Comfy UI K sampler Explained | How AI Image generation works | Simple explanation
ฝัง
- เผยแพร่เมื่อ 1 ต.ค. 2024
- This is my attempt to try and explain how Ksamplers in comfy UI work, while also explaining a VERY simplified explanation of how Stable diffusion and Image generation works.
written tutorial: endangeredai.c...
If you want to support the channel please do so at
Patreon: / endangeredai
One of the few channels actually breaking down the nodes. I hope you make more of these.
Thank you! I’m just trying to mirror my learning process
@@EndangeredAIthis is the way. Just do your learning and once you think you have a decent grasp, do a 5 to 10 minute explainer, and don’t worry too much about marketing. The likes and comments will generate views over time and before you know it, you’ll be like that Australian guy who does the best blender tutorials ever, with the donut gimmick, and I hope you can appreciate that your audience will always be a bit limited, you’re unlikely to become the next Pewdiepie, lol.
Looking forward to your channel’s ongoing growth and proliferation.
That is most well explaining video about the KSampler node I have ever seen! Thank you for giving an understanding of the nodes!
Glad it was helpful!
Thanks, I've strung samplers for upscaling, but is there any benefit in a basic workflow of using 2 samplers at 15 steps & next on 5 versus one sampler on 20 steps ?
Just wondered
There is if you want to change the model you use on the second sampler, such as on sdxl using base and refiner. You could do things like using a realism model and a cartoon one for the last 5 steps to get something with a more artistic touch
I have a LEGION 5 laptop, Ryzen AMD, NVIDIA GEFORCE GTX 1660 6gb graphics card, do you think it can work? I ask because I have tried several tutorials and they all give an error in KSAMPLER or UNSAMPLER. Do you have any help for my case?
I have tried to change video styles and it gives an error in KSAMPLER ADVANCE
Thanks for the details, please make more videos like this. I really helps to understand the concepts.
Glad it was helpful!
Does this mean you can use KSampler to actually denoise a noisy photo? Like: Input a noisy image (non ai-generated) and tell it to finish the last 4 steps out of 20?
Technically yes. I’ve never actually tried it. You could also inject a small amount of noise into the image , and then demonise that so that the sampler has enough to work with,
Thanks to great vid! Subscribed to Patreon!
Thank you! Appreciate the love! ❤️
Very good explanation, it helped me a lot!
Glad it helped!
Can't hear you.
Thanks for the explanation!
By the way - EULER is not pronounced "juler" but rather "OILER".
Leonhard was a German speaking Swiss: en.wikipedia.org/wiki/Leonhard_Euler
Aha! I shall try and improve my prononciation moving forwards!
you are amazing
but :
You are presenting a podcast on a video that contains an image that does not move
Yes, this is from early on in creating content! I’m trying to improve on it! Thanks for the input though! I’m trying to learn from network chuck!
Brilliant explanation!
Please more!
Could you talk about CFG next?
Sure! That’s a great idea! I’ll add it to my pipeline!
Could you tell me more about what you want to know beyond cfg is how much creative freedom the ai has?
About the quiet volume: sorry if you already know all this, I hope it doesn't sound patronising, but I thought I might try to help. I looked at the audio file from the video and it kind of looked like it just hadn't been normalised, but I don't know if it's different before and after TH-cam processes it. [More below]
If that's not the problem, I noticed at 0:32 there's a click sound, and noises like those tend to be picked up very strongly by microphones as the loudest thing they hear; so it may be that the voice audio is being picked up as a relatively much quieter sound than any incidental clicks etc. If that might be the issue, I usually manually lower any volume spikes while editing so that the voice becomes the loudest thing in the audio file, but applying powerful audio compression to the file might be another option.
In the past I've tried to upload stuff normalised to supposedly ideal LUFS (about -14) for TH-cam, but I found that it made no apparent difference, so I went back to normalising to -0.2 dB for simplicity. I don't really understand LUFS though, so maybe I'm missing the point there.
The video's very insightful, thankyou. 😊
Yes, I think I have solved the audio issue. Premiere’s new beta has some very good audio tools that have simplified the process! I wish TH-cam would let me re upload this video without having to create a new entry. 😩, and I can’t upload a revised version without risking being flagged for duplicate content
@@EndangeredAI IMHO the audio is fine.
@@OutOfWards it's really not fine, but at least the creator is aware of it and I'm hoping it's fixed in the next round. I can barely hear the narration on max speakers (granted, I don't have the loudest speakers, but they can usually fill a room), so have to use headphones to make it clear enough. Appreciate the explanation very much though creator. Thanks for making the video.
Hi nice video, but your voice is a little soft...
Sorry about that, haha this video is from before I sorted out my sound 😅
does keeping my steps upto 200 steps make sense?
Oh goodness no it does not. At most you shouldn’t need more than 50 and only for certain samplers. I usually work with 20-40 including refiner steps
I've been using a1111 and invokeai for several weeks, watching videos and reading blogs.You've taught me more about how sd works in your first three videos than any other sources. You are hitting the right balance of "Do this to make that" and how sd works to turn an idea into an image.
I'm done with sd and invokeai. There are simply too many balls in the air. The generation process is far from intuitive and it's too complex to represent in a traditonal, browser-based UI using menus, tabs, and panels.
Serious process management in industry and finance use workflows for good reasons.
that was very informative ! thanks. Can you do video explaing the latent image and what role does it play in image generation? Thanks
@jalpranal, that’s a great idea! I’ll add it to the list!
I have a question that I can't seem to find the answer to online so maybe you'd know. I'm testing out using ksamplers at different intervals. When I use a setup with 3 ksamplers (refiner for 2 steps, then main, then refiner again at the end) it seems to produce decent results. But when I split the main ksampler into 2 ksamplers (both with approprate start and end steps, no noise added, noise carried over, etc) it produces a different result than the 3 ksampler setup. Shouldn't the results be consistent no matter how many ksamplers there are, given that all of the external parameters are the same?
Is your seed the same when splitting the ksampler?
Brilliant explanation!
Please more!
Could you talk about CFG next?
Researching it now :) video will come soon
Damn that was good
Thank you! I’m so glad people are finding it helpful!
Hey! ... great explanation. Thank you for taking the time R&Ding this! Finally I understand the Latent part of this node!... Thanks!
... now I have to markbook this video and watch it 300 times until is burned on my head!
😂😂 glad it was helpful! Yes it was hard finding clean explanations for most of this!
very good!!!
Thanks!!
can't hear anything
I’m just going to risk TH-cam and re upload a higher volume version of this
a death risk@@EndangeredAI
thanks for explaining this so clearly!
Glad it was helpful!
I'd love it if you dived even deeper. What does karras do? What's the difference between euler and DPM. Or ddim. Or dpm and dpm 2m? 3m? ++ Sde? Lol.
I also really appreciate you deep diving into a node. I will absolutely watch as many node deep dives as you offer to create :)
Thanks! I have a video planned to compare the different sampling methods! Stay tuned!
@@EndangeredAI I've done some xyz plotting with different models to find the right settings and quality. It was really nice to learn what normal and exponential means :)
Exactly what I was looking for. Thank you!
Great to hear!
Great content! I have a side note question. How does the negative work? In automatic1111 the negative prompt REALLY has an impact while in ComfyUI it has little to NO effect at all. Am I missing something?
It should work the same in comfy, can yo double check it’s been wired in correctly?
@@EndangeredAI can yoy give me a sample negative prompt or a workflow that has it?
Such a great video just what I was looking for thanks
Glad it was helpful!
Very nice explanation, thank you.
Glad it was helpful!
Pls make a video on how to create video animation
It’s in the pipeline :)
Thank u for explanation. Subbed🫶🏾
Thank you! Appreciate it!
Thanks a lot mate ! Can't imagine how you've blown my mind
Glad it was helpful!
Great explanation. Can you do something to increase the sound of your voice in the future vids? They are very very quiet.
Oh no not again 😩, they sound fine on my computer but people have complained about this before. I need to figure out the levels to edit at
Sounds good on my iPhone. Just fyi
@@EndangeredAI interesting. I’ll try maybe in browser or on pc to see if there’s difference
Mic volume is really low
I know, I’m sorry about that. I’ve fixed my audio issue in future videos, however TH-cam won’t let me replace this video without creating a new one 😰. I’m planning to add captions to it to help
Great content
Thank you!
First time I understand the steps lol
Glad it was helpful!