The first part is how to update Stable diffusion to support SDXL 1.0. Let’s take a look at the current version of the stable diffusion web UI. Pull down the screen to show the version: 1.4.1 is the old Abbe. The current version of the stable diffusion web UI needs to support SDXL , you must update the stable diffusion web UI to version 1.5 or that old saying. Before upgrading the revised stable diffusion web UI, you must first copy the old version of the SD web UI as a backup . The updated SD web UI deletes and then changes the name of the copied backup folder\back to the original name. The advantage of this is that you can buy an insurance. The disadvantage is that your hard disk space needs to be enough. Assuming you have bought insurance, you will first install stable diffusion . Close and open the file manager, go to the directory where you installed the stable diffusion web UI, click here , enter CMD and press the enter key , this will open the command prompt under this directory Enter git pull and press the enter key , it will directly connect Grab the SD web UI files that need to be updated online and do the update. If you are interested in what file to update, you can pull up the screen to check it. If you are not interested, just close this command prompt. The official has mentioned that you need to add a The parameters are the same, go to the bottom of the video , copy this section of parameters, switch back to the file manager, click the right mouse button on the batch file to start SD, select edit, and paste the parameters just copied here. After pasting, you can save it and save it. Close the file to open the stable diffusion web UI. In addition to the normal startup, there will be some model sha256 caculating checks. After the inspection, the stable diffusion web UI will be opened normally to check whether the version has been updated to 1.5. If so, congratulations, you can start Models that support SDXL The second part How to get SDXL checkpoint and Lora resources to open AI drawing Good friend civitai.com There is a filter option here, click SDXL 1.0 , it will help us filter out all SDXL 1.0 models This checkpoint is an official checkpoint, and there will be two 1.0 versions available for download. This is the SDXL 1.0 base model, and this one is the SDXL 1.0 Refiner model. What is the difference between these two models? According to the official statement, refiner can generate more pictures Let’s take a look at the pictures provided on the Internet for more details. The next one should be prepared for the 4K image quality. Ok, back to civitai, this is an unofficial SDXL checkpoint, and it has already supported SDXL 1.0. There is a merge checkpoint here. It is the official version of merge. The author claims that it is better than the SDXL 1.0 Base model. Then it is the Lora of SDXL 1.0. At present, there are these three Loras. Click to have a look, and you will find that the Lora files of SDXL 1.0 are quite large. Each of them is about About 1.7G seems to be running out of hard disk space. As the name suggests, SDXL is XL size 1024x1024. XL Models are also XL above GB. The third part of SDXL 1.0 is a simple trial. Go back to the stable diffusion web UI and select SDXL 1.0 base model for the stable diffusion checkpoint. The prompt parameters of the steak map generated by version 0.9 use the default value first. If such a map is run out, it is a problem with SD VAE. Please select SD VAE to be closed or automatically closed and then run it again to see if you use SDXL 0.9 to run the picture SDXL checkpoint file. About 6G or so, the loading time is obviously much longer. Then run a picture with SDXL 1.0 refiner, and run a picture with chiloutmix for comparison. Seeing such a picture, do you know which side needs to be adjusted? It’s the size. Chilloutmix belongs to SD 1.5 , the training size is usually 512*512, so here you can use high-definition restoration to enlarge the picture size to see the comparison of these four models. As for which of the four is better, I will leave it to you to compare. Old Abe prefers to eat the middle Let 's see how much the memory of the display card will be eaten up under different parameters. The current one has no limit on the vram parameter. The memory of the display card can be eaten up to about 11.4G, and then we will test the stable diffusion. Adding the --medvram parameter shows that the memory card memory is consumed up to about 8.1G. In this picture, if the memory card memory is not set or the --medvram parameter is set, each it takes about 1.5~1.6 seconds and then the display card is used. The memory parameter is set to --lowvram, which will consume up to 5.7G of the display card memory , but the speed drops to 13 seconds per it, which is more than 8 times slower than the previous two settings. Here you can see the display card memory. How important is it to stable diffusion Finally, let’s take a look at how much graphics card memory will be taken up if you don’t limit the graphics card memory and use chilloutmix checkpoint to run the graph. All settings and parameters of the picture are the same . The screen on the side is cut, but the running speed of each image is about 1.5 seconds, which is about the same as the SDXL 1.0 base model. The current update supports SDXL 1.0. In addition to the memory consumption is too high, there are still some problems that need to be solved. For example, Lora often crashes. , Halfway through the picture, it will stop there and wait for the suggestions if you need it. You can wait to see if the official will make any corrections. The above is today’s sharing. If the video is helpful to you,
