That Seed Tip was fire! Maybe you could do a video going a little deeper into Seeds and their controls for people how are bascially forced to use Comfyui if they want to experiment in sdxl
Bro can you please 🙏 update your auto1111 colab notebook to support sdxl1.0, my pc cant run sdxl because of low vram, colab is the only option i have.🙏🙏🙏🙏
Bro i am trying to use Sdxl models (from civitai) in stable diffusion ( google collab notebook) created by you. but its not working??. Can you give a solution??
In majority of cases the problem is a proper link which directly download file. If notebook works with hugging face it should work with any website. Check a link and file in temp directory. And yes, model must be .safetensors or just rename it manually in .ckpt
@@marat_ai you are saying .ckpt works for SDXL but .safetensors doesn't work! Am i right? . So i have to rename model from (SDXL.safetensors) into (SDXL.ckpt). same for refiner model.
@@OtakuInspiration I meant just check proper extensions for model, ckpt should be .ckpt same with safetensors. Automatic1111 consume too much ram so colab works only with paid subs
I've been working with SDXL in A1111 for the past 2 hours. The amount of time it takes to generate a single image and the bad quality output leaves me quite unimpressed. The people look plastic, like the model was trained on overly photoshoped blurry faces. I know it's just the base model, so it's bound to get a lot better once the community starts doing it's magic. But after the low quality ick that it's been producing for me for the past couple of hours I'm going to have to give SDXL a hard pass until some decent community checkpoints (that address it's more serious issues) comes out. The quality that I get from the 1.5 models are just light years ahead of what I can get from SDXL right now, at least without waiting ages for a single image to generate.
Have you used refiner model? It'd drastically improve the final results. I find automatic unusable due to RAM consumption and comfyUI works much faster. I should have added refiner to notebook, but I think I will add it soon and make video about comfyUI.
@@marat_ai It's funny, I've heard that ComfyUI is actually slower than A1111 for SDXL. So I'm not sure which is true. Either way though, it probably isn't in my future to learn ComfyUI, I'm really not a fan at all of the spaghetti strings and all the nodes. It's a big mess and a headache. It's one of the reason I quit trying to learn Blender, same type of strings and node system that's headache to try and figure out. I did try the refiner model with the refiner addon for A1111. It didn't improve the image enough to warrant the extra render time. With my computer it takes 2-3 minutes to generate a 1024x1024 image with SDXL, when I add the refiner it takes an additional couple of minutes. So in total that's 4-5 minutes to generate a single image. And the images still don't look that great even with the refiner. The faces and skin look ultra smooth and plastic-like. With SD 1.5 I can generate the same size image in a fraction of the time. I tried a couple of the models made by the community and am still seeing the overly smooth and plastic looking skin. The results just don't look realistic to me, and it feels a bit like a step backward. For now I'll have to stick to SD 1.5. I'm getting amazing, highly detailed, and realistic results with 1.5 and the images generate WAY faster as well.
I'm using SDXL in Automatic 1111 with 16GB of RAM and 12GB of VRAM and it works really well, even with the refiner it takes less than 1 minute to generate an image. Check if you can optimize your settings a bit more. I've searched for specific settings for my hardware and it works pretty fast. I agree with you that the people looks plastic by default, it is somehow biased to create perfect fashion models but if you work the negative prompts you can achieve great realistic photos. Try for example "smooth face, blurred face, perfect face, soft face, no skin details" and such in the negative prompts. You can also specify an ethnicity so you don't get "perfect" american models. Also, I think ADetailer doesn't blend really well, you can get much better faces by just using hires fix or fixing them with inpaint if they're too far (takes time but it's worth if you got a really good composition). With that I managed to get some good skin details and realistic faces, also the refiner helps a lot as marat mentioned before. Make sure of use the refiner as 20% of the steps, for example I use 40 steps and refiner set to 8. Other tip I can give you, download the SDXL DreamShaper, it's much more versatile for both photorealism and artistic styles, def look better than the base SDXL. I hope this helps, I know SDXL has a lot of bugs but so far I find it really amazing and fun to try
@@latent-broadcasting Thank you for the tips. I've been using most of the tips already but I will give the ones I haven't a try. You have more ram and vram than I do, but maybe some of the tips can help. Thank you. :) I just don't think that SDXL is worth it for me in it's current state when it comes to generating people. Between the much longer image generation times (yes, I use Xformers and medvram), the plastic skin (I've tried many different prompts and negative prompts to get more skin realism and I've been using DreamShaperXL10 and a couple of other community SDXL models), and the need for a refiner which makes generation even slower on my machine, it's just not worth it for me to use SDXL right now for people. At least not for generating images that are meant to look like photographs of people. Toon type characters and other NPR styles are another story. I'm sure in the next couple of months there will be huge leaps forward in the image quality from the community's models though. There was a massive leap in image quality a few months after the release of 1.5 too. So I just need to be patient! lol One thing that SDXL does that really impresses me though is color and composition. I really like the image compositions better than with 1.5. I'm definitely looking forward to the next few months. I know we'll be seeing some amazing things. And I'm almost positive someone's going to figure out a way to get SDXL to run almost as fluidly for lower/mid range machines as 1.5 does. It will be a happy day! lol Thanks again for the tips! I'll give the ones I haven't tried yet a spin. I appreciate you taking the time to give some helpful advice!
