SDXL1.0 is finally out. You can use it to generate all of these amazing images and more. It can do various styles, types, subjects, backgrounds and so much more. All for free. Here are some more realistic examples, some more anime style examples, and some random stuff that I found interesting. So how do you use it? I'll show you three methods. The first is running it locally using automatic1111 the most popular UI for stable diffusion. I'll also show you how to run it locally with comfy UI, which had the best support for SDXL since launch and also natively supports the SDXL refiner. And finally, if you aren't able to run it locally for whatever reason, I'll show you how to use it online for free. So to run it locally, you first need to grab the SDXL 1.0 base model, which should be good enough for most people. But optionally you can also download the refiner model as well. This is not required for Auto1111. To do so, just go to the huggingface repository by Stability AI that I'll link in the comments and click the download button on the .safetensor files of the base model as well as the refiner model. To use it in automatic1111, first, go to the official GitHub repository with over 90,000 stars that I'll link below. Unlike my previous tutorial, you don't need to install python or git, this time since we■ll will be using the pre-built file. Scroll down to installing and running and under installation on Windows 1011 with nvidia GPUs click the v1.0.0 pre hyperlink. Now click on sd.webui.zip to download the compressed file. After the download has finished, please extract it using whatever tool you'd like. Now copy the earlier downloaded SDXL base model and paste it after going in sd.webui webui/models webui/models and finally the Stable-diffusion folder. Now all you need to do is run the update.bat script by double clicking it. This will just update the code to the latest version which supports the SDXL. Now, similarly, double click run.bat This step might take some time as it will download and install all the required dependencies such as PyTorch. After it has finished installing, it will start the server on localhost:7860 You can Ctrl+Click the link in the terminal to launch it in your own default browser. You can also close everything and now if you double click run.bat It will start the web server every time since it has already installed everything. Let's try generating an image by entering a positive prompt God of flowers fantasy setting and a negative prompt worst quality And it generates a cool looking artwork, but that's not exactly what we are looking for. Since SDXL was primarily trained on 1024x1024 sized images, it generates much better if you just change the width and height to that. This does double the time taken to generate, but it generates actual images based on our prompt. You don't have to stick to 1024x1024 , for example, if you want a landscape image you can use 1366 into 768 are any of these other resolutions that are also natively supported? Again, the time taken to generate will probably increase, but it does a pretty good job at generation. Now let's try a photo realistic prompt as well. Honestly, much better face and details than 0.9. Let's do some landscape generation as well and super interesting results I've only ever generated square results. So this was nice. Now if auto1111 doesn't work or if it's too slow or if you want to use the refiner model as well, ComfyUI might be a better choice for you. So all you need to do is go to the official GitHub repo of comfyUI that I'll link below and click on Installing ComfyUI or just scroll down and click on direct link to download or you can go to the releases page and this top download link will download the latest stable build and this bottom link will download the latest unstable build. We■ll just used its latest stable build for now. After downloading, we just need to extract the 7zip file using your favorite utility. After extracting go within the ComfyUI folder, then the models folder, then the checkpoints folder and paste here SDXL 1.0 base and refiner models here that you had downloaded earlier. Now all you need to do is double click the run_nvidia_gpu.bat file if you have an nvidia GPU or alternatively to run_cpu.bat file, if you don■t. This will launch ComfyUI in a new browser window on localhost:8188 If you had pasted the model in the correct folder, it would automatically be loaded here. If you have other models as well, you can use a dropdown to select the as the SDXL base model and reload. Now let's enter the same positive prompt and a similar negative prompt. Let's change the sampler to eulerA and let's change the width and height of the empty latent image to 1024x1024 Let's do a landscape generation as well. Wow. This is the first time I've had this color and design. Now let's add the refiner model. We can add nodes and create a new workflow, but for now, click on load and load this refiner.json file, which was for sdxl 0.9, but should work fine with 1.0 as well. This is courtesy of camenduru. Now make sure to load the actual refiner model in the load check point of the refiner and the base model. In the load chekpoint of the bas After loading the refiner, let's generate and compare. It might take slightly longer since we're using a different sampler, but feel free to experiment. It first generates a base image and then feeds that image along with the prompt back into the refiner. As you can see, the refininer model added a lot more details to the face, the background and small minor details to the variety and kind of flowers, textures, shades and lighting. Let's generate one more image this time a landscape one. Interesting results, slightly more detailed face and hand. At least one of them. Changed the crown from gold to plants and some minor details and touches here and there. Again, let's try the same photo- -realistic prompt that we had to used in Auto1111 earlier and another one with landscape resolution. Okay, if both of these were■nt able to run due to hardware limitation or software errors, for now you can try it on clipdrop.co which is a product by Stability the company behind SDXL. There is a limit (400) to the number of free generation. But this should give you an idea about its capabilities. I■ll make a video on how to run on colab properly. Since there are some resources available but they aren■t that fast or optimized for the SDXL yet. Do subscribe for more stuff with SDXL including LoRA training Dreambooth, ControlNet running on low vram devices, etc.. I'll see you soon, hopefully after I've gotten over this cold. Bye.
