap has those problems worked with mid journey before?
@@1littlecoder I mainly use Dalle3, SD and Pixart (I like experimenting). I tend to get results that I like, even if they dont follow the prompt 100%, but Flex seems to give very "instagramy" or "corporate" results. For me, those are really boring. I like dynamic poses, effects, paint strokes, etc and Flex returns 0% of that. It's too early, and I have to experiment more, but it seems this model is not for me.
@@1littlecoder a simple prompt like "watercolor, big hand on a small mouse" does not work at all. it does not get "watercolor", "big hand" and "small mouse". I tried 3 times, and I only get a hand on a mouse (well done, but not what I wanted.
Thanks for sharing
Tried to get API key on their site, but it says on invite basis only currently?
How can we run it locally safely?
@@MisterWealth they've released a comfyui node but I guess you'd need GPUs for that
Do we have to download it from hugging face?
If you have GPU and want to use it locally
is that paid or free? how to get acess?
ua-cam.com/video/rsSQJ-ACPaM/v-deo.html
@@1littlecoder thank you!!
Can its run on 2080 ,12 gb ram ?
Just tried a FP8 version on Colab. Don't think it'll work on 12GB yet
SCHNELLLLL
Bhai, please do a series on llama and langchain
All use cases.
Thanks
Does it do NSFW content?
I honestly don't get why in the world people even want all that disgusting stuff.
I'm obviously commenting so find out if there is actually a lesser known good use case for it and not something disgusting.
They explicitly mentioned that you shouldn't do that
@@1littlecoder tested, it gets close, but does not
That's a huge category, but if you mean just full anatomy, it 100% does in my very tame tests. (Museum art pieces: Michelangelo's David and the Venus de Milo.)
Lol, what? Proper release of SD3?
@@bolon667 nope a different company
Spiritually, that seems what it would be. Promising, promising, promising.
This works very well even better than ideogram, following the prompt really well.
Yep, even the low-tier model!