Stable Diffusion resize to 24K?? Easy method + vram tips to get upscaling with Tiled VAE extension
Vložit
- čas přidán 5. 06. 2023
- In this video I'm going to use Multidiffusion extension and UltraSharp upscaler for a very straightforward way to push high quality and high resolution in your images. I'll take an image from 512 x 768 to 16,384 x 24,576 and also provide some memory management tips to avoid out of memory errors when upscaling.
Items discussed in this video:
4x-UltraSharp upscaler:
huggingface.co/lokCX/4x-Ultra...
Multidiffusion on Github (Available as an Automatic1111 extension, but the documentation here is worth reading)
github.com/pkuliyi2015/multid... - Věda a technologie
Totally Underrated workflow for upscaling. Awesome video, has been working for me very well. Sometimes I have trouble choosing the right denoising strenght, but overall its a perfect preliminary way to upscale. I wish I could know more about the tiling (I'm not programming savvy), so that I can tweak and overclock it. Many would be greatful if you do a long form of upscaling techniques, and elaborate on the details in the settings - I'm sure the community would greatly appreciate it. Either way what you have brought is priceless
Thank you for the kind words and for watching!
Thankyou, been looking for how to upscale with tilled diffusion.
ps: if you're using control net with tile you can leave it empty since the control net will use your img2img image as an input
It's the only upscale method that works for me. For some reason without Tiled Diffusion and Tile VAE the image won't go larger than 4k. It processes without error, but the size remains. Your method works.
In your first upscale (the hires.fix one) you used 4x ultra sharp. Wouldn't it be better to use Latent for this first upscale? and then use the 4x ultra sharp in the img2img part?
Thanks for the tip, I'll give that a try.
Awesome channel, keep up the great work!!!
Thanks so much!
@@renaissancelaboratories5645hey buddy ..sorry for bothering you again but please I really need your help
Please please please 🙏 😢 😭
Why don't you use the Extra Scale directly, what is the point of going through Img2Img first? Thank you for this video!
Hi. I don't see the point, why you are using Controlnet here. When you activate Tiled VAE, you are losing details (I have rtx 3090, so I activate Tiled VAE only when need to upscale to 8k or more). Controlnet is for adding extra details (as creator of Tiled Diffusion says). But you put denoise in the lowest level, so there will be no extra details. So what is the point activate Controlnet ? And all in all I don't get the tile use of Controlnet as you already do that with Tiled VAE. If you want extra details, just add a little bit denoise and more sampling steps and you even don't need Controlnet. Controlnet just adds extra rendering time. Also 4k Ultrasharp adds some contrast, when sometimes you actually would like to avoid that. So I've noticed, that less contrast is being added when you uncheck Fast Encoder and Fast Decoder in Tiled VAE. It will add 1+ minute extra rendering time, but somehow I got less contrast. Don't know how it's connected, weird of course, but the result is better.
This is good info, I'll keep experimenting with different settings. Combing Tiled VAE with ControlNet is a suggestion in the Tiled VAE readme: github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/blob/main/README.md
@@renaissancelaboratories5645 Yes, so that's what I wrote in my previous comment: use Controlnet if you want to add details. But you totally killed denoise level, so what is the point to use Controlnet if you don't add any details ? You just increasing rendering time for nothing. But as I also wrote: you can add all those details without Controlnet just adding sampling steps and increasing denoise level.
Yeah I was kinda confused with the .01 denoise strength. Seems like not much detail would be added. I ran through this process following the video and also ran the txt2img hires fix photo straight to extra tab to upscale to 4x w/ ultrasharp and ended up with a very similar image compared to running through img2img w/ Tiled Diffusion, VAE, and ControlNet. Seems like you'd want a higher denoise strength to let tiled diffusion and controlnet spread their wings.
@@Kryptonic83hey I am absolutely newbie in this field ....I have a problem can anyone please help me with it It's really serious
In order to increase the sped just as he did and being able to use such features of upscale...am I supposed to increase my VRAM? Will increase RAM also affect the speed
Thk. Going to try to push my 3060 12gb to do 12k.
Please let me know how it goes. I got an out of memory error right away with a RTX 4080 using the default encoding tile size of 3072. As long as you use lower values you should be fine.
@@renaissancelaboratories5645hey brother I am in a big problem can you please help me
With rtx4090 I was able to set Encoder tile size: 2400 (base was 512x512)
I'm using different upscaler: 4x_NMKD-Superscale-SP_178000_G
How does this fair as compared to Ultimate SD uscaler?
should make a video on forgeui explaining what tf everything means lmao especially hypertile integrated
Thank you for the tutorial. Please would you mind to share the link for original tutorial on reddit?
Hello, Here's the Reddit thread and also some tips from the developer site:
www.reddit.com/r/StableDiffusion/comments/13w817d/kaneda_motors_new_superbike_model_workflow_in/
github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111
@@renaissancelaboratories5645 thank you so much!🫶
So for SDXL you need different controlnet models
Interesting workflow. An other Tutorial I watched by "Artificially Intelligent" used the Ultimate SD upschaler script in img2img to get to 8K
Using that method I get artifacting and new characters being generated in the 512x512 tiles.
That doesn't seem to be a problem here. ( denoising is low in both methods. )
Following this guide
I had to update the image in Controlnet , because I got an "ValueError: images do not match" error. otherwise. it wouldn't continue.
Thank you for the guide.
Glad you found it useful! Thanks for watching.
Do I have to change the image in contrlnet after upscaling, or can I only change the main image for upscaling?
controlnet in newer a1111 version, will put input image automatically into controlnet input. meaning you don't need to put it into controlnet anymore at all. just enable controlnet and set the methods (processor/model).
Hey! Does it work on Macbook M1?
when you are in img2img you do not need to place anything into controlnet, it grabs the image from above automatically
Good tip, thank you!
Where did you get ESRGAN_4x? I can't find it anywhere😭
This is interesting. I installed Automatic1111 from scratch on a different machine and I have ESRGAN_4x (and a few others) in the upscaler list. However, I searched the folder structure and I don't see the PTH files themselves. I'll keep digging.
Did you know you can change Automatic1111 to output .jpg files without any noticeable loss in quality but much reduced file sizes.
how?
@@ga5518 In Settings tab, the Saving images/grids sub-tab on the left. Second line where it says file format for images, just change png to jpg.
@@copenormcbrokelegsproperra2909 but doesn't png also save prompt inputs? Is this why its reduced? Like can I still use PNG helper for JPEGs reduced from png in the way you describe?
can you tell me pls. what lora is this
Hello, the is called 东方巨龙 Oriental giant dragon
Link to CivitAI:
civitai.com/models/81125
I had 8gb ram , but today i increased it to 32 in the hope that i will be able to generate inage in high resolution and use deforuma because at big size my laptop do not genrate the image . But after that things have got even worse if i try to make a image of big size even of 1024×768 my laptop is straight up showing that blue resstaring screen with that sad pc image..what should i do..what are the actual factors that increase my speed and fix this problem I will sound like an absolute idiot...but was I supposed to increase my VRAM instead of RAM
Indeed you need VRAM
Try setting low vram
These AI tools are heavily dependent on VRAM. There is a startup switch in Stable Diffusion for lower VRAM cards, --lowvram and you can experiment with other parameters here:
github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings
Feed your ceiling bird.
going to keep it 100 that doesn't look clean. I mean the work is tight but zooming in on your own work it doesn't look clean.
I appreciate the honesty and I agree. I'm still learning and experimenting. I recently used almost the same process to upscale another image that I had printed 55" x 33" on a canvas and it came out great.
You only need 9k for printing 24x36 poster guy.
Agreed. This was more of an experiment to see how far I could push the upscale using this process. Thanks for watching!
get a pop filter
Thank you for the feedback, I'll take a look!