The results are much better when you delete the prompt during the resize process. On your final picture, you can see different persons standing on her hairs. I had the same issues, but after removing the prompt the upscaling was perfect. :)Very nice video!
These are some great results! I agree to do some traditional Photoshop cleanup and fixes + generative fill and inpainting before the final upscale. Going to have to give this a go.
Thank you so much! The quality of 4K ultrasharp is so much better than all the defaults! Literally every image looks a few times better than what i've been producing previously.
@@ArtificallyIntelligentAi Awesome stuff. I have a strange issue where with certain loras the image comes out really noisy, but it's perfect with other loras. Funny thing is im using loras created by the same person.
Thanks for this guide, I've recently started doing some fidling arround with A1111 and you have provided some of the most usefull guides I have seen, with this one as cherry on top.
The most important step is the latent upscale, everything else is a mediocore improvement at best. I still have better results with GigaPixel AI. If you want to compare your results, you don't look at pixelation, you should instead upscale your lower resolution image to the same size using a good classic non-AI upscaling method like Lanczos and then you can compare the images. That way you'll most likely discover that the difference between your 4k version and the 8k isnt worth the computing time and the image may become actually worse. My suggestion would be to do latent upscale as high as your VRAM allows and use the control nets just to keep the image barely from falling apart (steps, weight). If possible use the exact same model and settings that were originally used to create the image, that has a huge impact. Also the Pixel Perfect setting doesn't do anything for the tile control net.
You kinda touched on what I was thinking as I use GigaPixel as well but I thought this would be better because it's adding detail that didn't existing as part of the upscale process and I'm not sure that's how Giga works?
@@Space-O-2001 GigaPixelAI also adds new details that makes it so far superior to any classic upscaler. But it can only do so much without any additional information. Yet it still can yield better results than upscaling with SD using suboptimal setting. The thing is the more you force SD to keep the upscaled image unchanged the more you restrict it (either by low denoising or with control nets) the worse the result. In worst case the quality is the same as with a non-AI upscaler. GigaPixel AI delivers by contrast always better version of exactly the original image. Now latent upscale is another story, SD can here work in latent space, that's were the original image is formed before it was transformed to pixel space. Feel free to look up how all this stuff works. ;-) It would be optimal if we just generate a higher resolution image without any upscaling right from start but even if you have enough VRAM, you don't have a model that is trained for such high resolutions and so you get deformed copies of things in the trained resolution. You could finetune the available models to high resolutions like 1024x or 1280x, that works in principle, the model will adapt but if you are not careful, it breaks. But the computing power required for high resolutions increases apparently exponentially, so you can get only so far. Using img2img and latent upscale is a nice trick. Without control nets the result would be trash if you go over the training resultion but with control nets you can keep it a bay. But this only works best if you use the original model and settings as a base and have the control nets only interfere as little as possible. If you take a different model you need a much stronger control net effect and get way worse quality. So what is the best upscaling method? It depends, I'd say, if you want an upscaled version that's as close as possible to the original and the images are not already generated but real photos, I think GigaPixel is the way to go. If the images are generated by diffusion or you don't care much about smiliarity, use SD and latent upscale and maybe GigaPixel on top. Though 4k or 8k upscaling is IMHO pointless with SD either way, at least when using a common 512px model.
@@testales Many thanks for taking the time type such a considered response. Half the battle with learning is knowing the correct terminology so my next project will be reviewing latent upscaling :P. Have you played around with ControlNet>Reference only yet?
@@Space-O-2001 Of course I did. :-) But I can't add much helpful information here fore this. It all takes long time to figure out what works best and in what situations for what purpose. For example I didn't have to time to get a feeling how multiple reference images work together. There are so many ways to guide the generation of an image, it's insane. For example I just recently saw what you can do with the loop script and that you can have parts of your prompt only have affect earlier stages of the generation. And of course I'm still in search of the perfect upscaling method and also one for videos that bascially recreates the frames in a better quality but keeps consistency.
@@testales I hear you, I also have played with the loop script. I appreciate it's a result of being early days and ever changing landscape but given the amount of permutations available during the creation process the UI could do with some 80:20 pre-sets based on what you are trying to achieve. That said it could be argued this is where the skill lies and it is, for the most part, faster than learning how to draw/paint/render yourself.
Let's take a look at Control Net Tile Upscales that allow you to render 8192 by 8192 Ultra 8k Resolution. If you don't have control net yet check out this video here - github.com/lllyasviel/ControlNet-v1-1-nightly/blob/main/README.md Also, I just uploaded the 4x Ultra 4x Upscale. It seems I forgot to add the link when I first released the video. It is in the video description now.
Thank you so much for this video. I have learned so much from you! I thought I was making ok photos, but I wasn't lol. But does make me wanna upgrade from my 3060 RTX to something better.
I tried this on a picture of the Martian landscape, but the picture came out with faint tile lines and lost a lot of detail. Not sure what I did wrong.
During the latent upscale or the tile upscale? It could be a diff model was used to create the image, custom models will add their own flare. Check out this site, stable-diffusion-art.com/controlnet-upscale/ the third method is tile upscale. This is go to site for all things stable diffusion
Great stuff as usual, I just wanted to point out that the Multidiffusion (Tiled Diffusion & VAE) is its own upscaler. I did not see you ever enabling it in the video. The ultimate SD upscaler script can work fine without it, aswell as Tiled Diffusion & VAE can work fine without ultimate sd upscaler. I noticed it due to an error i got (something about image does not match size) while trying to generate an image while having both of them enabled at the same time, which did not work.
I think I prefer Tiled Diffusion w/ Tiled VAE. Seems like it does a better job with not showing the seems that can appear with Ultimate SD upscale. I found an upscaling workflow from someone recently that I think does a good job. From initial 512x512 img to 1.4x HiRes fix (txt2img) or Img2Img resize by with low denoise strength like .15 to .3. Then drag that new img to img2img and again 1.4x, then again 1.4x, and once more 1.4x (now at 1967x1967), then Tiled Diffusion(or Ultimate SD upscale) 2x up to 3888x3888
Deng do you 24gb of vram? I should of forked up the extra cash. I got the 4070 ti but sitting with 12gb vram. Thanks for the knowledge. I always wondered if it was a stable diffusion 2048 limitation or a video card one.
i looked for a tutorial like this for months... Thank you so much. I only have 4gb Vram, but i trying to generate the best images i can... Thank you again!
@@ArtificallyIntelligentAi about 10-20 minutes to get 2k resolution, im not trying 4k or 8k for now hahahaha I leave it working and i go play with my son on the garden. He makes this time pass like a eye blink
Yes, I’m the beginning of the video I said I was only going to show how to install it so you can test it because I can’t get good results from it and I need more experience with it.
hmm ive got a question, how do you make the 'same' picture you've done at one platform to another platform? previously im generating pictures over at stable diffusion web, but now im trying to do everything over at automatic1111 but i cant seem to get my 'style' over at the 1111 side. anyone can help me with this? 'style' being the face and stuff, and the models im getting arent that close to what i did over at stable diffusion web. so i cant seem to replicate the work im generating over at the stable diffusion web. cuz over at the stable diffusion web you dont have to pick any models to begin with, its just adding your prompts and adjusting the guidance scale/seed etc.
No problem, I’m sure the tile upscale can but u prob want to keep the de noise low to prevent the photo from changing its details. My original pic was so bad I had to up the Denoise to create details that weren’t there. But I’m sure it’ll work.
i dont understand why i get always bad results, like worst than the original pic, what looks like macroblocking, the image just looks blown up to 400% with no visible changes
All of the tiles have differentiating edges, and it makes the whole image seems like a puzzle. Main model used: A-Zovya RPG ControlNet settings: Preprocessor: tile_resample Model: control_v11f1e_sd15_tile [a371b31b] Any idea? Would you be able to help if I dropped more of my settings? Do you think I should change the main model?
@sovietchild Yes! Ditch Ultimate Upscaler, and also ControlNet for upscaling. Under Scripts, run SD Upscale, put the first slider up to max (which if I remember was around 256), choose 4x Ultra, Esrgan, or whatever you like, and it should look amazing. Tell me if it doesn't work and I'll tr6 explaining it while I'm in Stable Diffusion
you can remove every prompt that describes the character and only leave some unspecific quality prompts, that way you should not get these small characters in the image
Ultimate SD upscale doesn't do the work, because it makes tiles. On 8k it's very visible. Go to Photoshop, add extra curves layer and play with curves to extreme. You will see that girl is under the grid. The only method which works is Tiled Diffusion. But I have other problems with it. I have rtx 3090 so I can upscale to 4k without problems, but when I want to upscale to 8k I have to enable Tiled VAE (to avoid Cuda out of memory) and then I get not so good upscaled result, but at least it's not tiled. I still didn't work out, how to get the same good result as it was on 4k without Tiled VAE. 1:45 even that grannie picture is tiled.
I'm new to this and I was wondering if this only works with Nvidia graphics cards and its Cuda cores, I have an AMD card RX 6800 with 16GB VRAM and I want to try these things.
Hi there, so it does work on AMD but it's more of a workaround and they hasn't been very many updates. I don't have an AMD myself so I can't test it but there is the link for the install and it's just an additional GIt command it looks like github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs then here is a install tutorial on it ua-cam.com/video/8xR0vms0e0U/v-deo.html. I saw the full video and the only part thats diff is at 2 min to 2min and 30 into the video. It seems you need to create subfolders for the webui then paste in the command that is specifically for AMD. That's the only diff thing about the amd tutorial I see. There are also free sites like playgroundai, leonardoai, and clipdrop that let you create stuff through stable diffusion through their sites for free with limitations on how many you can create like 500 pics a day or something.
Check out my video description I will put videos in order if your staring from scratch so it’ll be the stable diffusion automatic 1111 install, custom modes, control net, the canvas zoom extension and then this video
Ive seen other people do it w/o the img2img so I def agree. I haven’t done too much testing without images in ctrl net so I didn’t want to. Once sdxl is fully compatible with control net and ultimate upscale I will prob update this video and test it out
does this work with anime? I've tried it a few times but my picture was too blurry, any more tips? its hard to find good info on upscaling or inpainting. there's a lot but not all of it is very useable or relevant and the anime models are kinda different
Thanks for the video! Is there an extension you're using to zoom into those images to inspect pixels within auto1111? Or is it a different application?
OMFG! Thank you for this. An absolute gamechanger. I've been using 4x_Foolhardy_Remarci in the extras, which is pretty good and fast but this is next tier in terms of fidelity. It's far more time consuming but the results are worth it, especially for print!!! Legend!
Great video gonna keep watching your videos :) Btw the Interrogate clip seems to take a very long time on mine never seen it become finished, even after 10min how long should it take?
It will download a model the first time you run it. It's about 900MB, so it might take a while if you have a slow internet connection. You can look in the console window and see if it's downloading.
But it may vary on internet connection speed, img, size and VRAM but it your trying to interrogate a clip that's 4k size you can wait up to 5 minutes on average. 10 minutes is a bit long but possible if your have an 8k img. If you have at least 8gb of vram, a 500mbs or greater connection and a img under 4k you shouldn't wait more than 5 min. A 1024 pic will take 30 seconds or less. The Site Hugging face also have CLIP interrogators that will do the same thing, quite a couple sites have it actually.
When I did this I got some body horror on the level of SAW. The person looked so cut up and scratched. I literally got nervous from it. (The shock that is). I have no idea why that happened. I mean it was sharp just nothing like the original in skin damage.
Thank you sir! Will try it out later. I tried upscaling with the 512 tiles once, but that gave me strange color variations in the final image (basically, you saw the tiles). I will give it another shot with the tiles and denoise stuff once I'm home
I got that too one time , I forgot what I did but it was a bunch of tiles as well with some little people on each tile. I think it was when I upscaled with the preprocessed not matching the tile model
@@ArtificallyIntelligentAi I tried it yesterday and I can say: everything worked out pretty well. I did my very first 4k x 4k upscale and I'm impressed by the details, no boxy tiles or something. You changed my workflow with your video, thanks!
crazy results, thanks! I"ve been comparing results with gigapixel and the SD version is much cleaner on textures lightning and even the hair, I tried a lot of setting on gigapixel but they always end up blurry. is there any way to save the settings as a template?
I had this one time and I think it was when the pre processor didn’t match the correct tile model for control net. If it’a each tile that’s prob the issue if it’s the img has a whole lower denoise and ensure downscaling isn’t too high. It might allow u to use less vram but it lowers the quality of the image so it can recreate new details.
I got a secret, it's not bad trick - if you blow the image up to 4000 by 4,000 and you like it but do you want to inpaint. Resize it back down to dimensions that your computer can handle, so usually I'll put the back down to 1000 x 1000 or 512 by 512. Do the inpainting then blow the image back up to 4000 by 4000 or whichever resolutions that you were at. I only using 4,000 by 4,000 as an example.
Well... Unless you want to inpaint a half of the picture, lol. But I think if you've let it go this far, without taking any action, it's your fault ¯\_(ツ)_/¯
This is the deliberate negative embedding that I was using civitai.com/models/30224/negative-embedding-for-deliberate I would test it with that xyz plot script I used to find the perfect denoise before upscaling
how? when I use controlnet upscaling just like you did in the video anything that Im upscaling looks worse than before.... whys that? please tell me I m trying to get higher resolutions for months and nothing works.
My denoise was set to high so lower that first, hence the little people floating in my pic but it shouldn’t just be worse quality , make sure your upscaling with the same model you created the img with. I made that mistake before. Also if your using sdxl I’m not 100% sure if it’s compatible right now. I will have to test it when I get a pc again
I've never been able to get tiles to upscale the full iamge for me, it takes an hour or two to upscale an 800x600 image by x2 and then only upscales about a quarter of the image the rest is just the old image resized. Hope some of this will help but most of it i've already tried.
Hey there! Great video. Super easy to follow....you did a good job. Is it possible to run this with a 1080 Ti? I keep running out of ram even after following your suggestions. Thanks!
1. why did you also install pkuliyi's multidiffusion? it was not used at all in the video? 2. what does controlnet actually do when combined with ultimate SD upscale? I have been using ultimate SD upscale for a long time just by itself.
1. It was so you can test it out and see what results you get. I did mention it but it was brief. It was also recommended by most people but I didn’t get any results from it. 2. Provides the tile upscale when you go above and beyond your cards limits to process it in smaller chunks rather than giving u an out of memory error. So if you were running a 3090 you might not even noticed since it has 24gb of vram. Depends on your card and how large the upscale is though
Problem for me, error: gather(): Expected dtype int64 for index Time taken: 3.92s | GPU active 2741 MB reserved 3030 MB | System peak 4325 MB total 16384 MB
This use to work fine, but recently perhaps after some upgrades I am getting this error when I want to use to controlnet, any one else is getting it " AttributeError: 'ControlNet' object has no attribute 'label_emb'"
The results are much better when you delete the prompt during the resize process. On your final picture, you can see different persons standing on her hairs. I had the same issues, but after removing the prompt the upscaling was perfect. :)Very nice video!
What about setting Denoising to 0?
@@Ichigolificacioon Denoising 0 means no change 0.1 - 0.2 mean subtle changes, over 0.5 change the image
Very occasionally the YT algorithm recommends me something actually interesting, like this time. Subscribed.
These are some great results! I agree to do some traditional Photoshop cleanup and fixes + generative fill and inpainting before the final upscale. Going to have to give this a go.
It’s def my go too for hen I want a really clean pic I did one that was 10,942 by 8192 recently so it goes even higher
You are extremely thorough. Thank you for making these videos!
Thanks!
Anytime and thanks for the appreciation!
This is incredible! Thank you so much for the video! It was well done and edited as well. Going to check your other vids soon! :D
Awesome thanks. Definitely going to make more slowly but surely. Eventually I'll speed up the pace.
This is the upscaling workflow I’ve been searching for since I’ve gotten back into sd. Stuff was looking off when I followed other tutorials. Tyty
Amazing stuff ! Thanks for these tutorials !
wow! thank you for this amazing demonstration! Amazing what comes out of her hair, those small figures. I want to see my old digital photos upscaled 🤗
Thank you! It worked and I used 0.25 for denoise and I got the same exact photo but upscaled!
Nice! Ya, my denoise was too high on this one. 0.25 is a lot more reasonable and won't create the little people like in my example.
I have seams between the tiles .... can you help me ? :(
Good info. I'm currently upscaling a 4k rendering to 8k on my i7-10700k@5ghz + GTX1080 and man is it taking forever. Time to upgrade my GPU for sure.
Thank you so much! The quality of 4K ultrasharp is so much better than all the defaults! Literally every image looks a few times better than what i've been producing previously.
Awesome! Glad it helped. I can't go back to normal upscales either, even if I just doing a 2x.
@@ArtificallyIntelligentAi Awesome stuff. I have a strange issue where with certain loras the image comes out really noisy, but it's perfect with other loras. Funny thing is im using loras created by the same person.
Thanks for this guide, I've recently started doing some fidling arround with A1111 and you have provided some of the most usefull guides I have seen, with this one as cherry on top.
I cannot tell you how grateful I am for your videos, I've learned so much about SD and plan to upgrade for SD.
The most important step is the latent upscale, everything else is a mediocore improvement at best. I still have better results with GigaPixel AI. If you want to compare your results, you don't look at pixelation, you should instead upscale your lower resolution image to the same size using a good classic non-AI upscaling method like Lanczos and then you can compare the images. That way you'll most likely discover that the difference between your 4k version and the 8k isnt worth the computing time and the image may become actually worse. My suggestion would be to do latent upscale as high as your VRAM allows and use the control nets just to keep the image barely from falling apart (steps, weight). If possible use the exact same model and settings that were originally used to create the image, that has a huge impact. Also the Pixel Perfect setting doesn't do anything for the tile control net.
You kinda touched on what I was thinking as I use GigaPixel as well but I thought this would be better because it's adding detail that didn't existing as part of the upscale process and I'm not sure that's how Giga works?
@@Space-O-2001 GigaPixelAI also adds new details that makes it so far superior to any classic upscaler. But it can only do so much without any additional information. Yet it still can yield better results than upscaling with SD using suboptimal setting. The thing is the more you force SD to keep the upscaled image unchanged the more you restrict it (either by low denoising or with control nets) the worse the result. In worst case the quality is the same as with a non-AI upscaler. GigaPixel AI delivers by contrast always better version of exactly the original image. Now latent upscale is another story, SD can here work in latent space, that's were the original image is formed before it was transformed to pixel space. Feel free to look up how all this stuff works. ;-) It would be optimal if we just generate a higher resolution image without any upscaling right from start but even if you have enough VRAM, you don't have a model that is trained for such high resolutions and so you get deformed copies of things in the trained resolution. You could finetune the available models to high resolutions like 1024x or 1280x, that works in principle, the model will adapt but if you are not careful, it breaks. But the computing power required for high resolutions increases apparently exponentially, so you can get only so far. Using img2img and latent upscale is a nice trick. Without control nets the result would be trash if you go over the training resultion but with control nets you can keep it a bay. But this only works best if you use the original model and settings as a base and have the control nets only interfere as little as possible. If you take a different model you need a much stronger control net effect and get way worse quality. So what is the best upscaling method? It depends, I'd say, if you want an upscaled version that's as close as possible to the original and the images are not already generated but real photos, I think GigaPixel is the way to go. If the images are generated by diffusion or you don't care much about smiliarity, use SD and latent upscale and maybe GigaPixel on top. Though 4k or 8k upscaling is IMHO pointless with SD either way, at least when using a common 512px model.
@@testales Many thanks for taking the time type such a considered response. Half the battle with learning is knowing the correct terminology so my next project will be reviewing latent upscaling :P. Have you played around with ControlNet>Reference only yet?
@@Space-O-2001 Of course I did. :-) But I can't add much helpful information here fore this. It all takes long time to figure out what works best and in what situations for what purpose. For example I didn't have to time to get a feeling how multiple reference images work together. There are so many ways to guide the generation of an image, it's insane. For example I just recently saw what you can do with the loop script and that you can have parts of your prompt only have affect earlier stages of the generation. And of course I'm still in search of the perfect upscaling method and also one for videos that bascially recreates the frames in a better quality but keeps consistency.
@@testales I hear you, I also have played with the loop script. I appreciate it's a result of being early days and ever changing landscape but given the amount of permutations available during the creation process the UI could do with some 80:20 pre-sets based on what you are trying to achieve. That said it could be argued this is where the skill lies and it is, for the most part, faster than learning how to draw/paint/render yourself.
THAT'S FANTASTIC!!
Super Clean images 😁
"Look at all these details"
The details: *few humans growing straight from her hair strands*
Amazing work - thank you for the excellent descriptions!
Let's take a look at Control Net Tile Upscales that allow you to render 8192 by 8192 Ultra 8k Resolution. If you don't have control net yet check out this video here - github.com/lllyasviel/ControlNet-v1-1-nightly/blob/main/README.md Also, I just uploaded the 4x Ultra 4x Upscale. It seems I forgot to add the link when I first released the video. It is in the video description now.
👋
Very good results. Thank you so much for your tutorial!
Glad it was helpful!
Great video! Thank you for the guidance and knowledge. 🙏
bruh this can literally fix any crappy image i made even in 2021!! I have been in Ai since then!
😀
awesome job! I definitely subscribed --I look forward to more
Thanks, I got a lot more coming. Been playing with stable diffusion for a while and there is a lot of cool tools.
Followed this (admittedly in the newer version of A1111) but I dont get anything like this sadly
Thank you so much for this video. I have learned so much from you! I thought I was making ok photos, but I wasn't lol. But does make me wanna upgrade from my 3060 RTX to something better.
hm, under dontrolnet v1.1.234 I just don't have the tile under Model - lots of others you have are missing too - how to update that?
Excellent tutorial friend. Thank you very much
No problem! 😃
Thank you for this amazing tutorial!
Anytime !
very interesting presentation, good job dude
Great video. Thanks for sharing ❤
Glad you liked it!
awesome work! keep it up!
Appreciate it!
Subbed for the potc quote
On AMD GPU's this does now work.
Do you have a fix or a guide for it to make it work?
Wow!!
I tried this on a picture of the Martian landscape, but the picture came out with faint tile lines and lost a lot of detail. Not sure what I did wrong.
During the latent upscale or the tile upscale? It could be a diff model was used to create the image, custom models will add their own flare. Check out this site, stable-diffusion-art.com/controlnet-upscale/ the third method is tile upscale. This is go to site for all things stable diffusion
I never knew about that xyz plot script that is a great feature. Thanks.
I use it for everything. I have to know what the best setting is.
Great stuff as usual, I just wanted to point out that the Multidiffusion (Tiled Diffusion & VAE) is its own upscaler. I did not see you ever enabling it in the video. The ultimate SD upscaler script can work fine without it, aswell as Tiled Diffusion & VAE can work fine without ultimate sd upscaler. I noticed it due to an error i got (something about image does not match size) while trying to generate an image while having both of them enabled at the same time, which did not work.
I think I prefer Tiled Diffusion w/ Tiled VAE. Seems like it does a better job with not showing the seems that can appear with Ultimate SD upscale. I found an upscaling workflow from someone recently that I think does a good job. From initial 512x512 img to 1.4x HiRes fix (txt2img) or Img2Img resize by with low denoise strength like .15 to .3. Then drag that new img to img2img and again 1.4x, then again 1.4x, and once more 1.4x (now at 1967x1967), then Tiled Diffusion(or Ultimate SD upscale) 2x up to 3888x3888
I work on 8k pictures with inpainting for some time now, slows down the UI but it's okay. Rtx 3090.
Deng do you 24gb of vram? I should of forked up the extra cash. I got the 4070 ti but sitting with 12gb vram. Thanks for the knowledge. I always wondered if it was a stable diffusion 2048 limitation or a video card one.
interrogate clip ?, where i find that
This is cool. Imagine using this to upscale game assets nicee.
You making a game? I wanna see =)
Great vid, keep the good work
Perfect tutorial, thank you
i looked for a tutorial like this for months... Thank you so much. I only have 4gb Vram, but i trying to generate the best images i can... Thank you again!
Awesome glad I can help. This takes a while so just start it and play some ps5 or go to the gym
@@ArtificallyIntelligentAi about 10-20 minutes to get 2k resolution, im not trying 4k or 8k for now hahahaha
I leave it working and i go play with my son on the garden. He makes this time pass like a eye blink
i get lines on final image from tile after ultimate sd upscale, what to do?
Me too, did you figure it out?
You've really made progress. I need to catch up and learn how to make 8k AI art
Thanks!
Super helpful video. Thanks man!❤Have you noticed the three small characters on her hair?😄
my brain will explode
This is awesome 😮
This is fantastic. Nevertheless, I noticed a minor issue - I can see some tiles in the picture after upscaling. Is there a method to eliminate them?
This. Some images show the tiles
Learn how to paint. No tiling or artifacts!
@@skepticalpanda8862 What connection does your response have with the subject of upscaling?
Did you manage to get rid of that? I never got this before but it suddenly started to happen
@@willfreitas2832 If my memory serves me right, try reducing the denoising strength. The lower it is, the fewer grids should be visible.
The license of 4x-UltraSharp is for NON-commercial use, and attribution is legally mandatory whenever you use it.
Good to know. I'll look into it and put that in my video description.
i will definitely be trying to attempt this, i have been wondering how people get higher quality pictures then i was able to get
Thanks!!!
No probs!
I run it but the resolution only changed with the same pixelated imagen at a new resolution.
i dont have the window to uploais d my image on controlnet , is that important?
Why interrogate clip? Just send it to img2img using the shortcut in the txt2img tab
why does it continue to generate images within the image?
You never used multidiffusion in this video right?
Yes, I’m the beginning of the video I said I was only going to show how to install it so you can test it because I can’t get good results from it and I need more experience with it.
@@ArtificallyIntelligentAi Thanks yep same could not make it work either
so I have 16gb Vram, 32gb ram, AMD Ryzen 6900 XT and still getting GPU errors on even the basic Latent Upscale step... What am I doing wrong?
hmm ive got a question, how do you make the 'same' picture you've done at one platform to another platform?
previously im generating pictures over at stable diffusion web, but now im trying to do everything over at automatic1111 but i cant seem to get my 'style' over at the 1111 side. anyone can help me with this?
'style' being the face and stuff, and the models im getting arent that close to what i did over at stable diffusion web. so i cant seem to replicate the work im generating over at the stable diffusion web. cuz over at the stable diffusion web you dont have to pick any models to begin with, its just adding your prompts and adjusting the guidance scale/seed etc.
There is a way to erase the error of the tiny replicas around the image?
That is insane! Thanks for this! Can this be done for photo real landscape matte painting?
No problem, I’m sure the tile upscale can but u prob want to keep the de noise low to prevent the photo from changing its details. My original pic was so bad I had to up the Denoise to create details that weren’t there. But I’m sure it’ll work.
i dont understand why i get always bad results, like worst than the original pic, what looks like macroblocking, the image just looks blown up to 400% with no visible changes
primer comentario en español, me sirvió tu video, ahora puedo lograr mejores resultados, te lo agradezco mucho ✌✌
Utilicé el traductor de Google, por lo que algunas de las palabras traducidas pueden ser un poco extrañas, ¡pero me alegro de poder ayudar!
thank you awesomeee
No probs!
I don't have the esrgan folder, how can I install it?
your meme game is almost as strong as your SD knowledge
I love it, you get me. I got to hold back on some of the memes though, I find almost everything funny and meme worthy.
im just happy to go from 512 to 2k with this
i dont find the case "zone" for putting image in control net !!!
All of the tiles have differentiating edges, and it makes the whole image seems like a puzzle.
Main model used: A-Zovya RPG
ControlNet settings:
Preprocessor: tile_resample
Model: control_v11f1e_sd15_tile [a371b31b]
Any idea? Would you be able to help if I dropped more of my settings? Do you think I should change the main model?
A small follow-up: Yes the main model was the issue
Follow-up 2: Mess around with the Seams options under the 4x Ultra-Sharp Upscaler
@@sanyidupla8830 still has tiles showing up in the picture. Any other way around it?
@sovietchild
Yes!
Ditch Ultimate Upscaler, and also ControlNet for upscaling.
Under Scripts, run SD Upscale, put the first slider up to max (which if I remember was around 256), choose 4x Ultra, Esrgan, or whatever you like, and it should look amazing.
Tell me if it doesn't work and I'll tr6 explaining it while I'm in Stable Diffusion
you can remove every prompt that describes the character and only leave some unspecific quality prompts, that way you should not get these small characters in the image
I actually just prefer doing that now especially with inpainting. Pretty much going to recommend it moving forward.
Ultimate SD upscale doesn't do the work, because it makes tiles. On 8k it's very visible. Go to Photoshop, add extra curves layer and play with curves to extreme. You will see that girl is under the grid. The only method which works is Tiled Diffusion. But I have other problems with it. I have rtx 3090 so I can upscale to 4k without problems, but when I want to upscale to 8k I have to enable Tiled VAE (to avoid Cuda out of memory) and then I get not so good upscaled result, but at least it's not tiled. I still didn't work out, how to get the same good result as it was on 4k without Tiled VAE.
1:45 even that grannie picture is tiled.
I'm new to this and I was wondering if this only works with Nvidia graphics cards and its Cuda cores, I have an AMD card RX 6800 with 16GB VRAM and I want to try these things.
Hi there, so it does work on AMD but it's more of a workaround and they hasn't been very many updates. I don't have an AMD myself so I can't test it but there is the link for the install and it's just an additional GIt command it looks like github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs then here is a install tutorial on it ua-cam.com/video/8xR0vms0e0U/v-deo.html. I saw the full video and the only part thats diff is at 2 min to 2min and 30 into the video. It seems you need to create subfolders for the webui then paste in the command that is specifically for AMD. That's the only diff thing about the amd tutorial I see. There are also free sites like playgroundai, leonardoai, and clipdrop that let you create stuff through stable diffusion through their sites for free with limitations on how many you can create like 500 pics a day or something.
@@ArtificallyIntelligentAi oh thank you a lot mate. 🙏 i will try it out.
why didnt you only download 1 file.... it is only 63.9 Mb ????
I want to get into this do u have a video show how to start from step 0; get the app or program ur using n its extensions plug-ins and so on
Check out my video description I will put videos in order if your staring from scratch so it’ll be the stable diffusion automatic 1111 install, custom modes, control net, the canvas zoom extension and then this video
You don't have to put an picture in control net. Just in the img2img Canvas and control net will take it from there
Ive seen other people do it w/o the img2img so I def agree. I haven’t done too much testing without images in ctrl net so I didn’t want to. Once sdxl is fully compatible with control net and ultimate upscale I will prob update this video and test it out
does this work with anime? I've tried it a few times but my picture was too blurry, any more tips? its hard to find good info on upscaling or inpainting. there's a lot but not all of it is very useable or relevant and the anime models are kinda different
Have you tried Waifu xl ? It’s a free site and can do either 4 to 8x upscales.
Thanks for the video! Is there an extension you're using to zoom into those images to inspect pixels within auto1111? Or is it a different application?
I think he might just be opening the file from the folder and zooming in from there.
OMFG! Thank you for this. An absolute gamechanger. I've been using 4x_Foolhardy_Remarci in the extras, which is pretty good and fast but this is next tier in terms of fidelity. It's far more time consuming but the results are worth it, especially for print!!! Legend!
Great video gonna keep watching your videos :)
Btw the Interrogate clip seems to take a very long time on mine never seen it become finished, even after 10min how long should it take?
It will download a model the first time you run it. It's about 900MB, so it might take a while if you have a slow internet connection.
You can look in the console window and see if it's downloading.
But it may vary on internet connection speed, img, size and VRAM but it your trying to interrogate a clip that's 4k size you can wait up to 5 minutes on average. 10 minutes is a bit long but possible if your have an 8k img. If you have at least 8gb of vram, a 500mbs or greater connection and a img under 4k you shouldn't wait more than 5 min. A 1024 pic will take 30 seconds or less. The Site Hugging face also have CLIP interrogators that will do the same thing, quite a couple sites have it actually.
Im having an issue where the model for the any preprocessor is missing in the list. how can i fix that
I understand what we downloaded for and how we used the tile model for control net. But
Can you tell me what we installed this extension for 02:41?
When I did this I got some body horror on the level of SAW. The person looked so cut up and scratched. I literally got nervous from it. (The shock that is). I have no idea why that happened. I mean it was sharp just nothing like the original in skin damage.
So we installed that Tiled Diffusion / MultiDiffusion, but how are we actually using it? I don't remember seeing you check "Enable Tiled Diffusion"
Thank you sir! Will try it out later. I tried upscaling with the 512 tiles once, but that gave me strange color variations in the final image (basically, you saw the tiles). I will give it another shot with the tiles and denoise stuff once I'm home
I got that too one time , I forgot what I did but it was a bunch of tiles as well with some little people on each tile. I think it was when I upscaled with the preprocessed not matching the tile model
@@ArtificallyIntelligentAi I tried it yesterday and I can say: everything worked out pretty well. I did my very first 4k x 4k upscale and I'm impressed by the details, no boxy tiles or something. You changed my workflow with your video, thanks!
Mine still has boxy tiles, how can I fix that?@@neurosis404
what is the difference between the options "Balanced" and "ControlNet more important"?
look at the quality of this 13:23 🌚
crazy results, thanks!
I"ve been comparing results with gigapixel and the SD version is much cleaner on textures lightning and even the hair,
I tried a lot of setting on gigapixel but they always end up blurry.
is there any way to save the settings as a template?
I'm facing an issue each tile it render deferent variation, some time it show distortion, extra faces and hands. please help.
I had this one time and I think it was when the pre processor didn’t match the correct tile model for control net. If it’a each tile that’s prob the issue if it’s the img has a whole lower denoise and ensure downscaling isn’t too high. It might allow u to use less vram but it lowers the quality of the image so it can recreate new details.
Whats the point of installing VAE/Tile Upscaler when you just use Controlnet
Waay to much memes. But the tool walkthrough is clean and easy to follow. Thanks.
I'm working on my meme addiction. Your not the only that said this, but expect less memes.
I got a secret, it's not bad trick - if you blow the image up to 4000 by 4,000 and you like it but do you want to inpaint. Resize it back down to dimensions that your computer can handle, so usually I'll put the back down to 1000 x 1000 or 512 by 512. Do the inpainting then blow the image back up to 4000 by 4000 or whichever resolutions that you were at. I only using 4,000 by 4,000 as an example.
I prob got to do that for this image to get rid of the fairies 🧚♀️ thanks for the tip
Why not to just cut out the part you want to change, send it to inpaint, and then paste the inpainted part back?
Well... Unless you want to inpaint a half of the picture, lol.
But I think if you've let it go this far, without taking any action, it's your fault ¯\_(ツ)_/¯
Sir, Please give link to Civitai negative prompts, thank you. I'm doing everything exactly like you did
This is the deliberate negative embedding that I was using civitai.com/models/30224/negative-embedding-for-deliberate I would test it with that xyz plot script I used to find the perfect denoise before upscaling
how? when I use controlnet upscaling just like you did in the video anything that Im upscaling looks worse than before.... whys that? please tell me I m trying to get higher resolutions for months and nothing works.
My denoise was set to high so lower that first, hence the little people floating in my pic but it shouldn’t just be worse quality , make sure your upscaling with the same model you created the img with. I made that mistake before. Also if your using sdxl I’m not 100% sure if it’s compatible right now. I will have to test it when I get a pc again
I've never been able to get tiles to upscale the full iamge for me, it takes an hour or two to upscale an 800x600 image by x2 and then only upscales about a quarter of the image the rest is just the old image resized. Hope some of this will help but most of it i've already tried.
Hey there! Great video. Super easy to follow....you did a good job. Is it possible to run this with a 1080 Ti? I keep running out of ram even after following your suggestions. Thanks!
You need VRAM, a ton. Best would be 12+
i hv no control net session, anyone knows why?
1. why did you also install pkuliyi's multidiffusion? it was not used at all in the video?
2. what does controlnet actually do when combined with ultimate SD upscale? I have been using ultimate SD upscale for a long time just by itself.
1. It was so you can test it out and see what results you get. I did mention it but it was brief. It was also recommended by most people but I didn’t get any results from it.
2. Provides the tile upscale when you go above and beyond your cards limits to process it in smaller chunks rather than giving u an out of memory error. So if you were running a 3090 you might not even noticed since it has 24gb of vram. Depends on your card and how large the upscale is though
Problem for me, error: gather(): Expected dtype int64 for index
Time taken: 3.92s | GPU active 2741 MB reserved 3030 MB | System peak 4325 MB total 16384 MB
This use to work fine, but recently perhaps after some upgrades I am getting this error when I want to use to controlnet, any one else is getting it " AttributeError: 'ControlNet' object has no attribute 'label_emb'"
Ya SDXL wasn't fully compatible with Control Net. Hopefully I can replace the video soon
Once those powerful price manipulators do their thing all of a sudden TA doesn't work anymore.