Multi Diffusion for A1111 - Super Large + LOW Vram Upscaling
Вставка
- Опубліковано 8 чер 2024
- Multi Diffusion Tiled Diffusion Upscaler with a tiled VAE. This Low Vram Upscaler can give you 8x upscales with super high quality in A1111. And it works with Lightning models for fast upscaling
#### Links from my Video ####
github.com/pkuliyi2015/multid...
#### Join and Support me ####
Buy me a Coffee: www.buymeacoffee.com/oliviotu...
Join my Facebook Group: / theairevolution
Joint my Discord Group: / discord
AI Newsletter: oliviotutorials.podia.com/new...
Support me on Patreon: / sarikas - Навчання та стиль
Leonhard Euler was a Swiss mathematician (1707 - 1778). Pronounced oil - er. I know, not important, but it's good to know.
Respect
@@sorijin Did I say something without respect? Euler deserves respect. One of the most famous mathematicians ever.
@michaelballance1893 No, my guy, he was merely acknowledging what you have to say. :)
His ANCESTRAL is amazing!
This has been mentioned before....
Olive does his own thing.😂
"women, that are very gifted" :D
_men of culture... ?_
Thank you for always giving fun and informative lectures!🥰
I can't recommend it enough for SD1.5 models, I usually upscale initial 512x768 image 2x twice with it, first with denoising 0.45~0.6 and then about 0.3.
But for SDXL, ultimate sd upscale with CFG about 3.0 and denoising strenght 0.35~0.45 usually performs better in terms of quality.
Thanks for the video. Next time could you show the original image compared to the result? Without seeing the original full size, we can't see how good the result is.
MultiDiffusion upscaling has been around for nearly 1.5 years.
Thought I was losing my mind. Glad I wasn't the only one
Och wie geil, meine seit Monaten verzweifelte Suche nach LCM Lightning o.ä. Upscaler zu finden endet mit Deinem Video. TOP DANKE!
Hello MR> Olivio your channel is the best AI channel bro ..thanks for all !!! ... cheers !
I’ll try it thanks again!
You're gifted too, Olivio. Thanks 👍
Agreed
Pause😅
Hi, interesting as always, I wasn't able to find the noise inversion, at first :)
Thanks Olivio for another awesome tutorial!. Is there a way to fix the eyes artifacts or "reflections"?
Olivio, could you do a video on Layer Diffusion, please? Generating images with transparent background?
Kindly make a video on SUPIR 12gb VRAM version installation for SD1.5 and SDXL plz 😊
Hello and thanks for your videos, I was wondering if exist an upscaler not only for the resolution, at my eyes often the result is so poor realistic, but a way to get a realistic skin, with some details... Thanks!
Cool!
When I do this, the details of the skin are lost. Like small moles and natural skin texture.
Cops gonna love this (The upscaler).
@Olivio Sarikas SD WebUI Forge as Multi Diffusion integrated to it but this seems a bit different to that the only options the Forge version has are method 1: Multi diffusion And Method 2:Mixture of Diffusers then Tile Height,Tile Weight,Tile Overlap and Tile Batch Size those are the only options in Forge Version it confuses me
Fine. Please make a video about SUPIR, the best upscaler with ComfyUI, In order to understand what changes can be made to the default settings. Thank you!
Hi, do you have for comfui? also do you have only video upscale?
so nice....although not there yet with skin,... as MJ in most cases gets there. would love to challenge you on the face skin one day, if you are up to it? Thanks for lettings us see all your tricks and magic.
When using sd15 use tile controlnet too
I can see visible seams upscaled images, but it is really fast :O
img to img, but only for images created there? If so, what is the best offline upscaler? For Comfy, preferably.
SDXL-Lightning method is not working on AMD (ROCm on Linux). But the SD1.5 version works fine. Thanks for the video.
interesting, but what is the difference between this method and SD Ultimate Upscaler?
In my upscaled version I can see the tile lines. What adjustments do I need to make?
Does it exist for ComfyUI?
interesting
It is weird, when I use Juggernaut_RunDiffusionPhoto2_Lightning_4Steps using your example settings the output is soft and over saturated and is clearly not working correctly. But when I use an SD 1.5 model or the regular (non lightning) Juggernaut-XL_v9_RunDiffusionPhoto_v2 model it works fine. Olivio are you 100% sure your upscale result you showed came from the exact settings you showed in the video when you used the Lightning model?
I'm getting the same result, changing the cfg doesn't help either...
everytime i run the SDXLX model in the text2img, the result is a noisy image. so you can still make out the face / portrait, but it has a ton of noise on it. any help or suggestions with this?
Hi, can we find similar things for confyUI?
Is it available for Comfyui as well?
I don't think so, I installed and it's not showing.
@@BrunoMartinho hopefully it will be in the future.
I doubt, it's actually very old extension. Iv been using it for close to a year. I'm surprised people don't know about it.
I installed this but all I see is tile size under Multidiffusion integrated under img2img....weird. I don't have the noise inversion or tile diffusion tabs at all...
Is it better than Coyote A
Is there some working Discord invite link? The one in video description has expired.
Does this Upscaler work with Forge
Adding a comment to boost the algorithm!😂
Hi Olivio! Is there any reason others use Multi Diffusion in combination with ControlNet tile for upscaling? Why are you omitting ControlNet tile? I would appreciate any insights. Thank you!
You don't controlnet tile for it get great result. And you use multi diffusion because it's fast, other option is super slow and not necessarily better.
@@wykydytronthanks! So what is the ControlNet for then? why does it work without?
I tried installing this extension in forge and it refuses to enable, what am I doing wrong? is it bugged? broken?
I was hopeful on this one, but unfortunately my results are too airbrushed looking. Ultimate SD Upscale is far superior as you get more realistic details.
If I use my own pictures and workflow/settings shown in this video, the pictures sort of come out blurry, so you have this highres 4000 x 3000 image which when you zoom in just a bit is, sort of misty/blurring, not sharp at all like the result in this video. That's using the 4xUltrasharp in the Tiled Diffusion tab. I've tried changing the settings and the results just get worse. How could the workflow be adjusted to address the blurriness of the picture? And if you follow this workflow with anything except a head, the output is again blurry.
i am also having these issues, all the details from my initial picture gets washed out and when i zoom in the entire picture is covered in tiny lots that look like kinda like pores
I'm using ForgeUI and I cannot see the TiledVAE and TiledDiffusion, it is definately installed. Obviously I restarted it as well. Any guesses?
It doesn't work with forge since forge has a built in version.
I tried this and I got me slightly worst results than just using highres fix, maybe this is intended for low vram but this isnt better than the standard highresfix in A1111
@ 7:27, you can see the tiling. Halfway up the cheeks, the skin texture changes.
didnt notice that at first, but thanks for pointing that out, is there a way to reduce this?
Is that totally free like upscayl?
Does this upscaler run in Forge UI? Because I have only 4 Gb of VRAM and I'd like to know that I can use a powerful upscaler.
from what i saw forge ui its different completely but on the vanilla ui , i am able to follow along
Nope, it doesnt, please write a request in the issue section to get them noted haha, the developer of Tile diffusion are having an exam in China atm
The project's been around for a long time, but the native Readme is quite bad... this is the first time I'm seeing set CFG to 1.
I love gifted women
How is this better than the Ultimate SD Upscale script that's been around for ages? Is it faster? Or just more control?
It also has been out for ages, Olivio just missed it. It's faster for me, gives better results and you have much more control.
Has anyone here managed to get it installed for Forge?
Is this different than just using ultimate SD upscale? Seems like the same to me...
Yes. It does a better job of staying consistent across tile borders at high denoise.
It also has other features such as regional prompting that Olivio didn't get into this video.
It's like that but gives better results, is faster, and gives more control. I'm surprised people here about it just now it's quiet old extension and I can't imagine using any other method after I started with it.
Ultimate SD Upscale is better for photo realism.
🙏👍👌
OK, ok - but can it show us some molecules at the surface?
looks great but you can see a tiling line 7:12
Correct!
Is this same as the multidiffusion tab in forge?
yes
@@TheP3NGU1N Not quit, Multi diffusion does not have Noise inversion, which can be use to generate better result
you may want to check again XD@@CharlesLijt
@@TheP3NGU1N i used them every single day.... you mean the forge author zhang add that option today?
@@CharlesLijt - You are right. There's no option for noise inversion. The author, lllyasviel, addressed this directly and said parts of the extension were implemented directly into forge and that he couldn't figure out how to implement the noise inversion and some other options that he left out.
Hmm, what brought this on?
Tiled Diffusion has been around for over a year already.
He must have missed it.
👋
Not working with 4gb vram. Not enough gpu memory...
reduce tile size
I Tried to upscale X2 using a SDXL model but got Vram memory error (12gb vram, need at least 18-20gb) So by 1.5 or less would work if you lack for a 40XX card or more than 12gb vram or enable Tiled Vae
Change tile size to smaller one, if you don't fiddle with it even my 4080 can get out of memory.
you are a major gift yourself!!!
Well, without SOURCE image I can assume it was upscaled with MS Paint 😛
I tried this, and all i got was less details in the upscaled images, or visible outlining of the tiles, or sometimes even artifacts in the images. And i followed the steps in the video. For me, this wasn't really that great unfortunately...
Receiving a nans error.
I don't know if it's just me, but the end result looks too smoothened out.
Don't know if the original photo actually had any detail to begin with.
Result will depend on upscaler too, you have to test it as some will produce too smooth or too rough results with certain models or styles.
A1111 - Yeah!!!!
A1111 for Google Colab is Dead.... Do You know another SD project (A project using WebUI) for Google Colab ?
Fooocus
@@onurerdagli A project using WebUI
I am willing to give you a like, although you didn't showed the gifted women
This extension is outdated.. like come on.
Is there a confy UI version?
sadly doesn't work on forge
It comes with Forge as default, you don't need to install anything.
@@Elwaves2925 Nah, The Multidiffusion does NOT have Noise inversion like this Tile diffusion do
@@CharlesLijt Fair enough, I've never used it but remember seeing it installed from the start. As Forge and A1111 both work on the same base, it was fair to go with them being the same. I guess not.
@@Elwaves2925 Forge is on dev branch ver 1.8.0. TiledDiffusion (the one that has noise inversion and not the preinstalled one on forge) is currently only programmed to work on A1111 release version 1.7.0 This is why it doesn't work acording to the developers of the extension and forge devs.
@@closedforever844 Cheers for the info. 🙂
Quality of the upscale is fantastic, but the skin is too perfect, it needs more skin details and imperfections for it be a believable portrait. This also probably depends on the input file but still, many upscalers make the skin too perfect and fake.
SUPIR killed all upscallers.install him
@ 7:11 the hair has a very weird texture
equivalent for comfyui ?
not impressed (sdxl, same model and settings as here). my initial image is superior to the upscale which is just overall less detailed. played around a lot with the settings. nah. just generating 2048x2048 without upscaling is just miles ahead of this + WAY faster. Upscaling is too slow to be useable imo + requires immense amounts of VRAM in comparison to just doing 2048x2048 normally. I couldn't really generate much above 2048 here without running out of vram lol, so what' even the use.
I had the same issue, but I fixed it by closing down Automatic1111 (not just using the restart UI button) now I can make 4K and 8K images without running out of Vram. Also you need to uncheck "Keep input image Size". Also, you have to enable the tilted VAE or you will run out of Vram and the image Resize setting is needed or it will not upscale the image. Olivo made a few wrong assumptions when explaining it I think.
@@jibcot8541 even so, i don't think it's adding any detail really, quite the opposite for me.
Yeah, I haven't found any settings I like that don't make it cartoony, also it doesn't "honour" TensorRT setting so it cannot be used for the 60%+ speedup like it can with the Ultimate SD Upscale tiled upscale script. @@Vurt72
What's so special about "raw photo" prompt? I'm always using "realistic photo" prompt instead.
nothing, depends on what your model is trained on
Raw photo = real photo
Realistic = art style mimicking real photo.
It does give different result. If you want photo and not advenced digital art always go for raw.
Not very impressed tbh. Details are mushy, skin texture looks like it has some patterning artefacts and I can see the vertical and horizontal seams going through the center of the image. I think other upscaling methods such as ControlNet Tile are superior.
The tiling is a bit too obvious
"Look at those skin details"
Literally zero details on the skin, like nothing.
8 minute render and the results are terrible, I did everything you did. Extras is done in seconds and works great! What a waste of time!
I don't agree it isn't amazing - skin looks super fake especialy corner eyes and lips - the most focus point in the portrait image.
Ultra super duper 16x resolution for ... plastic skin. Sad, but true. It's render, painting, anything. Admit it already.
Is it possible we could use our own pictures rather than generated we don't like to up scale generated pictures we want to a scale our own captured pictures is it possible to of scale like that
Try it through upscaling in the Extras tab or through ImgToImg. That might work.
what if i want to use that one upscaler for upscale our own images not genrated is it possible with this upscaler ??@@Elwaves2925
Isn't this 5 month old extension already obsolete? This feature was added to base A1111 already. Why use this and not the built-in hypertile u-net/vae?