FASTER than Stable Diffusion. FREE & JUST RELEASED!

a new generative AI model that's two times as fast as stable Fusion Excel and 15 times as fast to train well say hello to version oh and it's free and you can use it today let's go check it out oh and by the way I took up origami for a while but I gave it up it was just too much paperwork these are all images from version the new generative AI model and it has just been released and you can use it well today fast diffusion for image generation so it's a diffusion model and it works similar to wall stable fusion and other models but it's also a little different and the main difference is actually the compression and where stable Fusion can compress their images to the latent space about four to eight times versus Jam or however you pronounce that can actually achieve 42 times the compression so you know you have a big image you compress it to the latent space so let's say stable Fusion does that version does you know super small the reason why that is important is it's going to be much much faster to to train and you can see that here actually so this is a GPU hours thousands of hours so stable Fusion 1.4 was straight train in about 150 000 hours and version here version two which is 25 000. oh so it's much much faster and you can still keep the solution pretty high now stabilization 1.4 it was just trained on 512 by 5 12 and version was actually trained it says here resolutions going up to 1536 so they did it much faster much cheaper and with greater quality or high resolution and if you look at this little graph here you can see that the time it takes to to render an image twice as fast even in some examples here where the batch the batch size goes up it's more than twice as fast as sdxl so you have them all up spin trained actually in higher resolution than sdxl and can do it faster both in training time and in render time or inference time and I mean this would be great news even if it was just paper but it is available now today there's actually so you can try the demo here it says error here but if you press the little link here you come to this page and here's an example astronaut in a jungle cold color palette music colors detail 8K so you know pretty sweet images we can do Sim cinematics still of a Viking Warrior in Nordic woods so we're running this and we will see here in a second or two so now I'll run restarting here you can see our Viking Warrior popping up and here's the end result now this face because of the messed up this image is kind of cool let's do cyberpunk woman in neon Blade Runner City and we are getting that image in right now there's actually a little bit of a queue as people have been starting to use this this is a demo yes the the diffusers are available and I saw a feature request actually well I didn't see it I asked my good friend roon about it if he could get version in his ruined focus and he found a feature request for for comfy so people are starting to take notice and I honestly expect it to be available in your favorite UI any any any second now here are some examples an anthropomorphic chicken dressed as an officer and the examples here are ranging from 10 24 by 1024 actually 2 15 36 by 1536 here and that's the highest resolution this model has been trained on however you can keep training this and raise the resolution and fine tune it for high resolution because well just how easy it is to train so I expect this to be very welcomed in a lot of areas in the model creating community and also for people that don't have like a high vram GPU because this has less requirements to run so if the quality can keep up with what we're seeing here in these examples and stay on are with the infusion XL and some of the custom 1.5 models I expect this to really really take over because this is a great implementation and if you want to know more in detail on the blog page here there's actually a link to detailed explanation that goes more in depth into the math behind this and actually how how it works so I'm not going to dig deep into that you can read the paper or you can check out the video so check it out I'll see in the next video have a good one see ya

Channel: Sebastian Kamph

Views: 22,464

Keywords:

Id: lIuCLY2Fkzw

Length: 5min 15sec (315 seconds)

Published: Fri Sep 15 2023