Run Stable Diffusion real fast at up to 75 it/s on Linux with RTX 4090 (TensorRT)

hi my name is and I'm going to show you how to run image generation with stable diffusion model 1.5 at speed 75 iterations per second there are a lot of videos on how to do it on on Windows but there are none on how to do it on Linux it's important because especially Mac users don't have an option to run stable division generation model efficient ly accept renting a Linux server with a GPU on it and do your exper experiments there one of the GPU Linux providers are is RP pod and I'm going to show you how to do it there so you have to go over the link that I left under the video description of this video select for 4090 RDX 490 press deploy press continue again press deploy and go to your paths wait for a bit I already did that so I'm going to just stop this and delete because I don't need need that because I already have one then go to connect go to 8888 board then go to your sidebar go to stable diffusion web UI directory open run a 1111 notebook focus on the cell press command to return run the cell and it's going to run automatic 1111 webui it's going to download some model which is stable Vision 1.5 which is required for you to have imag generation because without it you cannot and I'm think if you are watching this video you already know that but in case you're newbie or something like this never used automatic 1111 and only used something like diffusers for image generation I'm just telling you that it's downloading automatically base model which is stabil efficient 1.5 okay so go back to your run po account and press on HTTP connection to Port 3000 and it's going to open Automatic 111 web UI go to transfer RT tab press 512 512 image size and you have to you have to keep in mind that you have to pre-bake this kind of dancer model for each model and also for for each size I guess you probably can also generate some like Dynamic model but I'm I think that is going to be slower so I'm just showing you how to do it in some basic way for some basic resolution of images so press export engine the same as usual the same as in any other you tutorials for Windows and check your notebook tab to make sure that something is going on and I'm going to continue this video once it's finished I'm just going to pause this and go back to you once it's finished and we'll proceed from there thank you after it's finished scroll down to the bottom make sure that there is line telling t RT engines has been saved to disk go back to your automatic 111 webui press refresh near sdn drop down press on it press trt and name of your base model then go to text to image type something image of account and press generate the first time it's going to be slower like twice as slow it's going to be like as regular image generation without Daner RT but like flower the second time and all consequent times it's going to be much faster like as you can see it's 75 not 75 for some reason but it's like 60 let's do some testing of 30 images press generate and check that it's like 65% reason and I have no idea why because previous time it was 75 but even though it's 10 iteration per second slower it's still okay 15 iterations per second slower it's still very fast and we have a lot of flower pictures amazing but again keep in mind that have to pre-bake tensor RT model for each stable diffusion model so it's not that like you can use this thing as a magic speed up thing so you have to like carefully pre-bake transer RT model for each stable def model and go from there thank you

Channel: Oleh Kopyl

Views: 197

Keywords:

Id: MPD4OBWD_So

Length: 5min 52sec (352 seconds)

Published: Sun Oct 29 2023