Thank you for this video. Barely getting into SD and while this is outdated it was super useful. I am an artist (Creative director) so seeing the grids you put together were really useful to see how the values affect the images. Thank you for taking the time.
I learned quite a bit with this video, and have already adjusted my workflow. Actually turned off highres.fix..Went 768*768. Chose a good 'discovery' sampler like DPM ++2m (pumped out an image in 10 seconds)...chose what I liked. Sent the image to img2img. Upscaled using your Heun (You stated fantastic quality), to 1536x1536, each image taking 1 minute. This is producing fantastic results in 1/2 the time as before. You literely increased my workflow 2x!
Is sendingntonto img2img as effective as the hires fix? I worry that doing the step later somehow loses information about the prompt that helps the upscaling
Just letting you know that this video is supper useful for me rn. I just started using SD some weeks ago and this is helping me a lot. I often find myself doing the same thing over and over trying to figure out which sampler is better for which task. Thanks a lot
Great recommendations, thanks. I've mostly been using Euler a and DPM++ 2M Karras so I'll try branching out into some of your other recommendations. Was able to clean up my sampling list at least haha.
This tracks with the little bit of experimenting I have done, mainly with Euler a, DDIM, Heun, and DPM++2M Karras. DDIM, Heun, and DPM++2M Karras sometimes have very similar output, but sometimes they are different.
This is awesome! Would you also be able to turn all these matrices into a CivitAI article that could be quickly scrolled through? It would be great to be able to consult these via a page and scroll through them
Samplers keep getting better and better. Due to the audio quality, I will probably redo this video and include newer samples at some point though I'm focused on ControlNet tutorials in the near future.
Yer doing God's work, son. Bless you.Oh, and because I'm gonna keep watching your vids I gotta straighten you out right now: it's pronounced "OILer" sorry. That was driving ing me nuts! Like, "foilage" "nukyular" or the "Specific Ocean".
Not very deep, just basic information really. You can add ancestral property of re-applied noise on to non-ancestral samplers with new extension, which in my short experience leads to better outputs overall. Part 2 required. Also there are more variations of samplers with older known extension - you can apply properties of Karras, exponential and variance preserving samplers to any of the existing with variable parameters. To see if sampler converges or not it's enough to look at it's name, because ancestral(with "a") samplers don't, they re-introduce noise to image regularly. You can easily see that in your converges/not bubbles with the only outlier of DPM++ SDE, for whatever reason. But hey, at least i enjoyed some memes. I'll sub for now, let's see what you'll do next.
It definitely sounds like you are way beyond the majority of users in terms of experience with samplers. I think for now I'm going to stick to features available in the base Automatic1111 GUI, because my primary goal right now is to create good, concise videos around intermediate and advanced features. Though I'm definitely gonna have to take a second look once I get the low hanging fruit out of the way.
@@siliconthaumaturgy7593 Eh, i might be, but i really don't think so, because it's just extensions that everyone can get and use without any experience. Im not very advanced, considering that people familiar with python coding will have a huge advantage and can make their own features e_e Or, well, if we talk about majority, which includes everyone just using SD, then yes, of course, because majority is just using sites and bots that don't even let you see what sampler is used most of the time. We here are all beyond majority tbf.
Disagree. It’s quite bearable. It’s a tech deep dive. Most of these don’t have the best audio. His is ok. Sure. Could be better. But good content and it’s bearable.
![Blox Fruits Dragon Rework Update [Full Stream]](http://i.ytimg.com/vi/EqDAp8udhm0/mqdefault.jpg)
Your videos are a gold mine, thank you
Thank you for this video. Barely getting into SD and while this is outdated it was super useful. I am an artist (Creative director) so seeing the grids you put together were really useful to see how the values affect the images. Thank you for taking the time.
You are the only one doing these kind of usefull deep dives into SD,thank you for your many helps!
I learned quite a bit with this video, and have already adjusted my workflow. Actually turned off highres.fix..Went 768*768. Chose a good 'discovery' sampler like DPM ++2m (pumped out an image in 10 seconds)...chose what I liked. Sent the image to img2img. Upscaled using your Heun (You stated fantastic quality), to 1536x1536, each image taking 1 minute. This is producing fantastic results in 1/2 the time as before. You literely increased my workflow 2x!
Is sendingntonto img2img as effective as the hires fix? I worry that doing the step later somehow loses information about the prompt that helps the upscaling
Using 'Euler a' as opposed to regular Euler, seems crazy fast & seems to produce very good quality images considering how quick it is.
Same experiences! It rarely messes up the quality and I love using it!
Just letting you know that this video is supper useful for me rn. I just started using SD some weeks ago and this is helping me a lot. I often find myself doing the same thing over and over trying to figure out which sampler is better for which task. Thanks a lot
Thank you for your research, you really helped me choose the samplers!
Great recommendations, thanks. I've mostly been using Euler a and DPM++ 2M Karras so I'll try branching out into some of your other recommendations. Was able to clean up my sampling list at least haha.
This tracks with the little bit of experimenting I have done, mainly with Euler a, DDIM, Heun, and DPM++2M Karras. DDIM, Heun, and DPM++2M Karras sometimes have very similar output, but sometimes they are different.
great video. Lots of examples and a nice and concise tl;dr at the end!
Just what i was looking for, bookmarked this for whenever i can't decide on what to use
Concise and super helpful. A critical understanding for SD
Amazing job man, congrats
This is awesome! Would you also be able to turn all these matrices into a CivitAI article that could be quickly scrolled through? It would be great to be able to consult these via a page and scroll through them
thx for your time and dedication, you make a great service to us !
Thanks a lot for your work! 👍
What is best sampler for text to image ? In your thought ? Can you rank top 3 favourite ,for stable diffusion 2.1
So I guess UniPC is probably really new. I was looking for comparison with this one. I am under the impression that it is giving good result fast.
Samplers keep getting better and better. Due to the audio quality, I will probably redo this video and include newer samples at some point though I'm focused on ControlNet tutorials in the near future.
this is so insanely useful
Yer doing God's work, son. Bless you.Oh, and because I'm gonna keep watching your vids I gotta straighten you out right now: it's pronounced "OILer" sorry. That was driving ing me nuts! Like, "foilage" "nukyular" or the "Specific Ocean".
Haha I'll try to correct myself in the future.
can u add the new samplers "3M" and "restart" to the cheat sheet? :)
Unintentional ASMR
I hope this analysis is also valid for SD v. 2.1, as that is what I'm using.
I got smarter after watching this video.
hiiiiiiiiiiiiiiii plz help me about Gyre i am in much tension help me please! professor?!
What does Karras mean?
It's the name of a particular algorithm used in those samplers
Johanna Karras. 😏
sigma chrun, eta ancestral, eta ddim ?
Not very deep, just basic information really.
You can add ancestral property of re-applied noise on to non-ancestral samplers with new extension, which in my short experience leads to better outputs overall. Part 2 required. Also there are more variations of samplers with older known extension - you can apply properties of Karras, exponential and variance preserving samplers to any of the existing with variable parameters.
To see if sampler converges or not it's enough to look at it's name, because ancestral(with "a") samplers don't, they re-introduce noise to image regularly. You can easily see that in your converges/not bubbles with the only outlier of DPM++ SDE, for whatever reason.
But hey, at least i enjoyed some memes. I'll sub for now, let's see what you'll do next.
Link to this extension please?
It definitely sounds like you are way beyond the majority of users in terms of experience with samplers. I think for now I'm going to stick to features available in the base Automatic1111 GUI, because my primary goal right now is to create good, concise videos around intermediate and advanced features. Though I'm definitely gonna have to take a second look once I get the low hanging fruit out of the way.
@@siliconthaumaturgy7593 Eh, i might be, but i really don't think so, because it's just extensions that everyone can get and use without any experience. Im not very advanced, considering that people familiar with python coding will have a huge advantage and can make their own features e_e
Or, well, if we talk about majority, which includes everyone just using SD, then yes, of course, because majority is just using sites and bots that don't even let you see what sampler is used most of the time. We here are all beyond majority tbf.
your audio quality is unbearable
In what way? He's got a bit of room reverb but that's not really an issue for this, maybe it's your system?
Sounds good enough to me. Not every video creator has a high dollar, tricked out video studio. Nor do all of them need such a studio.
Disagree. It’s quite bearable. It’s a tech deep dive. Most of these don’t have the best audio. His is ok. Sure. Could be better. But good content and it’s bearable.
Not everyone needs a 10k audio setup. He has 171 subscribers and is just starting. Give the man a break from the internet rage JFC.
The sound quality is good, you can understand everything without effort, checkout your hardware