Stable Diffusion 09 How to Choose a Sampler
Вставка
- Опубліковано 24 сер 2024
- There are three groups of samplers:
1. Samplers that converge to a stable image with increasing number of steps
2. Samplers that do not converge but keep changing the image with increasing number of steps. Tthey have an 'a', from ancestral, in their name.
3. Samplers that may converge but do so to a different image from group 1. (They often have SDE in their name, Stochastic Differential Equation)
In general the workflow is:
1. Generate a large number of images in the shortest possible time to see if there is one that you like. For this phase we need a converging sampler that is fast and needs little steps to show the image it will converge to. Seven samplers are available in this group that converges, need only 9~12 steps, and are fast
2. Now select the image you like and work further on it to improve quality and upscale it. For this phase we need the sampler that gives good image quality in a reasonable amount of steps. My personal favourites are Euler and DPM++ 2M Karras,
Most samplers do not need many more steps 30. Increasing step count further may sometime render some more details but in general with the converging samplers the image won't improve much anymore above step 30.
When I first started using SD I was very disapointed since DALL-E did generate better images, but now that I've been using it for a couple of weeks, and really start to understand AAALLLL the details and settings, it's just crazy what it can do! These detailed analytical video's really help with that, thank you!
you're welcome.
Damn! Thank you for every hour you put into this!!
this video is the gold i was looking for. thank you multiple times :D
You're welcome!
Until now I've been pushing Euler to 30 because I didn't know it was getting decent results as low as 9. Very useful information, especially if you have a slow computer like me. You have a good logical approach. Thank you very much 🙂👍
Thanks for your feedback. I only use 9 or 12 for a quick scan. For higher quality I usually go to 24, or more, but past 40 there is almost no visible gain anymore.
@@rudyshobbychannel I mainly work in 16:9, 960x544 portrait or panorama, then double the size with ControlNet-Tile to 1920x1088, so this kind of information is useful for me, because I can't experiment much myself due to a slow computer. I intend to buy an rtx 3060 12Gb, I think it will be enough for average speed. Thanks again for your analysis. 🙂
Thank you for this great tutorial, it helps me a lot. I was so confused by all these sampling methods 😅
Interesting and clear info. Thanks
Fantastic explanation! I am less confused about this than I was prior to watching!
Thx for your feedback, it's appreciated.
It's a very useful test and save me a lot of time in future. Thanks
Thank you so much for putting in the work on this. Most sources I've read on which sampler to use say, "it depends", but never really explain why or how it depends.
Thank you very much for your detailed breakdown of your experience with samplers!
Thanks for your feedback, it's appreciated.
simple to follow and very useful!
Good explaination
Thank you for the explanation!
Ive started to use txt2 img from a free online service and have just instaled a1111 on my computer and i found out that dms's samplers give a more crisp img, and the 2m follows the prompt better (butt sometimes it gives some anormalities), while other methods are more "creative".
I will give a try to euler as the finder, but i liked more the dpm++ karras as the better one as it gives a crisp image, natural details and follows complex prompts very well, and its also fast to create and to get to a stable img.
Spectacular video!
Thanks for the feedback, glad you enjoyed it.
Hola, justo andaba buscando un análisis como este! muchas gracias compañero! :D
You're welcome!
I don't have DPM++_2M Karras in ComfyUI. What I do have is DPM++_2M and Karras as a scheduler... but I also have Karras with Euler so not sure if it's the same thing
Hey thank you so much, your explanation was so detailed and really enjoyed the video!
Excellent and efficient workflow...! Thank You Ruud for this series...!
Looking forward to get my own SD / GPU / PC...! Will rewatch all, when I get a machine.
Have tested, found And been using models, just like this, on Playgroundai...
Euler Ancestral for variations, changing steps, and ddim for uniformity...Learned AI Converge today...!
Greetinngs.
Thx for the feedback. Get yourself a good one, luckily video card prices normalized some.
Great video. very helpful
I have a 3070 myself and with ComfyUI the speed is OK.
I have gtx1650 with 4gb vram, and I usually let run my computer for hours 😂
It needs about 2-3 minutes to generate two 512*768 picture in good enough quality, but I using easy diffusion, so I can leave when i set up the batch work.
On your excel sheet you showed 22-24 for the high number for both of the samplers you used at the end of the video the ones you said you liked the best but you put 36 when you did the higher image. You said earlier in the video that after 22-24 there is hardly any change so I'm confused why you use 36 to finish the images instead of 22-24?
That was only to test if you were paying attention. :) No, really, I don't remember, there is no special reason. And it is not important anymore ... many of the best checkpoints nowadays are Turbo versions (I use Dreamshaper Turbo a lot, all images at www.deviantart.com/rudybai/gallery are done with it) and they require only 6-8 steps, or Lightning versions even 4.
@@rudyshobbychannel I can't use SDXL version they are too big for my graphics card. My graphics has 6GB but I can only load a 4GB file into it and comfortably be able to render images too. So I don't have a choice but to stick to these.
@@DavidBrownFishing Ah, I see, yes then 1.5 models are your go to ones. Best experiment and judge for yourself to find an optimum between quality and speed. Euler and DPM++will always be safe choices. You might still be able to increase image size by using Kohya Deep Shrink: ua-cam.com/video/E3Ss9-QZ7Cw/v-deo.html
Great video! Is there a way to get a copy of this worksheet? Thanks!
Yes, this the link to my Sampler Comparison GDrive folder:
drive.google.com/drive/folders/1ZDYL-T9ZErTUuA6RdCHBz26LFkHTzFZb?usp=sharing
Thank you!@@rudyshobbychannel
Thanks.
You're welcome
thankyou
This is off topic, but I had to laugh out loud at "girl, age 30", haha!
Very helpful video. Just FYI, Euler is pronounced "oiler" not "yooler." Thanks for putting this together.
Yes, after the Swiss mathematician and physicist,astronomer, geographer, logician and engineer Leonhard Euler. I am guessing Rudy did not study any of those things, or he would have known and/or recognized the name. 😂
Well, I have studied and I've known 'oiler' since I was 16 years old ... it's just that I'm Dutch, and we call him Euler. :)
Hi! thanks for the video, it was very interesting and helpful. could you share the spreadsheet you showed?...hope I'm not being too tactless 😄
I placed a copy in the GDrive folder with workflows:
drive.google.com/drive/folders/1C4hnb__HQB2Pkig9pH7NWxQ05LJYBd7D
Thanks, Rudy, you're a big help
👍🏻
Why are you talking like Captain Kirk?
great video! thanks