Optimize Automatic1111 V1.6.0 For SDXL 1.0
Вставка
- Опубліковано 6 сер 2024
- Optimize Automatic1111 V1.6.0 For SDXL 1.0 . Hey everyone! So, I think this will help a lot of people with 8GB GPU's or less! I will say the performance gains didn't come from the update to Automatic1111 V1.6.0 as this also worked for V1.5.2 however SDXL 1.0 is running quite faster now for me! For all the technical details, check it out below!
Automatic1111 (install instructions and update list)
github.com/AUTOMATIC1111/stab...
Trouble shooting tips for Nvidia GPU's github.com/AUTOMATIC1111/stab...
Tips for AMD users github.com/AUTOMATIC1111/stab...
Tiled VAE extension can be installed via extension tab
github.com/pkuliyi2015/multid...
⏲Time Stamps
0:00 Settings and results
1:24 Updating to an existing Automatic1111 version
1:51 Command Arguments to use (if applicable)
2:42 Install Tiled VAE Extension
3:41 Shout out to Double Bob and Live Previews
4:13 About Tiled VAE Encoding
5:10 Speed comparisons to be revised!
*Disclaimer Affiliate Links Below*
📸 Gear I use
Sony A7C CAN amzn.to/3spWX8C
Sony 35mm 1.8 CAN amzn.to/36Apekr
OBS (FREE) obsproject.com/
Editor Davinci Resolve www.blackmagicdesign.com/ca/p...
Audio: Rode Podmic amzn.to/35sSnxv
🔦 Find us on:
Discord: / discord
Instagram: / monzonmedia
Facebook: / monzonmedia
Ok folks, I think I will have to redo my tests with my new findings, the previous test video will be removed as it is now misleading. Going to test over the weekend and come back with a more accurate result. There are many variables to consider so perhaps a bare bones test out of the box? What would you suggest to make it reasonable and fair?
I'm just here to watch and learn. XD
You don't need to delete the previous video, just rename it with some sort of "Outdated - " prefix and update the description to tell a little bit about why it is outdated, and maybe a link to the newest video. At the speed this technology evolve, you would end up deleting a lot of videos, while they are still good to keep as reference (where we started from). o/
@@EskaronVokonen That's true and plus if I delete it or make it private I lose my views and analytics but yeah would be good to keep for reference. I like the idea of putting something in the title. Good call!
Hey dude! Big thanks to you! You really helped to a lot of people.
Glad I could help!
Thank you for the great tips! I always get happy when I see you've posted a new video because I know I'm about to learn something useful! Thank you for the great content! :D
You’re very welcome as always and it’s my pleasure! I’m glad that you learned something as that’s always my goal! 🙌🏼👍🏼
This! Thanks! Down from 1+ min renders to 29s. Also, first time seeing your channel. Great speaking voice & great content! Subscribed.
Glad it was helpful!
Thanks for the update! I apologize if I missed you mentioning it in one of your earlier videos; but, if you include "git pull" to the webui-user.bat file, you'll have the updates automatically. Thanks to your comments the other day, I knew a new version was coming and was very excited when my interface was all shuffled around a bit upon opening Automatic1111 this morning. If anyone's interested, after applying your settings, my render times went down to about 23 seconds. I'm using an old GTX1080 with 8GB of VRAM. Thanks again for keeping up with all the changes!
Hey Daniel! Yeah good point and no worries about git pull being in the webui user file. I’m so used to using my launcher it had slipped my mind to add it as well.😬 glad to hear it worked out for you! Mind you other samplers like some of the new ones do take longer. But a world of difference now! I will be adding more tests with samplers when I do the updated comparison.
Have fun bud, more to come on this update!
This took me from 30 minutes to 6 minutes, Thank you!!
Many thanks. Love your clear and relaxed presentation.
You’re welcome! Glad it helped 👍🏼
You the man a lot faster, I was waiting 5 to 6 minutes. for one image.
Whoa that’s way too long! Glad this helped you you out! Curious to know your specs and what you are getting now?
Thank you!! It's fine!!
Fantastic! This helped me a lot!
Awesome! Glad it helped! SDXL for A1111 has been a mess hahaha!
great video - just on time!!! GJ
You’re welcome my friend!
this actually helped lets goooo!!!
Great to hear! You should check out Web-UI Forge, it's a bit faster than A1111 github.com/lllyasviel/stable-diffusion-webui-forge
A good test might be out of the box like you said to see how they compare then maybe attempt to optimize each platform and see how that affects render times. As others have stated, no need to pull the old video just put in the notes that was a preliminary video and the updated one is...
Great effort. Me, I'm going to stick with ComfyUI for speed and ED at the moment, but it's nice to see a video about how A111 does, saving this video for the info! Great job!
Yeah I think I may go down that route...maybe throw in a bigger aspect ratio and perhaps use some of the new samplers too. Appreciate you bud!
@@MonzonMedia same!
Much appreciated boss, its so much faster
Glad it helped! Now if I can find a better way to optimize for SDXL ControlNet, that would be great! 😁
saved the day!
Glad it helped! Have you tried Forge? It's the same as A1111 but better optimized. Highly recommended. Much easier to install and no need for those command arguments. ua-cam.com/video/FKzvHFtc8N0/v-deo.htmlsi=J5dxVaUM42ZBcK2K
Hi ,this works really well for the last section of the render .It use to pause for about the same amount of time as the rendering took to reach 95% .Now a couple of seconds and it's done .However I find that using the refiner stops the tiled vae from having any effect .Strange .
Thx for this video too ;)
You're welcome! Hope it's helpful.
I have no idea what those numbers are, but wow props to you Kuya. 👏🏻
hahaha! Well thanks for the support regardless 😁👊🙌 I can explain it to you one day!
@@MonzonMedia . Will be waiting for it! 👌🏻
Note. Tiled diffusion is also WAY better than ultimate Sd upscale for tile based upscaling. No tile lines
It's a great point! Looking to do content based on this.
Great video, ComfyUI has a tile VAE type nodes, but i haven't tried it out or tested what speed difference it would make
Hey Ramon, always appreciate your support my friend! Yes after I made this video I went to check the other platforms, ComfyUI was the first one I checked and you're right! Will be testing them all this weekend once again. 👍
You are a hero of sorts.
😬 😊 I’m assuming this was helpful? I’m glad to hear it! 🙌🏼
It zooms now. Almost as fast as 1.5 and I HAVE A 3060 TI 8GB TOO! Needless to say, I'll be using SDXL a lot more in the future.@@MonzonMedia
Amazing bud! I kind of feel silly making that previous video now but I’m glad I was wrong! Hahaha! Enjoy man!
Thanks for this info. Brought me from ~50s down to ~30s per image on a 3070 8GB using SDXL base , dreamshaper XL, and a couple others @ 1024x1024 30 steps. Sadly though, for me SDXL+A1111 is still unusable, as using anything beyond just a base checkpoint alone results in unbearable generation times. Like 5+ minutes for 1.5x latent upscale per image.
No problem at all! Makes a big difference doesn't it? But yes, sadly especially with version 1.6 it's so buggy. I'm discovering a lot of the extensions I was using on V1.5.2 are no longer compatible with 1.6 and there seems to be very little optimization for SDXL. I've been testing ControlNet for SDXL all day and it's taking way to long. Even with my optimizations. 3-5min generation times! See my post in the community tab. I tried SDXL CNet in ComfyUI and it's done in less than a minute! So much faster! I love A1111 but the lack of optimization is too much at this point. And when it does get optimized it takes awhile, Still great for SD1.5 but a headache for SDXL. I'll let you know if I find some solutions.
Thank you very much for this helpful video!
I ran a test on my RTX 4070 (12GB, AMD Threadripper 1950x) with your prompt:
Without --medvram-sdxl in the webui-user.bat render time was 40.9 sec. With --medvram-sdxl in the webui-user.bat the render time was 25.2 sec. And after activating Tiled UAE with your values (i.e. Tile Size 1536 and 96) the render time is now 18.6 sec.
You could also say: doubling my lifetime 🙂
So glad to hear it was helpful! I'm really hoping that SDXL get's better optimized soon as much as this helps, when you start using other samplers and of course generate more images at a time, it's still not quick enough.
@@MonzonMedia I have since found that the Tile Size of 96 gives me slightly blotchy images. At 64 it becomes even more obvious. I have to go to 128 so that's not the case. As a result, generating the images takes just as long as without Tiled VAE, i.e. approx. 8-10 seconds more, depending on the image size.
Just in case the same thing happens to anyone.
Interesting, I haven't had that problem. What dimensions are you generating and what are your settings?
@@MonzonMedia I use checkpoint RealVisXL V1.0 from civitai. I generate the first picture (close-up of the blonde beauty's face) with the prompt given there. As VAE I use sdxl_vae.safetensors. Resolution is 1024x1024. No refiner. No Restore faces. If you zoom in on the resulting image you can clearly see the spotting. After disabling Tiled VAE, the image is as clear and clean as the image seen on civitai.
As I said, this happens with the setting 96 for Decoder Tile Size.
@@lesslyrics I've generated hundreds of images this week with no issues whatsoever, however I suspect it's the SDXL VAE as RealVisXL has a VAE already baked in. Have you tried it without the SDXL VAE? Now I'm curious to see if that's the case but I'm almost positive that is the case.
Beautiful video and big like that ❤😂❤😂❤😂❤😂❤😂❤😂
Appreciate the support 🙌🏼
Go Bob !!!!
Not sure who Bob is but ggggooooo!!! 😁☺
Anyone having issues make sure to update your torch to 2.0 and remove venv folder after you do it will reinstall xformers for you after
Very helpful, wasn't able to run any SDXL derived model on Automatic1111 until I did this. Geforce RTX 3050 8GB here. Thanks! It had to be possible, because it's working with Fooocus. Maybe you'd like to cover how to get the same photo realistic results as in Fooocus. I know the model is realisticStockPhoto_v10 and SDXL_FILM_PHOTOGRAPHY_STYLE_BetaV0.4 is used as a LoRA with weight 0,25. But then there are 3 "Styles" actived in the Style tab Fooocus V2, Fooocus Photograph and Fooocus Negative and probably some more hidden magic. It would be awesome to be able to reproduce the preset --realistic in Automatic1111 (with 8GB RAM). Maybe an idea for a future video ;)
PS: Fooocus takes about the same time for rendering a 1152x896 image as Automatic1111 with the same model and the settings explained here.
Glad it was helpful and A1111 should be a tad faster now with the recent 1.7.0 release. As for the photorealism in A1111 it's no different really. The styles that fooocus uses are just predetermined text prompts. I'm not sure about Fooocus V2 those if that style is available as text, I'd have to check.
My findings on a 3090 24GB and your exact version/variables in your video are: 1) --medvram-sdxl SLOWS down each generated image with aprox. 2 - 3 sec. 2) Tiled VAE does not speed anything things up, with OR without --medvram-sdxl. 3) Tiled VAE does however slightly slow down each each generation, and that slowing down becomes larger the smaller the tiles are. I've done plenty of testing and conclusion is that this is mainly geared - as you informed - towards cards with lower specs (and it looks great for that). 4) An image with --xformers and nothing else, per your settings, excluding Tile VAE and --medvram takes aprox 21,x seconds (1024x1024, Euler A, 30 steps). 5) Best way to cut down time on higher end cards is to tweak steps and use diff. sampler - are way faster ones out there, but that of course would not solve VRAM issues et.c if one might have those.
Hey Stefan, I appreciate your feedback. Yes these setting are mainly for people with GPU's with 8GB or less, maybe even 12GB. Medvram will not be needed for 24GB GPU's. These commands are to efficiently use RAM for lower end cards. As indicated you would only need xformers.
Helpful information...I'd love to have a GPU with more VRAM eventually to do more testing! 😁
This is great for people who would otherwise not be able to use SDXL. Sadly, tiled VAE is not without quality cost, and the smaller you make the VAE tiles, the more obvious it becomes.
Using Tiled VAE adds noise to the image, leading to a less smooth, more grizzled / textured look.
Definitely best not to go too small though, personally I haven't had any issues. If A1111 and SDXL was better optimized than none of this would be necessary but alas here we are. ☺
I just figured out (because I'm stupid) that you can use the refiner to switch Checkpoints mid rendering
Yes very handy now! And no you’re not stupid. Even though I know most people have watched the update videos from other creators I’ll still do my version 😊
hey! does anyone know how to get xformers to install? because even when i put it in the command window it doesn't install on launch and says it is not found :(
let me know if anyone has any fixes!!
HELLLOOOO!!!
Hey dude! Up late!
Great and Thanks. Got the Tiling successfuly installed. I noticed you have Roop installed. I've been unable to get that installed properly, I got nothing but error upon error. Simply getting the Extension does not work, even after installing 3 components from Visual Studio. Help from Github from the version specifically written for A111 also failed to get it right. Lots of people are having trouble with installing it, so I wonder if you could possibly do a short video on getting it installed right. In any case Thanks !
Hey Dan! Sure I can do that. Actually I'm editing a video right now where I show an example on how I'm using it to achieve consistent characters but because I don't want the video to be so long I did not go over the installation process however I do have footage already recorded on how to install it. I will post that one in the next few days. It's kind of a buggy extension, I find that I have to uninstall it sometimes and re-install it.
@@MonzonMedia That will be great. It seems all of the face-swaps depend on insightface, and some other projects, plus M Visual Studio Components.
I was going to type: "The Tutorial show the file name extension "bat" but my stable diffusion fold shows the WebUi-User but the "bat" is drop (missing) I'm a newb and just finally installed Auto1111 last week using your easy installer, (by the way, that link has updated to a new link for easy installation, not the same), would like for someone to shed light on my issue" but then after reading the comments, I realized my launcher can force an update if the update is available, and that's what I did. If you redo this video as a hybrid of this and the "easy installer", I hope you make a mention of this. Thank you for awesome videos, thumbs up
Hey Gomez, yeah I need to do an updated video as I noticed that installer is now obsolete which is too bad it worked so well but the new one looks great! I'm glad you got it figured out and thanks for the suggestion. 👍
When you do videos about performance, I think it would be helpful is you stated your system specs.
I admit this was a last minute video but I did mention my gpu. At the end I refer to this video that I have to redo. ua-cam.com/video/C97iigKXm68/v-deo.htmlsi=g3Lfm65nmBkTcXd- and I mention my specs there. But I hear ya, updated video will have all that info 👍🏼
@@MonzonMedia Thank you for the consideration.
@@cyberhard and thank you for the feedback!
The tiled VAE extension lagged out me and my friend's computers when we tried using it for some reason :c I didn't notice much of a difference in speeds anyway, at least on my system
What are the specs of your system? A1111 has been a hot mess since the update to 1.6
Hi, thank you for this update, Ermin remember the 1111 installation that you record a video, "How To Install Stable Diffusion Automatic1111 Easily" I did it so the question is can I upgrade that 1111 to this new version? and if I can how can I do it?
Does the installer not update it to the latest version? When you launch it there should be an option to auto update. Then check at the bottom of automatic1111, the version is listed there. See if it says 1.6.0. If not you can do what I did in the video. If it's still not working there is another similar way. Let me know.
@@MonzonMedia Hi, thank you for your attention answer and help, I just did what you said and voila now is 1.6, thanks.
Oh great! How’s it working for you so far? Did you get the 3060 yet?
@@MonzonMedia Hi, thanks for all you help and tips, I already both the 3060 but it does not arrived here in Brazil, as I asked for a cousin to buy it in USA, he lives in Texas. The 1111 is working very slow and I still does not tried this 1.6, I just made the update I will work with it this night and after this I will tell you my impressions.
Oh exciting! I’m happy for you! Please keep me posted when it arrives! Congrats my friend! 🙌🏼👏🏼
I'm having 4GB vram on RTX3050 so I want to know if I can run SDXL on it?
Yes but it might be a bit slow, on Automatic1111 you would have to use --lowvram in the command arguments as I showed in the video. Comfy UI is better optimized for lower VRAM cards but there is a bit of a learning curve or even Fooocus-MRE which is very easy to use. I have a couple videos on both. ComfyUI ua-cam.com/video/z8efDtdBZn8/v-deo.htmlsi=2CtISyr8lbaquGEh, Fooocus-MRE ua-cam.com/video/z8efDtdBZn8/v-deo.htmlsi=2CtISyr8lbaquGEh.
Sadly i'm still on lowram, but not eight minutes anymore, now I'm done in 4 Minutes🎉😂
That’s still quite the improvement! 😆😊
Did you abandon ComfyUI? :p
Not at all! hahaha! I've been using A1111 since the first time I started using SD and because of the nature of my channel, I do use most of the main SD platforms.
Fooocus is better, xformers is obsolet, sdp is better too
I like Fooocus and use the MRE version however A1111 still is the most versatile due to having so many extensions. Did you see there is in/outpainting now too for the original Fooocus?
@@MonzonMedia sure, But I've never had good results with sdxl on a1111, even though I've been using it since its inception and I think I've mastered it. So I'm switching to fooocus for sdxl.
How do you get fooocus? I use xformers for right now but looking to try other alternatives.