Це відео не доступне.
Перепрошуємо.
Stable Diffusion XL Turbo's Real-Time Text-to-Image Generation is Amazing! 👀
Вставка
- Опубліковано 7 сер 2024
- Stable Diffusion XL Turbo is a new real-time text-to-image generation service. The pictures appear literally as you type! Stability AI has managed this speed bump by developing a new distillation technology called Adversarial Diffusion Distillation. Using it SDXL Turbo performs single-step image generation, reducing the required step count from 50 to just one.
---
00:00 - Intro
01:54 - SDXL Turbo demo
04:47 - How it works
09:19 - One more demo!
09:50 - Outro
Let Me Explain T-shirt: teespring.com/gary-explains-l...
Twitter: / garyexplains
Instagram: / garyexplains
#garyexplains
Thank you for another excellent video Gary! You do a great job of explaining the material and keeping your users up to speed on this fast changing technology much appreciate it buddy
Really liked how you explained it all, not super overly complex, yet still enough that i know have a pretty good idea how TURBO models are made from the original SD model. Very cool !
Very well explained. Thanks
How does the model know how to combine the various images in the text query in a sensible manner?
Can you use this to rapidly design/prototype logos?
Would an RTX 2070 8 GB GPU be enough to run it with this speed in my desktop?
how do i install or what is the link?
2:45 Why does the moon duck wear boots with claws?
Why is it playing a guitar on the moon?
@@GaryExplains And then there is the old, philosophical question. If a guitar is strummed on the moon, does it make a sound?
@@GaryExplains Because you ordered it. But you didn't order claws.
@chrisarmstrong8198 Good question! 🤓
@RagHelen How you do know that I didn't subliminally request it? 😲
The speed is quicker than doing a web search, amazing.
Next AI will recommend re humanity!
The model is great, except for one thing: in the style of [artist name] does not work with it.
That's a good thing
@@TyQuinn , heck why? This is essential for Stable Diffusion models.
@@MeinDeutschkurs this model is intended for real time text to image not for standard image generation. Also I've never used an artist name since the very early days of SD, they're just not needed.
@@weirdscix well, they are. Especially if you need something in the style of [name]. It‘s quick but for my use case unfortunately useless.
What GPU was used for that? Does it mean that a lower end GPUs are now able to do this in real time as well or does it still requires some high end GPU?
Vram will always help these models run faster. This might be faster on a low-end card than previous models, but >8GB is still recommended. Ideally for SDXL models 12GB or more.
@@DigitalJedi On a 8 GB RTX 2070, can I expect a similar speed to this?
@gaborkiss1425 It's impossible to say for sure. You definitely have enough vram to mess around with most stable diffusion models if you wanted to get into it, but training will take a while on s 2070.
I'd recommend SD 1.4 or 1.5 models as they tend to be content with 6-10GB depending on the resolution and any add-ons you use such as LORAs or Control Nets which will probably add a bit to the memory needs each. For some reference, my 12GB Titan Xp is fine to run at 768x768 with 2 loras and a heavy control net active. Memory need grows with the square of resolution so 512x512 should fit in 8GB.
@@DigitalJedi Thanks!
This is insane. With great power comes responsibility (that a lot of lazy people will abuse). These tools are awesome, hopefully the will be water marked so folks know it was made with an AI tool.