ULTRA SHARP Upscale! - Don't miss this Method!!! / A1111 - NEW Model
Vložit
- čas přidán 10. 07. 2024
- This Ultra Sharp Upscale Method will blow your MIND! In 60 Seconds or less, I show you how to get the Results. Then we take a deep Dive into A1111, Img2Img, different Sampling Methods and Upscaling Strategies. This deep Insight will give you the best results with the lowest render Time
#### Links from the Video ####
Join my Discord: / discord
Join my Facebook: / theairevolution
Upscale Database: upscale.wiki/wiki/Model_Database
4x-UltraSharp.pth Download: mega.nz/folder/qZRBmaIY#nIG8K...
Support my Channel:
/ @oliviosarikas
Subscribe to my Newsletter for FREE: My Newsletter: oliviotutorials.podia.com/new...
How to get started with Midjourney: • Midjourney AI - FIRST ...
Midjourney Settings explained: • Midjourney Settings Ex...
Best Midjourney Resources: • 😍 Midjourney BEST Reso...
Make better Midjourney Prompts: • Make BETTER Prompts - ...
My Affinity Photo Creative Packs: gumroad.com/sarikasat
My Patreon Page: / sarikas
All my Social Media Accounts: linktr.ee/oliviotutorials - Jak na to + styl
#### Links from the Video ####
Join my Discord: discord.gg/XKAk7GUzAW
Join my Facebook: facebook.com/groups/theairevolution
Upscale Database: upscale.wiki/wiki/Model_Database
4x-UltraSharp.pth Download: mega.nz/folder/qZRBmaIY#nIG8KyWFcGNTuMX_XNbJ_g
👋
Why when I set the Denoising Strength to 0.3 the image that comes is blurred? Is 0.5 the minimum? Because I don’t want to change the image so much, especially the face, usually my face 😂
I can see the upscaler in extras but not in img to img. Any idea why?
Came for the AI stayed for the fire raps
Sensational!
Highlight of the video
Same 😂❤
He 110% asked chatGPT to generate that for him XD
As a recent ai explorer I have to say your channel is just fantastic, east to follow instructions and a wealth of technical data. Keep up the great work!
One of the most important videos about SD I've came across so far, this tip is truly game changer if you want to not only improve the quality of upscaled images but also finetune details that weren't as prominent (like different eye colors, for example) in the initial render
Love yours videos!
Olivio - thanks to you my images have seen a ten-fold improvement in both quality and artistic merit. I'm addicted to your channel!
Olivio thank you for your easy to understand turorials so far. You are doing an amazing job!
Thank you very much
Amazing results, excellent tutorial, thanks for sharing 😊
I'm from Brazil and your videos are too amazing. Your rap's flow is amazing too, awesome job.
Thank you so much for sharing this method! It works beautifully.
Thanks a bunch for sharing those upscaling methods! 🙏
Very grateful for this 🙏
This is why I follow you, thanks Olivio!
That looks great!
As always superb content Olivio... thanks for all you do. Also, the I am enjoying the raps / rhymes. Keep them coming! :)
Thank you, my pleasure :)
Wow! Thanks for this guide, it is most appreciated! :)
dude got bars, nice vid man !
thank you so much for the 1 minute tutorial in the beginning, that was all I needed :D
that is so cool, thanks for sharing this!
Lol love it Olivio - the rap, the video.
Thank you for the upload
Thank you 😍
Thank you! This helps my workflow so much!
As always , super insightful
Glad you enjoyed it
Thanks for another great video, this is really helpful to create more quality images, Thanks Olivio.
Olivio; great tutorial and information as usual; i have been using this post processing on my images now to great effect; thanks!
This is absolutely amazing. Thank you very much for your videos and information for the public.
You could keep this to yourself but you didnt. Thanks a bunch!
Now images are way better. Was actually checking out how to not make these low blurry details on clothing.
Your method really fixed this!
Thanks for you videos!
Great stuff Olivio!
subscribed !
thanks for the tutorial. its so easy to follow.
and i love ur rap bro
Thanks for the suggestion, I just tried it with Ultimate SD upscale, and in the case of Euler A, it pretty much worked like black magic, as it can clean up most of the depth of field in the distance and bring out all the details and textures in a convincing way.
You are a legend for all of these guides. Thanks dude.
Mypleasure, thank you :)
Thank you for your videos you've been a big help to me.
Thanks for this video, was really helpful!
I love that you tell the technique super quickly at the beginning of the video, then tell us more about the exact process. It's very efficient!
I also love the 🔥🔥🔥 raps. You get more stunts and props than Bruce Willis.
Wow that is amazing, you are the best. Thanx
Wow that really slows down my render but it is very impressive! Thanks for posting this.
Awesome video! Thank you so much!
Olivio got bars, lol. Nice job my friend. :)
This is incredible! Your channel is the holy grail of automatic 1111, bless you Olivio! I have been creating pictures since the very early stable diffusion days and come from shoddy bad (but cool) results to actual art pieces and I totally grew along with you and your channel. Today was the first time I actually wanted to go deep in how to get the best upscaling and voila there is the new video. I hope I can have a future in this field. I have never had more focus and fun while working towards a goal that is not purely pleasure.
Beste Grüße aus Deutschland :)
P.S.: That rap came out way too impressive! Did you come up with that or did gpt help? :D
Thank you very much, that means a lot :)
@@DioHard All my own lines. I'm pretty good with simples rhymes :) I also like the more complex ones that break words into different parts like: Headphones to mad tones. Or Mansion to hench men to pension 😋
Excellent video! I would love to see the Ultimate SD Upscale extension explained too😊
@@OlivioSarikas why when I set the Denoising Strength to 0.3 the image that comes is blurred? Is 0.5 the minimum? Because I don’t want to change the image so much, especially the face, usually my face 😂
this is great, thank you!
you are a legend! watched alot of videos of you and you got me curious about AI art
Amazing Video, literally the best upscalee
I have to say this is THE "industry standard" channel for AI render 😀If there is an update or new thing, THIS is THE channel to see! Keep it up!
Thank you very much. Always trying to bring you the best
It was great! Thank you!
You're the best, Olivio! 😄
Thanks for the Info!😊
I really love this tutorial!
Thank you, amazing video❤❤
You're welcome 😊
Best instructor!!!
Thanks! And the rap alone deserves the thumbs up!
Very helpful and thank you!
Amazing. Thanks for sharing :)
Love the quick summary of the method at the beginning of the video and the explanation afterwards, way too many people on youtube try draggig out the information from their video for as long as possible.
Amazing! Thank you!!!
Thanks for this method. It really helps!
You're welcome!
@@OlivioSarikas !!!
Thank you Olivio!
wow awesome thanks for this tip helps me a loot
Excellent this is really helpful
thanks so much this really helped me to figure it out! though you could just do 4x right away but you need to do the first step then the 2nd for better resoults
Love your rapping 😂! More please.
That's fantastic! I may have to re-render some of my favorites so I can get some prints done for my wall.
Great idea, that sounds awesome :)
Olivio! Did not miss a beat, taking all the heat, leaving everyone dead on their feet!
Best video ive seen on the Ai scene! 😂
Thank you so much!
thank you for this tip!
love the side by side comparison,
straight forward and 3 years old me can understand how to play with this.
*have a LOT of images that would greatly benefit from this advice, thank you*
Olivio.... you the MAN! 🫶
very nice tip! thanks
im not using anyother scaler, this one awesome, thank you!!!!
this might be the video of all times! My results improved drastically. And a massive timesaver!
O.....M....O (Oh My Olivio!!!) 😍🙌 subbed and part of discord.....
PS. actually doing it with just uploading previously completed images to give them more detail. Had been using GigaPixel to do so......
I've been using the Img2Img method to get to 2x already. I even use it for non-SD images I want to upscale. I haven't played around with the extras tab to get it higher, thanks for that!
magic~thank you!!!
Thank you very much
You can usually get better end results by sending the img2img result back to img2img and redoing the generation. You can repeat this adjusting the denoising as appropriate to narrow in on your preferred result. Then send it off to extras upscaling when happy.
Also, if your computer/vram can handle it, use xformers (to minimize vram usage) and use latent upscaling as large as you can in img2img. Then for the final upscale, while still in img2img, select the "SD Upscale" script (in the script section at the bottom), set the scale factor and upscaler as desired, make sure denoising is low (around .3) and generate as usual. This produces the best results I have seen.
This just changed my workflow ❤❤❤❤❤❤
awesome, happy to help 😀
I've been using a similar method to this for awhile now, and can often get to 8k without the need for inpainting. I start with resolutions of 704 and below(though 768 vertical can be reliable with closer portraits). I prefer the Latent(nearest exact) and Latent(bicubic antialiased) for Hires fix.They can operate at slightly lower denoising strengths than the standard Latent upscaler, 0.4-0.45 respectively. So, you have a lower chance of unwanted distortions appearing in your image. Using ControlNet, can help with this too, provided Guidance is ended early so full detail can be achieved. Next I use SD Upscale to double again, to get a little more detail than the Extras tab can offer. If you're dealing with an image containing a figure, you want symmetrical changes. The easiest way to achieve this is by using full final width tiles. Then, you want to minimize the total number of tiles, so take the overlap into account when selecting a vertical tile resolution, and select the maximum your GPU can handle. I do the second pass upscale with 4x-UltraSharp, and 0.2-0.3 denoising strenth. I usually bump up the CFG by 1.5-3 points with each SD Upscale pass. Using DDIM in the second upscale pass can also help with detail, but it also adds noise, so I like Euler a or a DPM for subsequent passes to smooth that out. Next, attempt to double again, with a denoising strength of 0.1-0.15, reduce to 1.5x if distortions appear.
Also there's a new Loopback Scaler script on CivitAI, which may be useful for a step in between Hires fix, and SD upscale, to go even bigger with more detail, if you have a GPU with 24 GB VRAM.
Pls make a howto video ;-)
In what way do you use ControlNet to upscale?
sounds too complicated
This is the ONLY way to add detail, rather than using a "gaussian blur sharpening" which the Extras Tab uses. Similar to Photoshop.
Also, in later videos in this channel using ComfyUI, Olivio actually does this using Latent Upscale nodes.
@olivio - you should take this video down... OR point to the correct ComfyUI video for upscaling.
Works great!
Thank you Olivio! I'm anxious to try this new Upscaler. I've recently swapped my workflow from Hi-Rez fix over to Img2Img / Extras to save time as you suggested. I've been using the stock ESRGAN model (vs the 4x which I've found flattens and mutes the colours too much). I like the skin texture, tone and details shown in this model.
I wish they would add a "late hires-fix" so we can decide on it once we see the image and like it.
@@OlivioSarikas I usually generate a batch of 5-10 images and then choose the seed I liked best and generate it again with Hires-fix
11/10 THAT RHYME IS FIRE!
I love you Oli!
Omg 😂 those bars cracked me up!
MUy bien Olivio!!! sigue adelante!! saludame.. DJ HUNTER
Thank you
This is similar to my latest method. Also a 2 step upscale, but I use an upscaler (universal sharp) instead of latents for hires fix (keep this below x2), then send to extras, and now use a combo (also keep below 2x) (I use swinr or swin2r, and universal sharp at 50%, but you can play with ratio as it is quick), then finally another pass into img2img at 0.1 denoise, cfg/2. This works great for photo realsim of people.
Thank you, I will give it a try for sure. If you are on discord, message me to go deeper into that idea :)
THANKS AGAIN AND ALWAYS for your videos. How do you go about when you want to perform "only masked" in-painting after upscaling ? Do you decrease the size and do it or keep the same size set when you sent it from extra ?
I subbed because of the rap. LMAO, it was good.
0:59 I wasn't expecting you to be spitting straight fire I'm shook
i was looking for upscaler that adds such grainy effect to the output.Thanks a lot,Olivio! It would be nice if you explore another upscalers too. as far as i know there are some that add metalic feel to the metal texture its called 4x_Nickelfront, also there is one for fabric texture excitement called 4x-Fabric and for realistic skin texture x1_ITF_SkinDiffDetail_Lite_v1.
Thank you, i put them in my notes and will check them out asap. Feel free to post examples in my Discord or Facebook Group ;)
I’d like to see that too !
@@Vigilence the link is in the description
Olivio amaze me in every video, i dunno how but he spot all my settings precisely. 😊
This time also the 0.5 denoise spot, wich i use often ❤I guess all this micro settings and adjustment are those things you find only when using stable diffusion for so long so tryal and errors talks itself to improve the workflow. 😊
testing for hours. that's why a 10 minute video can take half a day to create ;)
I just found out that if you send an upscaled image back to img2img you can go beyond the usual 2048 limit. So you can do a latent upscale of any size, as long as you have the vram. If you use xformers to minimize vram usage, you can get around 3000X3000 with a 24GB card. You will need to lower denoising to about .3 though, or you will get strange mutations since Stable Diffusion is not designed for that resolution.
So the best results will be attained by ramping up the resolution and lowering denoising step-by-step in img2img, then using the SD Upscaling script to get your final resolution, then send the upscaled image back to img2img and generate the final with denoising around .3. This produced the best results for me.
To keep small details, I recommend the Siax model (it's on the wiki page too).
I only use his one for realistic style of images like real faces
Sounds good, i will give it a try
came for the upscaling tips, stayed for the hard bars
Thanks for making this guide! I've had a few great results using this technique but I've noticed that maybe 60% of the time, my latent upscaled image is really blurry and after some googling it appears other people have had similar issues with Latent upscaling. I was wondering if you'd encountered that in your experimentation and if you'd discovered a cause and/or solution. Thanks again!
Thanks!
Thank you very much
I didn't knew we could add the pytorch models from wiki database. I use a lot with chainner and now i will copy them into stable diffusion :)
Great!!!!!
under img2img, don't forget you can load the SD upscale from the scripts and use it to set the scale to x2
Use the extension to lock the aspect ratio, helps a lot. Is one of the firsts available on extension tab.
My computer only allows me use the 1st method, but it's amazing! THANK YOU SOOOO MUCH!
You're welcome!
okay spittin some barz at the beginning i like that^^
word 😅
Great video (as always, great explainations)
Is there a colab link (with my AMD, I don't think NodShark can use this)
awesome rap!
Thank you :)
Can you compare this method vs. the other gold standard upscaler: Ultimate SD upscale with LDSR.