First i recommend reading the Part 1. It explains the extension and settings. Remember to leave some ⭐(～￣▽￣)～
I have planned to expand more on multidiffusion tutorials:
Workflow on multidiffusion + controlnet tiling
Maybe inpaint workflow tutorial
If you have something you would like to see tutorial on let me know in the discussion.
Please leave feedback and images you manage to create with the tutorial :)
⬇Updates and information below ⬇
Updates:
05/07
V1.0 of Region prompt control
Tutorial focused on Region prompt control.
05/01
V1.0 of Multidiffusion IMG2IMG Workflow
Tutorial more focused on scaling in IMG2IMG
04/29
V1.2 of Multidiffusion upscaler how to use + workflow
Clarified few things in the tutorial
04/25
V1.1 of Multidiffusion upscaler how to use + workflow
Fixed some typos, uncompressed images, wording
04/24
V1.0 of Multidiffusion upscaler how to use + workflow
multidiffusion upscaler for automatic1111
Disclaimer
This is information i have gathered experimenting with the extension. I might get something wrong and if you spot something wrong with guide, please leave comment. Any feedback is welcome.
I am not native English speaker and write such text. I can't do anything about that. :)
I am not the creator of this extension and i am not in any way related to them. They can be found from Gitgub . Please show some love for them if you have time :).
Description
FAQ
Comments (122)
Thank you for sharing, this is great helpful.
No problem! Please remember to leave stars from the right side next to the images (～￣▽￣)～ ->
Very helpful for understanding the prompt regions and how they work, thank you! I was trying to use this in replace of Ultimate SD Upscale, but anything over 2x can not allocate enough memory for some reason. Even with much smaller tile sizes and batch at 1, I can't go past 2x. With the Ultimate SD Upscale I can go all the way to 8x with 1024 tiles. I don't know why it's not working there.
Thank you for your feedback :). Did you enable tiled VAE? Try it with settings like Encoder Tile Size 2048 and Decoder Tile Size 112. Try lower if it still fails.
Thanks for taking the time to write these up. Just read through them and can’t wait to test it out!
Thank you. Please leave review from next to the images "Add review"! :) And let me know if you have some problems,
Using tiled vae keep giving output with partial blank tiles. How do I fix that?
This is due to the AMD DirectML large tensor bugs. It may not happen for specific tile sizes, so please try more tile sizes & image sizes. And you may update the DirectML to see if they fixed the bug.
I keep getting this message "[Tiled VAE]: the input size is tiny and unnecessary to tile." despite following the recommended "split to 2x2 = 4 tiles. Optimal tile size 96x112, original tile size 112x112". Despite I have the exact same settings as yours in the pdf file. I noticed too that yesterday it was fine and I can generate images in 50seconds-1minute. But today it goes up to 1m30s up to 3minutes sometimes.....Any idea what's happening?
There are 2 different kind of tiling:
1.VAE tiling and then the tiling of the image. If the VAE is unnecessary to tile it will not tile, it will kick in if the image is big. I have most likely gotten this with the same settings i have in the tutorial.
2. Tiled diffusion tiling, this will tile the image and will give different kind of error if it is too small "Tiled Diffusion] ignore tiling when there's only 1 tile :)"
Reducing the Latent tile overlap, will reduce generate time, but can make some weird things happen in the image.
You might have forgot to put Latent tile batch size to 8.
Those are things i can think top of my head now. Let me know if these don't help :)
@Potatovision Hi thank you for responding! I think I may have found the cause. I used to always have "hires steps = 10" before but I forgot to turn it back on. Apparently being 10 makes my generation time faster, and turning it off (0) makes it longer which is pretty weird to me. But I did google some stuff, apparently setting hires steps cancels out steps on "2nd pass"? (I normally use Sample steps = 25. So setting hires steps to 10 cancels 10 hires steps. I'm sorry it's quite hard to explain for me). On your response, does having "[Tiled VAE]: the input size is tiny and unnecessary to tile." considered like a downside (like it's slower, inefficient, etc)? I have my Latent tile overlap set to 56, same as yours but sometimes I set it back to 48 which I believe the default when I was testing. I also have the Latent tile batch size to 8, but I only have 8GB VRAM so I set it to 1 or 4 when I get out of memory error. Hmm, if having the tiled vae error not a downside, then I supposed it's safe to ignore. Sorry this reply has gotten too long.
@ddoscv Yeah no problem.
Have not lately used much hires. Didn't know about that 2nd pass thing 🤔
It takes a bit longer if it tiles the VAE. So there is no downside on the "no tiling" on the VAE.
Also I recommend testing around with the tiling settings, i rarely use same tiling as i mess around with resolution a lot. Same with the overlap, if you start seeing double heads, 7 hands, 3 legs it means you have too low overlap most likely :P I used a bit higher on the tutorial as it works all the time.
If you have GPU that supports xformers, remember to put that on if you have not done it yet, it helps a lot with VRAM usage.
@Potatovision From what I read, setting hires steps to 0 means it'll apply the same amount of sample steps on your image on the upscaling process. I found this to be true too using the antiburn extension, there's an option there to save snapshots of each steps (I set Sample steps to 20, and hires steps to 0, it gave me 40 snapshots. 20 as "first pass" using original resolution 512x768, and another 20 as "2nd pass" in the upscale process, I set it to 2.5x for 1080x1920 final size). I'm gonna go slightly off topic here, but usually hires messes up some of my image like having multiple bellybuttons or anuses......But my guess is that it's because I'm using a lora not trained for 512x768.. Going back to the topic, thanks for the advise, I will keep these in mind. For now I'll just ignore the tile vae error message in the cmd, though it quite bothers me such error exists even it's harmless.
@ddoscv Ah, understood wrongly the message before. Yeah it uses same amount of steps if it is set to 0 :)
It is a bit tricky to use tiling with Hires, if you go too high with the upscale, you have too many tiles. Usually raising latent tile overlap and trying to maximize tile height/width helps with it. Maybe try creating higher res image before hires and scale it like 1.5x with little higher denoise. And also check out 4x_NMKD-Siax_200k upscale model. I personally like it. Helps with detail. Can be found in https://upscale.wiki/wiki/Model_Database under "NMKD Siax ("CX")"
Have you tried the img2img method without using hires? Or after smaller hires upscale -> img2img upscale. The tiled diffusion really shines in img2img.
What i have noticed is, the resolution does not really matter in the end (at least with the tiled diffusion) You can get good results with any resolution. I very rarely use the specific resolutions that have been set to "stone" so to speak. And if some resolution does not give good results, just jump few pixels up or down and it usually fixes it :P
Anyway gotta run, let me know if you have something you would like to know. I have spent too much time with this stuff (～￣▽￣)～
@Potatovision Hey, thanks again for the reply! I would really like to have something slightly permanent but I guess it's really impossible due to the complexity of each loras being used. One just works fine without tweaking the settings much and some just wont.
As for the multiple anus or bellybutton problem I have, I spent almost the whole day trying to figure out what's causing it. It turns out it was just the denoise. Although I've been using only 0.4 for several months now and I thought it was already "low" enough. I only lower it when I do img2img. So my default now is 0.3 which seems to moderately fix the problem but not completely but it's much much better now compared before, which can then be easily fixed through inpainting.
My usual workflow now is to render without hiresfix at 512x768 then send it to img2img for SFW or softcore themes. Or sometimes do 1.5x hires with 10 hires steps (takes only like 30 seconds to generate), then upscale it again in img2img using 4xUltrasharp at 1.7x upscale since it's the closest to 1080x1920 (which then takes like another 1.5minutes). Also sure, I'll give the upscaler you mentioned!
And I think I'm pretty set about my settings for future image generations. Right now I haven't encountered any other problem yet. So many thanks for this guide, the tiled VAE and multidiffusion setting is so helpful for me with only 8GB VRAM. I was only able to generate 1024x1536 before, but now I can push it to 1280x1920 consistently without getting memory error. Sometimes It can go up to 2048x3072 with a little bit more tweaking in the settings. Thank you for all the advice!
@Potatovision And thanks to the entirety of this conversation. I was able to find a setting as a middle ground.
@ddoscv The upscaler you use affects the denoise you can use. With some upscalers you can go higher without getting bad results and some give bad results on low denoise.
And no problem i mostly like helping people and the extension is really nice and needs some love :) Let me know if you have some other problems. I usually have some answers (～￣▽￣)～
@Potatovision Hello again, is there a way to make the tiled vae and multidiffusion setting save or be loaded when I click the "Read generation parameters from prompt or last generation" button under "Generate"? I kinda find it annoying to always set the parameters up for these two everytime I boot up.
@ddoscv Sadly this is not possible, Technically you could edit the settings from the files to match what you want. I asked for it in the github, throw thumbs up or something so it might be added sooner than later
(～￣▽￣)～ https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/issues/183
@Potatovision I see, I really hope it gets added eventually. Thank you!
@Potatovision Hi, I'm back lol. It seems SD got another update that makes it read the lora metadata instead of file name (which I find really annoying). I've made a post about it on reddit, and someone said the lora will still work even if you type the filename in the prompt region...which again I also find annoying. I already tried updating my extensions and reinstalling pytorch and xformers but didn't fix the "problem". I'm curious if there's a workaround for this.
Edit: Sorry I keep asking you, there really isn't much people who reply to my questions. Not even in reddit or civitai discord server.
@ddoscv Yeah no problem, i usually answer everything people ask from my knowledge, does not take much time. Don't really research on stuff i don't know and people ask :P
I am personally still sticking to 1.1.1 . Not updating for week or so for now, as i don't see much point in doing so. Extensions need time to catch with SD updates and the update always breaks stuff sadly.
I would revert back to 1.1.1 unless you really need something from the 1.2.0
@Potatovision I really hate how delayed civitai notifications gets updated...
And I just forgot to remove the git pull line in the webui-user bat, and this what happened.... And I didn't know you can do that, I'll looked up how to revert to an older version but I haven't done it before so kinda quite scared of messing it up. It mentioned like getting the hash you want for older versions but I don't know where to find the previous version. Github's ui looks intimidating to me since I don't really use the website.
Edit: I found a solution ^^ I finally removed that git pull from my webui-user.bat lol. Thank you
@ddoscv Haha, nice. Yeah the civitai notification is quite delayed. I feel sorry for everyone who follow these some what idiotic recommendation to put git pull into webui-user.bat ╰（‵□′）╯ Seen too many youtubers recommend putting it there. Not really good idea for most of the people who are not tech savvy. One day it just stops working and they don't know why.
@Potatovision It wasn't really a solution but rather just a temporary workaround. Of course, it's without any downsides.... it's in the same reddit link I posted here, same with the downside I just mentioned.
I hate not being able to receive the latest optimizations but the most recent updates lately just brings more headache than convenience. Even it's more optimized in the future but the lora metadata usage stays the same, I don't think I can keep continue using that. But I heard from someone that trigger words aren't really necessary for the lora to work. Maybe if I can get used to having garbled numbers and letters lora name, then fine....But I'd rather have a clean-looking prompt.
Just like you said, if I wont be needing the new features, I'll just stay with the older versions and wait longer til it's more stable and doesn't have shitty "features" that nobody asked for. I know it sounds like I'm being ungrateful, but it's quite like a common theme with big companies nowadays.
@ddoscv Yeah optimization updates are nice. Don't see much in 1.2.0 so no hurry there. New features are nice, but it would be nice if there was "Stable version" and "unstable version" instead of just push, wait for feedback, fix :P
BUT can't really complain as this is free work from so many nice people, big respect for them for continuing to do this! (～￣▽￣)～
@Potatovision Hi, I'm back with another stupid question. I have started using region prompt control lately but I've encountered some problems:
1. Sometimes when generating background-only pictures, the image is like split into 2-3 parts horizontally, and repeats the image like this one https://puu.sh/JI5mP/a06043dbf4.jpg
2. When I have a background and a foreground at the same time, the generation time takes like a 30x longer, from 30-40 seconds to like 12-15minutes. Is this normal?
3. Same as #2 but my foreground objects tends to become cursed like this one https://puu.sh/JI5lr/aa3e54fbda.jpg I just interrupted it immediately when I saw it in live preview looking like that.
Any clue what's going on? Here are my settings https://puu.sh/JI5mc/3fffbcab28.jpg
I also typically use a lot of LORAs since it's quite my set-in-stone settings, so I'm not sure if that's another factor...
@ddoscv Can you show the region settings you are using
@Potatovision Hello, it looks like this when I was generating a background first: https://puu.sh/JI6Mj/b3ce5bb7ff.jpg
Then I add the seed in the region-prompt-control once I found the one that I like. Then I start to enable region 2 and set it as foreground like this https://puu.sh/JI6Nb/b53cfe6a73.jpg (also note that I just simplified the seed in region 1 and the prompts in region 2 because I already deleted the image and the notepad with the parameters in it. But I used around 5 LORAs).
I thoroughly followed your region-prompt-control PDF guide, most of the settings are identical. But I'm definitely just doing something wrong lol.
@ddoscv Are you using the latest version? I saw some problems with lora with the Hires. fix. with other extensions. Have not personally used region prompt control in a while.
@Potatovision No I am not using the latest version. I rolled back instead to the previous one prior to when they changed how lora names appear on the prompt, I think it's pre 2.0, I forgot. I am also quite afraid that updating my plugins will break something, but I can try to make a plugins backup and try updating those tomorrow. I am going to sleep. I will post the results tomorrow, thank you. https://puu.sh/JI74A/4ff5d465ca.jpg
@Potatovision Hi are you using lora block weight extension? https://github.com/hako-mikan/sd-webui-lora-block-weight
When I updated my extensions and even I restored the backup extensions, my block weight extension just won't work anymore...... Mine had values in this part https://puu.sh/JIa41/b48043c4b7.jpg but it's blank as you can see in that screenshot.....
https://puu.sh/JIa84/fd55173258.jpg This part in the cmd usually have lots of 0s and 1s and specifies the lora I have with ":MIDD/:OUTALL/:INALL" but it's not there anymore...
This is another major extension I use, without it functioning, I can't make good images with multiple LORAs.... :(
@ddoscv Ah, i have used that, but have not used it in a while. Sadly i cannot help with it. Maybe the extension github page has some mentions about how to solve the issue. Seems like the hires update in the newest A1111 messed up many extensions. :/
@Potatovision Does yours have anything on that blank area I've mentioned? Because even reinstalling the extension didn't fix the problem for me sadly, and up until now, I'm still looking for solutions. I left a post/question in the github from the creator too and also on reddit but no replies yet. Thank you.
@ddoscv In middle of something atm, i'll try to remember to check later :P
@Potatovision Sure no problem. If you don't mind, can you check the values on these areas too? https://puu.sh/JIejs/2737954751.jpg This is really the first time in months that I can't work with new images ever since I installed SD and this extension. I'm sorry if I'm asking too much of you.
@Potatovision Hi, I finally fixed the problem and the extension is working again. If ever yours won't work anymore if you do a rollback, all I did was to download the zip file in the github page, and deleted the old Lora Block Weight extension folder (mine is from E:\stable-diffusion-webui\extensions) and replace it with the new one that came in the zip file. I just thought I'd let you know too. Sorry for all the troubles.
@ddoscv Nice that you got it figured! was really busy day and forgot. Sorry about that 🥴
@Potatovision No problem! I got kind of busy on other stuff too anyways and I wasn't checking civitai much when this has happened. I was ready to completely give up if I won't be able to fix it lol. Thanks again.
Thank you very much for your intelligent idea on controlnet! I have been long for this function.
But I want to know if it's possible to draw some region in irregular shape?
Hey, glad to see you liked the tutorial. Do you mean the prompt control? Sadly you cannot shape them differently
Check out this extension, https://github.com/ashen-sensored/stable-diffusion-webui-two-shot it lets you draw on canvas. It works together with multidiffusion.
where CFG scale?
Sorry i do not understand
@Potatovision Thanks, I saw it.ESRGAN UltraSharp
@Shivae Hey, you can put it in ...\stable-diffusion-webui\models\ESRGAN
@Potatovision Good morning, working day
Glad you got it working :)
@Shivae Did you restart stable diffusion?
@Shivae Ah, it needs to be in \stable-diffusion-webui\models\ESRGAN folder
@Potatovision I put it into ESRGAN at first but couldn't see it, when I downloaded it there were three files, I put them together and then couldn't see it, now I put the .PTH file in separately and I can find it in SD, I need to do some ultra realistic scene classes now and need this kind of zoom support. Thank you, big brother/sister.
Hi, thanks for the work! Unfortunately i'm getting error messages when i restart UI after doing the installation. Here is what i get(after the restarting UI notice) : Error loading script: tilediffusion.py
Traceback (most recent call last):
File "C:\SD\webui\modules\scripts.py", line 218, in load_scripts
script_module = script_loading.load_module(scriptfile.path)
File "C:\SD\webui\modules\script_loading.py", line 13, in load_module
exec(compiled, module.__dict__)
File "C:\SD\webui\extensions\multidiffusion-upscaler-for-automatic1111\scripts\tilediffusion.py", line 71, in <module>
from tile_methods.abstractdiffusion import TiledDiffusion
File "C:\SD\webui\extensions\multidiffusion-upscaler-for-automatic1111\tile_methods\abstractdiffusion.py", line 11, in <module>
from modules import devices, shared, sd_samplers_common
ImportError: cannot import name 'sd_samplers_common' from 'modules' (unknown location)
Error loading script: vae_optimize.py
Traceback (most recent call last):
File "C:\SD\webui\modules\scripts.py", line 218, in load_scripts
script_module = script_loading.load_module(scriptfile.path)
File "C:\SD\webui\modules\script_loading.py", line 13, in load_module
exec(compiled, module.__dict__)
File "C:\SD\webui\extensions\multidiffusion-upscaler-for-automatic1111\scripts\vae_optimize.py", line 71, in <module>
from tile_utils.typing import Processing
File "C:\SD\webui\extensions\multidiffusion-upscaler-for-automatic1111\tile_utils\typing.py", line 13, in <module>
from modules.sd_samplers_kdiffusion import KDiffusionSampler, CFGDenoiser
ModuleNotFoundError: No module named 'modules.sd_samplers_kdiffusion Do you have any idea of what i'm doing wrong ? Thanks !
Hey,
I can't really help with the errors. You could ask help in github page https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/issues
You could try to delete the extension folder \extensions\multidiffusion-upscaler-for-automatic1111\ and try to reinstall the extension. Sorry that is only thing i can recommend testing.
@Potatovision Hi, thx for answering, i did a full clean installation and it seems to works. I'm not quite sure but maybe i was using a wrong version of python.
@Pepo02 Could be, never know what breaks what (～￣▽￣)～
Just finished following your multidiffusion tutorial. Great stuff! Small question though, after trying it out, it kinda felt like the tiled diffusion and tiled VAE are making gens a lot slower than when you don't turn them on, especially during inpainting since you don't often need to inpaint(only masked) at too extreme resolutions. Does this mean that if your GPU has enough VRAM, it's better to not use the tiled diffusion and VAE in the workflow?
The VAE tiling does not really make the generation slower, unless you use full preview.
Tiling makes the generation longer, as it will tile the whole image and generate the tiles separately. Lowering Latent tile overlap can help with the generation tile, but might implement some random things in the image (～￣▽￣)～ .
The tiling helps with the VRAM, but also increases the quality/detail/consistency on the image. Especially on the img2img upscaling.
Thanks for the comment!
Interesting.. it just feels difficult to find a solid reason to use tiled diffusion instead of just the normal workflow of
1. t2i(with hires.fix)
2. move to img2img for inpainting for some modifications
3. finally use ultimate SD upscaler with controlnet tile preprocessor
within those three steps above, there just doesn't seem to be anywhere I could try and incorporate tiled diffusion/tiled VAE(which I tried but just ended up making generations slower so I don't think it's worth it). Unless there's something I'm missing..? Anyway thanks for the reply and still a great guide!
@cleopatracleo If you find your workflow solid, there is no reason to change it. Tiling in t2i does generate more detail but the idea in i2i is same for both ultimate sd upscaler and tiled diffusion in the long run combined with controlnet tile preprocessor. :P
Tiled VAE comes in handy if you run out of VRAM because of VAE, it does not slow your generation that much and only tiles it if it is necessary.
Thanks for comment
In fact, if you feel too slow, you can use mixture of diffusers and set the latent tile overlap to a very low value e.g., 4 or 8 (this is equivalent to the overlap in pixels in the Ultimate SD Upscaler).
Because our extension support denoise 8 tiles simultaneously, you can get significantly faster speed than Ultimate SD Upscaler if you have enough VRAM. The default overlap (48) is too high from today's perspective, especially when you use ControlNet Tile model.
Stupid question, but where is the guide? Is it on the github link?
On the top you can see 3. region prompt control 2. img2gim... 3.... And when you have one selected you can download from right, with the blue download button
@Potatovision Thanks for explaining. What would you say is the difference between this and doing ultimate SD upscale script?
@handicream This is answer from the creator.
"liyi20129744: Because our extension support denoise 8 tiles simultaneously, you can get significantly faster speed than Ultimate SD Upscaler if you have enough VRAM. The default overlap (48) is too high from today's perspective, especially when you use ControlNet Tile model."
On top of that multidiffusion can Tile VAE, has region prompt control and own noise inversion (which i have not used much)
Both do upscale in the end :)
@Potatovision yeah both do the upscale but this is much better in my experience and without any visible seams so far in my experience.
Great guide, seams hard djaskdjsa, gonna try it tonight. What's BREAK means on your prompt?
Hehe. BREAK starts new set of 75 tokens to put it short.
You can read more info on here https://civitai.com/models/54917/breakdomainrealistic?commentId=117309&modal=commentThread
robproctor83 explains it pretty well.
I don't quite understand why this extension is needed. The original resolution is quite low, and Hires.fix already works with tiles. Then what's the point?
As far as i am aware hires. fix does not tile the image. If you are talking about the "Tiling" checkmark next to Hires. fix, it produces image that can be tiled.
@Potatovision Tiling checkbox in txt2img is a different kind of tiling. It is to create an image that can be infinitely attached together in every direction like chessboard patterns for wallpapers.
@GuybrushPluto Indeed :)
Hi, I find your workflow really interesting, I'm trying to follow this guide, going step by step as you show, blocking the seed of the element I like and moving on to the next one. Only I noticed that the key to recycle the seed inside the region control doesn't work, and I blocked the generic seed of the img because otherwise the main prompt would conflict with the background prompt... where can I get the seed of the single region then? am I doing something wrong?
ps: I'm not using the "draw full canvas" option
This is a bug and I fixed it on 5.23.
@liyi20129744 i have update the addons and now work! Thanks for the tip!
@N0tt Sorry, somehow i never got notification for this comment.
@liyi20129744 I apologize for the inconvenience. I have encountered the same issue. I downloaded the compressed installation package from your website and also tried installing directly from the provided URL. However, the problem still persists. I am unsure where to find the download for the fixed files.
There is a direct and proportionate relationship between 'Encoder Tile Size' and 'Decoder Tile Size' found in Tiled VAE and the output image resolution. This should maybe be made clear in the guide.
Oh. I was not aware of that myself. Thanks for pointing it out. Have to look how it calculates.
@Potatovision It looks like the encoder/decoder tile size cant exceed the pixel width/height of the image, or it returns the "input size is tiny and unnecessary to tile". Do you know if the "optimal tile size" thats qouted, how it derives this?, is it calculating this to round up/down to remove any decimals from the tile size?
@Roscosmos As far as i am aware it is good to keep as high as possible, If it does not tile it necessarily is not problem as long as you don't run out of VRAM. I do not know how it is calculated. Maybe better idea to look around in the github page if you can find more information on this.
Hey, can you to a tutorial of best settings using tile diffusion + tile vae with controlnet tile resampling + controlnet tile colorfix + vae, I just cant find a good settting and my system crashes ?
Hey, seems like the tile sharp + colorfix thing is using a lot of VRAM, try tiled VAE settings: Encoder Tile Size: 512, Decoder Tile Size: 96. It is running out because of the VAE so no need to touch other tile settings. If you still run out of VRAM try lowering the Encoder Tile Size even more.
@Potatovision thanks. I will try it. I tried tile sharp + colorfix in txt2image and then simply used img2img with tile diffusion + tile vae & controlnet tile resampling and i am getting much better colors and details.
would you be able to put the tutorial in a google doc also?
Hi ! First of all, thanks for the great jone done ;-)
I'm starting at SD and i am in the part 1 of your guide. I did a seed, i use first highres and now i'm at the pint of img2img enhancing. At this point, i've a problem, after clicking "generate" nothing happened in SD. On the wd command, tiles load perfectly, next tiled VAE and the program seems to struggle at "mutlidiffusion sampling : : 0it [00:29, ?it/s]. Nothing move since 10-15mn now.. I restart SD (cause i can't interrupt) and a lower the Tiled VAE significatly fut nothing... do you have any idea what parameters should I move ?
Does your VRAM hit max? In the new nvidia driver there was nice feature that uses shared RAM if it does that. Try with less batch size.
Hey! First of all many thanks for this tutorial it's very helpful!
I've a question, when a have a decent image in txt2img with region prompt control enable, how do i have to prompt for img2img to keep the same image ? Cause when i send a txt2img to img2img the results are not the same when i upscale.
Thank you
Hey, try using controlnet tile_resample preprocessor in img2img with tiled diffusion, that helps a lot. Also, too high denoising strength changes the image.
Hey there. You say - First i recommend reading the Part 1. It explains the extension and settings. But which is that? There's no link and in your profile there's nothing that resembles apart 1 either.
Hey, you can see the different versions above the example images.
sorry may i asked where should i put the file into?
sorry may i asked where should i put the file into?
Nice work, just have one question, when I was applying CN tile processor into the upscaling process in img2img with tiled diffusion, it crashes and says "ValueError: too many values to unpack (expected 3)", what does that mean and do you have any ways to avoid this? Thank you!
Hey, i don't remember ever getting this error. Have been really busy lately and have not updated anything for weeks. I wonder if this is some new bug 🤔 You could try to ask help in the github page https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/issues
@Potatovision thank you for your quick response, I'll look into that.
I noticed that you don't mention or show anything for the Noise Inversion settings. Maybe that wasn't in the older versions? Now it is recommended on the GitHub page for upscaling and if I don't use it, it doesn't work very well.
I personally never really used the noise inversion.
is it possible to generate BG then add 1 by 1 object after that.. or we can only to render all of them at same time?
If you are talking about the region prompt control, you can do it 1 by 1 like i do it in the tutorial. Though the background will have slight changes depending on the other areas.
I like this tutorial. Only read it and now going to try this extension. See that in 3d part you want to do 4 but it's not appearing. Do you plan to make another? Sorry for my Eng.
Thank you for your comment. Sadly i had no time to do more. Maybe eventually i will do more,
Has anyone managed to successfully combine After Detailer with this?
I tested it once, and if i remember correctly it was throwing errors (～￣▽￣)～
Throws errors because there isn't enough space on the graphics card. A 4k pic requires nearly 100 gigs of vram. Although I'm currently writing this without testing it because I'm waiting, an image is saved before Adetailer. Using that image, you can crop, in my case, the face out of the image an then pass it through img2img with Adetailer. Should work. The extra step is worth it.
Where and how can I read the tutorial?
Is there a link for it?
click the download button, you'll get a zip file that contains a pdf file
@maeyhim907
Oh yeah I could read it, thank you so much:)
@hkuroda0222543 Seems like i didn't get notification for these comments, there is 3 different tutorials you can read. You can change which on to download from the top under the main title, just above the images.
(～￣▽￣)～
@Potatovision
Thank you so much!
Will try to read🔥
Partition control has been very helpful in my work, but after updating the XL model, it is not available, whether the plugin will fit the xl
Good to know it has been helpful, sadly i do not personally use A1111 anymore so i do not know how well it works with xl.
These tutorials help me a lot regarding how to make a good upscaling. Currently learning how to use regional prompt.
Do you plan to continue with more tutorials?
Hey, thank you for liking them! The tutorials might be a bit outdated, but should give good general information.
Sadly i have personally switched on using comfyui so i wont be doing any tutorials regarding A1111 . ;(
how to install your extension are there any prerequisites required to install prior to this .
And please can u give me the link for the first part of the tutorial how to use?
Hey, you can see all the parts of the tutorial from top. Clicking on the 1. "How to use + T2I workflow" you get on the first part and can download from there.
In the tutorial it does explain how to download the extension for Automatic1111.
Note: The tutorial is over half years old and can be outdated, but should give some general information still.
Thank you for taking the time to explain this.
Looks like seed in regional promt didnt work, it only looks on base prompt seed
Hey. Don't know the current state of the extension. For regional prompting i recommend regional prompter extension https://civitai.com/models/339604/how-to-generate-multiple-different-characters-mix-characters-andor-minimize-color-contamination-or-regional-prompt-adetailer-and-inpaint-or-my-workflow or forge couple if you use forge.
@PotatCat thx for the help mate
@capzed293 (～￣▽￣)～
you say "no need highres fix" to get good pics in part1 -6.2 Second step, Hires generate,but the demo pic show highres fix parameter in text2image,I wonder why?
I don't exactly remember what i wrote there over 2 years ago. Most likely was something about that you do not necessarily need highres fix if you upscale in img2img with multidiffusion or when doing t2i with multidiffusion enabled.
To be honest this guide is over 2 years old, back then there was only 1.5 models and the knowledge on image generation was much much more limited. I recommend looking something more up to date.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.