ComfyUI Advanced - 3 Pass Workflow
ฝัง
- เผยแพร่เมื่อ 8 ก.ค. 2024
- This tutorial is all about multi prompting in a 3 pass workflow, it should help you with prompt control, advanced image production and help you with refining and upgrading your end outputs!
Come visit the discord if you are find my tutorials helpful :D
Workflow: drive.google.com/file/d/1EK3_...
My Discord:
/ discord
00:00 - Intro
01:27 - About Sampler Passes
04:21 - 3 pass overview
10:07 - Build It Up
13:42 - Ksampler Build
19:25 - Multisamplers!
20:59 - prompt framework
24:01 - wiring it all together
28:32 - Sampler Settings
33:21 - Multiprompting
39:55 - Prompt Shenannigans
46:54 - Ending message!
Love your tuts and your approach to teaching and creating!
The only ComfyUI instructor who makes me excited for a 47-minute tutorial😂! Can’t wait to dive in. Haven’t watched it yet, but I know I’ll learn something amazing. Thanks for taking the time to share 🤝
yehhh... was going to be 58 but then I cut some fat off it XD. They are usuaully like 20 minutes shorter if I don't build the workflow XD
should split vids maybe, build and use. mmmm nahhh.
Thank you. Best instructor.
Wow, you have a great understanding of ComfyUI. Thanks for making these videos. Neat stuff!!
Really good explanation of what it's all about. thank you
Always look forward to your tutorials! Thank you for the noise gradients tutorial, the results have been really fun!
:D I should put up the upgraded version of that module :P
Woop woop - You are the best! Thank you for this
Thank you for such in depth explanations and examples. Being knew to comfy, I much appreciate it.
Thank you very much for all the explanation videos. You are my main ComfyUI tutor. 90% I know about how to work with it, I know because of you.
:D happy to help. keep up teh good work
Super sick thumbnail. Your work is looking crazy good!
love you Bro !!
Gotta say, I just found you but you have one of the best channels on how to properly utilize ComfyUI. I really enjoy the pacing and explanation of everything, and the way you think about images is clearly different than the standard "type words and watch image come out" kind of setups. Really appreciate you taking the time to make and share these videos and I'm looking forward to what else you put out!
working ona compositing video at the moment although im deep in a rabbit hole of nodes lol.
This is insightful information thanks for taking the time sharing this🎉🎉🎉🎉
Great inspiration! Thanks for the tutorial, it was well explained & thorough. Cheers!
Saved for later! Will have to try this out. Glad to see you back! I was just looking at your other masking video and was wondering when your next video was coming out.
takes a while, I like to really understand a thing before im willing to do a video on it.
awesome video, watched lots of your videos they are all really nice. hope you keep making them, lots of stuff to learn
Got a few good tips from this, especially.the buses. On the first pass SDXL lightning is.great for creating your draft image
This video content is also great
Thanks...
This was very helpful, thank you! I threw a problematic character Lora in there that was trained on a small set of images with the same white studio background in every shot. And whereas in a single pass I cannot get away from that background, with this approach I can make an image first in any setting I want with the Lora only at weak strenght, and then reintroduce the Lora at full strenght in round 2.
:D you might try the noise tutorial I did, you can use these to override your lora's desire to make particular environments :D blood red skies? no problem once you master that.
This is a very interesting and advanced idea, using different models and different prompts in stages in order to arrive at a result that would otherwise be impossible to get at. Perhaps an analogy might be rock climbing, where you employ different techniques to reach different stages/platforms, ultimately getting to the goal that would otherwise be unapproachable with a straight line single technique approach.
Thanks for sharing!
Great tutorial, and loads of fun to try along, thanks.
Just wondering if you've tried the Anything Everywhere nodes instead of having a bus? I like them because it cuts down on the noodles, and also means I don't forget to make connections.
Im aware of it. There's two reasons I don't use it.
1 - I am a visually oriented person so seeing where things go makes it easy for me, plus I don't have to type things in or remember what is going where.
2 - teaching people in a tutorial is best done with low level nodes and simplistic systems if possible, having the noodles visible also makes it easier to follow what i"m doing without me zooming right in on a node so you can see what it is connected to.
Still, i know several people who use the anywhere everywhere nodes and they clearly love them.
me, i find other ways easier.
I really appreciate this new tutorial! I'm going to implement it. Also, are you having "clip missing / unexpected" warnings on console for all workflows (sd 1.5, sdxl, cascade) since last update?
I update usually bi-weekly because of comfyUI's frequent node breaking updates. so im not having this issue.
Question: in one of your earlier videos you went to great lengths to split out the positive prompt to send parts to the text_g and text_l seoerately, have you decided to abandon that idea? Really enjoy your ComfyUI videos, thank you.
I genuinely thought « oh I must make Chris 🇦🇺my Australian friend listen to this because I love the accent » and wanted to know what he thought about this because Chris has a special accent but I just learned that it is so much 😅😅😅😅 the same thing 😂😂😂😂
:D im from melbourne so the accents not as bad as someone from queenland or who builds buildings all day :P
Our countries so young we haven't yet developed regional accents the way britain and other countries have, its getting there though.
Simple smart explanation, very good
I want to ask you what the difference is between upscaling between latent and vae decode, because you said that the upscale with latent is a lossy process. I thought earlier that the vae decode is a lossy comprehensive methode.
And btw creating things like a sniper house we can use conditioning average i think but iam not sure. With combine and time step maybe something like that, buti didn't try.
Appreciate
If any one gets ,( error Error occurred when executing CLIPTextEncodeSDXL: 'g' ) its a problem with your Load cheackpoint model, change it to another cheackpoint and the error should go away.
hello,
the Bus node is broken, now I load my workflow and it won't let me run it.
any suggestions? Thank you very much
Thanks mate from your viewer from indonesia. I hope you can make tutorial about 2 person photo couple and each person have lora face
you would be wanting to use Ipadapter and regional prompting for that kind of a workflow :D
I'll get around to those eventually but they are one of those nodes thats really hard to get right (mostly its getting the right models to work with eachother thats the issue.)
Can we get a video on using this with something like instantID? Would love to know how to add insane levels of detail to my face
im going to do an ipadapter video soonish, faceid is basically an extension of that. its not too hard to set these up, you basically stick them between your samplers and plug model lines through them mostly just like how you do controlnet using conditioning lines.
soonish for me can be a while, hopefully not like a month like the last vid.
Can't wait. Love your videos@@ferniclestix
Seems the MarasIT node broke yesterday March 27 - or perhaps my Python is broken...hmmm....
yeah ive chatted with the creator, apparently it just doesn't want to do certain things? at the moment ive phased it out for comfyrolls 'pipe' which while less flexible does provide some of the same features. ill switch back once its working again.
Hi, the MarasIT node became MaraScott node, so you might need to remove the MarasIT one from your Comfy then add back MaraScott node. the node should not trigger a red alert but might not work, you ll need to replace it with the V2 version the one with a Mara Face icon in front of it - sorry about that
а может лучше управляющие сети использовать, будет сильно быстрой, накидал эскиз максимально простой как дети рисуют.
_Euler_ is not "you-ler" - - it it a term named after Leonhard Euler, a great mathematician, and is "oy-ler" - - many get this wrong.