How to UPSCALE with Stable Diffusion. The BEST approaches.
ฝัง
- เผยแพร่เมื่อ 29 ก.ย. 2024
- Learn about ControlNet, ESRGAN, Superscale, "Hires Fix", the scripts "SD Upscale" and "Ultimate SD Upscale". We use Stable Diffusion Automatic1111 to upscale an image created by using the epiCRealism model.
epiCRealism on CivitAI:
civitai.com/mo...
Model database (Superscale, etc.)
openmodeldb.info/
Ultimate SD Upscale:
github.com/Coy...
ControlNet:
github.com/Mik...
--
Useful videos:
• How to use Stable Diff...
• How to create a bootab...
![CAMPปลิ้น | EP.82[1/2] แก๊ง 3 สาวเพื่อนซี้บุก CAMPปลิ้นครั้งนี้ห้ามเสียอาการ](http://i.ytimg.com/vi/_6WUjXHe1TM/mqdefault.jpg)
Thank you for providing these techniques for scaling applying Stable Diffusion! As someone who has experimented with numerous AI techniques for image enhancement, I am always seeking for the most successful ways to maintain quality.
Thanks a lot for your feedback!
Most of the time I use Ultimate SD Upscale to improve images. My usual workflow is: Generate a set of images at 512x768 -> select the best one -> send it to img-to-img -> upscale by 2 with the same settings and prompts and a low denoising at 0.15 -> send the upscaled image to inpaint -> mask the face, set a prompt that defines the details of the face and generate a set of new faces at 512x768 and a denoising around 0.25 to 0.55 depending on how much you want to change the face (combine with face-swap if neded) -> send the enhanced image back to img-to-img -> upscale by 2 with the same settings and prompts as the original image and a low denoising at 0.15 -> done! Since a blurred face stands out much more negatively from clothing, for example, this is a good way to increase the quality.
The woman in the example picture is very typical for the model. It has a tendency to always produce these or very similar faces. My method also helps against this. But it works even better with face-swap.
Indeed several custom models generate one certain type of face or even body or object. Thanks for sharing your workflow.
Why not generate in 521x768 and then use Hires Fix..? Hires Fix creates more details but with low denoising it keeps the image faithful.
my workflow:
544x960 if I like the image ->
Hires fix with the same suit at x2 to have an FHD and have more coherence with low denoising ->
Extra tabs x2 to have a semblance of 4K.
I create with a 4060 base.. the whole process takes less than 3 minutes.
讲的很棒，谢谢老师
Thank you, I'm glad that my video was helpful.
My VRAM is 6Gb and I can easily hires.fix upscale to 2x using SD 1.5, you should be able to do so, too.
By the way I came here because the text said, I would learn about ESGRAN, Superscale, etc... but I really did not other than to download and use basic settings with it... I was looking for a more detailed info on how these work, etc... and not just another video of someone who used the settings "everyone" does in the 1000 videos before on this topic. I would have liked to see more upscalers and how they differ, what they are used for, what to pay attention to, etc... The video was a nice superficial touch on the subject.
Thanks for your feedback.
Thank you for the great video. Liked and subscribed. I recommend people play with much higher denoising strengths, even up to .7 or .8 if the image is very low resolution or has a lot of macro blocking from older compression of video stills or still images.
Thanks a lot for your feedback and the sub - and thanks for sharing your experience regarding denoising strengths!
i dont have ESRGAN folder, what should i do man?
Simply create it yourself like other viewers in the comments - it will work.
SDXL is a waste of time, I can generate an FHD image in less than two minutes with the SD1.5 models. Then I go through the Extras tab and bring it up to 4K "even if it doesn't add details".
I have tried all the extensions to do Upscale, but when I want to create a 4K they create too many artifacts in the face of realistic images.
Thanks for the tutorial. I've tried many times the last option you shared with controlnet but I keep receiving this error message: " RuntimeError: Given groups=1, weight of size [320, 4, 3, 3], expected input[1, 16, 72, 72] to have 4 channels, but got 16 channels instead Given groups=1, weight of size [320, 4, 3, 3], expected input[1, 16, 72, 72] to have 4 channels, but got 16 channels instead". I have no clue of what that means😅
Thanks for your feedback. I have never seen such an error message. Nevertheless, if you are using a Flux checkpoint, this could possibly lead to such an output. Automatic1111 does (currently?) not support Flux.
@@NextTechandAI Hi， the new version now supports Flux Dev😃
@@JojoBenta Thanks for information, but are you sure we are talking about Automatic1111? According to the readme and the open/closed issues there is still no support - in contrast to ComfyUI.
@@NextTechandAI Sorry I was talking about Forge Webui, since the interface is the same, I thought it was more or less the same. After some research Forge Webui do not provide controlnet for Flux only to SD models so far.
@@JojoBenta Right, Forge WebUI does support Flux. Although it's not as fast developed as ComfyUI, I'm quite sure it will see ControlNet for Flux, too.
Title Is Incorrect: This Is an AI source upscaled Image, Not an Actual Photograph.
How do you come up with a photograph? It's not mentioned in the title or the thumbnail.
@@NextTechandAI yes, but by default, when you're searching for something to upscale and improve the quality of a photo, you're referring to a photo taken by yourself, not generated by an AI source. TH-cam's search engine also suggests this video assuming by default that you want to enhance the quality of any photo rather than one generated by the same AI model. :)
@michal5869 Well, that wasn't my intention; by mentioning "Stable Diffusion" both in the title as well as in the thumbnail it's clearly hinting to AI. Nevertheless, there is no default in TH-cam's search engine (in fact there are several), it does what it wants :)
You can use --opt-sub-quad-attention on amd gpu's to avoid crashing mentioned in 6:30.
Thanks a lot for the hint. Especially with 4x it's slow, but it works on my machine, now.
Looks like SD3 now is the future :X
I'm hoping for advanced custom models based on SD3. From what I've seen, SD3 has improved in terms of art and text, but still happily produces 6 finger hands :)
Brother, don't forget to inform us whenever rocm and paytorch come for windows
No worries. We still have to be patient, MiOpen is not ready yet and this is the first step in the chain.
I think it is close there are 12 pr‘s left but all of them are either very small or already done
18:22 if you've installed the best extensions and want to know how to run it
So, what exactly is your question or remark? :)
@@NextTechandAI it's more for me to bookmark the best method
@@no-handlesInteresting. Well, good luck with upscaling methods :)
Hi, I have a question for you that might be a good idea for your next content.
I'm new to using Automatic1111 and the file size is getting bigger, around 40GB. I'm using ControlNet and have downloaded the extension or model seg, depth, etc. for ControlNet.
When I use depth and see the preview of the depth image, will it download again in the CMD section? I thought that once I had downloaded parts of ControlNet like depth, seg, etc., it wouldn't download again, but it turns out it still downloads other parts in CMD. This also applies to extensions. Is that how it works? Or will it only download the remaining parts in CMD when using depth, etc. for the first time?
Hi, this sounds strange. Indeed you download model files etc. for extensions like ControlNet once, only in case of an update you should download new versions. That's why we download these files manually in the vids, they're usually not automatically downloaded.
Thanks for Video, just subscribed, because latest video about AMD gpu generation, hope to see more tutorials around AI, and maybe something for beginners without knowledge about using Linux.
Thanks for the sub! For sure there are more videos around AI to come. This includes Windows and Linux. As we still don't have a full blown ROCm on Windows, currently Linux is the best choice for AMD GPUs. Anyhow, I try to be as detailed as possible in my vids.
Have you tried any workflow to get the same results as Krea or Magnific?
Beating those two non-free tools with free software? Not yet :)
5/5
Thank you :)
what is your specs?
AMD RX6800 16 GB VRAM, AMD 3700x, 32 GB RAM.
demn lol tough for me to run then@@NextTechandAI
Great video and crystal clear explanation, you saved me a lot of time, thank you!
I'm glad that my video was helpful. Thanks a lot for your motivating feedback!
Not just like but I also give u sub, very useful info!
Thanks for the like and the sub. I'm happy that my video was useful.
Using SD ultimate upscale, once you reach a resolution higher than 2048x2048, the image gets split into multiple pieces, leaving you with seam issues. I cannot really get rid of these, i fiddled around with padding and mask blur, it got better but once you reach resolutions of 16.xxx times 32.xxx and you zoom in where the tile was split e.g. an eye, you can easily tell something is not right.. Do you have a good way of creating super high definition images?
You have to use the ControlNet-way from the vid. Upscale by only x2 each run.
How would you go about, upscaling an image 32 times from e.g. 512x1024@@NextTechandAI without getting blurry tiles seams?
See ultimate upscale with ConrolNet in my video. Do it five times with 2x, resulting in 16384 x 32768. Nevertheless, I have only tried three times with 2x so far (8K).
GREAT VIDEO! Loved how you covered different methods with examples while covering different bases of what you can do. Great job. Keep up the good work. :D
I am glad that my video was useful, your inspiring feedback is the best motivation. Thank you!
Great video - with help with the parameter weights such as denoising,etc
Thanks a lot for your feedback! I'm happy explaning the parameter weights was helpful.
Why don't I have the ESRGAN folder in my "Models" folder? I only have 5 there : *deepbooru* , *karlo* , *Stable-diffusion* , *VAE* , *VAE-approx*
Please help 😥
That's strange, there should be more than a dozen subdirectories including Lora and ControlNet. Do you have a different WebUI version? You need a current version of Automatic1111 WebUI.
I didn't have it either. I just added the folder myself and it worked
Thanks for hinting on this.
@@NextTechandAI Using latest version, it only has a couple folders. Think you're mistaking sir.
@@fatallyfresh3932 Yes, this changed in one of the latest versions, but as already hinted by one of the comments above you can simply create e.g. models\ESRGAN yourself and put SuperScale etc. there. It works.
Thank you! I’ve been searching for an upscaler and I’m glad I found one.
I'm glad that my video was helpful. Thanks for the feedback.
Very helpful, thanks!
Thanks a lot, I'm glad it's useful.
Subscribed, TY
Thanks for the Sub!
Thanks but what is that extension that shows you when you are resizing this orange box on the picture?
Thanks for asking, but could you give me one more hint or the timestamp regarding the 'orange box' you are referring to?
@@NextTechandAI 10:56
Now I understand. That's automatically done by WebUI. When you adjust the height or width in the Hires/Resize section, the changes in size are visualized with that orange box in the referring image.
@@NextTechandAI Thanks! :) I did not know that when you scale with the tool it does this because I manually inputted numbers.
That makes sense :) In a video I prefer the sliders cause of better visibility.