ComfyUI Advanced Understanding Part 3

[@latentvision]

It takes a lot more time to make the "basics" videos compared to my standard content so I'd really appreciate your feedback. Let me know if you think they are too long, too short, too advanced, ... and if you'd want more 😁

[@LouisGedo]

👋

[@enthuesd]

Thank you Mateo Good to see how the upscaling works. Appreciate

[@christianblinde]

It is great as always. And it cannot be too detailed. Keep creating this great stuff. I guess many people would be interested in animation and videostuff. Especially consitency.

[@RademenesVG]

👍more, more...

[@appu3765]

Definitely more...

[@ArrowKnow]

the videos can never be to long as they are full of useful info like "double click the dot to create a primitive!" Learning about how everything works in great detail is the best way to understand how to use the tool. Thank you so much!

[@chrisfreilich]

Playing a new one of your videos is like opening up a present on Christmas! Thank you so much for all you do!

[@Skyn3tD1dN0th1ngWr0ng]

I've been playing around the injection node testing multiple checkpoints with and without lora and I gotta say, this is in the same league as IpAdapter (excluding mad scientist node that can be combined with this method for even more control), is extremely powerful to fine tune generations and single-out issues like bad training of the models, over training, over representation of style ethnicity, etc. Makes some distillation outputs almost indistinguishable across the board, can make some CKP act as not intended when fine tuning. This is dangerous, sir... For companies with strict copyright ideas... You are hero, sir

[@Billybuckets]

Excellent show of integrating noise injection into a non-dedicated workflow. To anyone with a spinning head at the tile part: Matteo is showing us how tiled up scalers work under the hood, not suggesting you do it this way. At the end, he suggests upscale packages that utilize this method that streamline it greatly. But his example here teaches you how those work.

[@latentvision]

exactly right

[@ericruffy2124]

basic ha? this series might be the most detailed explanation for the SD via Comfy if you've watched the episode 1 and 2 you know my meaning, that's what a thought-to-be-an-advancer say "well I know that" to be shocked at the beginning. Believe me guys this series is really good to face on deep of the SD via Comfy, and thanks Matteo again for the contribution to the community, it's going deeper and with easy understanding, looking forward to episode 4, you're the best, Thanks Matteo~

[@jccluaviz]

Thank you, Dr.Mateo. There is nothing that gives more peace to my soul than your homilies. I am sure that you are an apostle sent by the dark gods of latent space to instruct us and prepare us for what is to come. Someone who is capable of not only creating but also perfecting and turning an idea from a few words into a work of art and poetry, cannot be classified by anything less than... a master. Thank you very much for sharing your wisdom with all of us. My most sincere thanks.

[@latentvision]

thanks... I guess... should I be worried?!

[@jccluaviz]

@@latentvision 🤣🤣🤣 Don't worry, the latent space will free us from all evil. Now seriously. Thank you Mateo for sharing your experience with us. Your videos help people like me, without much knowledge, to understand clearly how this tool works that seems so wonderful to me and that allows me to be a little more creative.

[@ZIOJONES]

Matteo, your tutorials are EXCITING. I've watched dozens of videos on ControlNet and Tiling, but I never fully understood them until I saw your video. I can't wait to see the one about Detailers. Thank you for everything you do.

[@dck7048]

You really do an incredible service. It's a good thing to point people to "packaged" solutions like tiled upscale nodes, but the crux in making good use of this tech is the fine control, and til now I had never seen a broken down tiled upscale workflow in which you can control individual pieces with this much detail. Bravissimo.

[@hiddenbesidesone]

Thanks

[@maxehrlich]

I appreciate you taking the time to make these longer form lectures to give us a better understanding of what it going to be on "under the hood". I watch them several times to take it all in and go back to them frequently to review how I build my own workflows. 😊

[@alecubudulecu]

Thank you for this!!! We love and appreciate the work you do with these videos

[@fungus98]

This is great stuff! I meticulously went through this entire tutorial and I am slowly starting to get it. I can't wait for your detailer tutorial

[@madmushroom8639]

Love the advanced videos! Thank you so much for putting this together! Video length is perfect for me.

[@step1420]

Nice!! Going to start the vid but just wanted to tell you I really appreciate your efforts! They're perfect btw. I do desire to learn in-depth so Just stick with whatever length you feel is good and you can do.

[@simohee391]

Thank you Matteo, all your videos and work on SD is fantastic! The longer the videos are the better imho and your explanations are very clear. Some concepts are harder to understand than others but with a second watch and some tries by myself on comfy I’m able to understand them all.

[@ptnpnx-rvkinh]

This is the most comprehensive guide to the subject so far! That guy describes almost every single notable aspect, that I'd miss by my own. I'd say it's underrayted AF! Gonna subscribe

[@swannschilling474]

Thanks so much! There are not a lot of channels where I sponge up so much condensed information and enjoy the way it is presented so much!! This is great!! 😊

[@murphylanga]

Thank you for your videos, it's always a pleasure to learn something new.

[@S.Korolev]

Thank you! Watching your videos after having spent a few months with comfy is like going through a grammar book after having had a thousand hours of input. Same hilarious feeling! Also your accent is cool! I've spent some time around one Italian dude in India, so watching your videos brings back those fun memories

[@latentvision]

well thanks 😅I guess. Glad I brought you hilarity and fun 😛

[@eslamafifi1020]

Hi there, Your tutorials are amazing! I’ve already learned a lot from them. Even though they are a bit too advanced for me right now, I love them. Please keep up the great work and share more! I’m currently taking some other beginner tutorials to catch up with you, but what KZbin really needs are some more advanced ones like yours. Thank you so much!

[@milesjohnson9003]

Not to long, the level of detail is great, takes all the concepts and practices that people gather from all over the place and brings it all together using core concepts. SUPER helpful for the understanding portion of the learning process

[@steveyy3567]

bravo,this is what i am looking for these weeks about cn tile model and upscaling, and you help me get through the tiling, thanks! and looking forward to your videos in the detailer nodes!salute!

[@papabravo21]

Thank you SO much for continuing this series. Very much what the new user community at large needs!

[@gentran1991]

Such excellent tutorial. Must be one of the best I have watched so far because you explained the reason and intuition behind your steps as well, not just showing them. Thanks for making this tutorial series 🙏

[@caseyj789456]

Merci Mateo. Great "basic" wow ! 15:27 : upscale

[@MrBenjaminRhoades]

As always, your videos are amazingly informative and always get my creative juices flowing. Thanks for making these!

[@AIKnowledge2Go]

Hey Mateo, your ComfyUI tutorials are amazing. As a fellow KZbinr, I tip my hat to you. Most KZbinrs only cover the basics of ComfyUI, but you really dive deep into the details. One little critique about your pacing if i may. Sometimes you dive deep real fast and quckly built complex workflows where one has to pause, rewind, play, pause, rewind to follow your steps.

[@latentvision]

you blink... you die! 😄 It's hard to strike a balance. This is already a 35m video and the average people watch is like 5-7m so I tend to move fast. But I get the "go slow!" thing often. I'm thinking of doing a relaxed series maybe in streaming. Cheers!

[@lukeovermind]

@@latentvision Somebody mentioned slowing down the playback speed in one of your previous videos, thought it silly but tried it and it actually works a charm lol!

[@adriands8207]

This is amazing, thank you so much. We need more of this series, please!

[@JimmyGhelani777]

Thanks!

[@niknah]

For tidying up the workflow, you can get the quick-connections custom node in comfy manager and it'll auto tidy things up for you if you leave some room between the nodes.

[@santiagocondemorales5567]

you dont know how much i need this kind of tutorials to understand things thank you

[@alexgrigor7290]

This is amazing! Please continue! The more advanced and detailed the better. You’re doing god’s work:))) Can’t wait for the detailer video

[@ceegeevibes1335]

always the best tutorials, the star in the sky

[@runebinder]

Watched all 3 parts this week, very informative and learned a lot, thank you :)

[@TriNguyenKV]

when it comes to ComfyUi, your videos are the most educational!!! Thank you so much for spending time making video like this.

[@johnconnor7778]

Oh my god, your channel is what I was looking for. You explain very deeply and clearly, thank you so much!

[@ysy69]

Appreciate this deeper dive into the basic concepts. Really helpful!

[@shape6093]

Thanks for helping me understand this. Your videos on SD are the best on youtube.

[@thuthao9402]

You are a very thoughtful and wonderful teacher! Thank you so much for your detailed foundational tutorials! I really appreciate the time and effort you've put into them 🙏🙏🙏

[@JanKrummrey]

Thanks a lot for all your work. Great to learn from an expert.

[@RokSlana]

Seeing how these work flows are created and each node being accompanied by some explanation is superb albeit a bit overwhelming... There is so much new here. I'll have to go over some of this stuff and recreate it to get a more intuitive feel for this stuff.

[@delizade]

I watched like a cow observing a master farmer making cheese. I watched like a dog observing Michelangelo carve a marble block as if it were a silk dress. I didn’t understand much, but I’m thankful to have a brain developed enough to at least respect it... A fantastic instructor. All in all, a fully loaded engineer, as I've seen around here. Thank you for all your efforts. I bow with respect, saying Moo! and Woof! Woof!... For years, I had trained myself to create materials for 3D visualization. I had overcome all the complexities of node systems and started producing, only for AI to render it all meaningless. Now, everything starts from scratch... Damn.

[@latentvision]

Moooooooo!

[@bustanilyasir753]

Amazing video. I love the way your explanation mattew. I need more basic and fundamental video.

[@AnNguyen-pd2xi]

Great video. These videos explaining the fundamentals of SD and AI tools have taken my understanding to the next level

[@MyWhyAI]

Thanks Matteo! Great content as always!

[@ariescordova846]

It's wonderfull how you can transform someting complex in a very easy content to learn. Thanks a lot!! And obviously, we want another video of ZERO 2 HERO. Muchas gracias por el contenido que traes! Se agradece muchísimo, espero haber escrito correctamente en inglés. Tus videos son geniales y tu trabajo al organizar todo en cuanto a lo que enseñas es aún mas asombroso. Gracias de verdad!

[@lemda500nm8]

I really like your videos, it really helps to know more about the basic of image generation resulting better image generation. Looking forward to learn from your next video.

[@alexijohansen]

Thank you. Looking forward to the next one. Do you know what’s up with SDXL in-painting?

[@JuliusZlioba]

Your videos are amazing! So much useful information.

[@Guus]

So good!!! I'd love for one of these next parts to cover some video stuff!!! Big thanks for making this, it really helps me

[@latentvision]

yeah sooner or later we'll talk about video too, but it's such a fast moving tech

[@dflfd]

thank you, this is gold! great balance between speed and content.

[@1lllllllll1]

You share soooo much great info. What would be great, if you slice these videos up a bit more, to keep them focused on ONE TOPIC at a time. I noticed you cover several related topics in a single video. All I’m saying is, maybe slice them up as separate videos.

[@Amithyst-MAL]

Thanks Matteo! Your videos are super helpful. I am experimenting with SDXL by adding noise to the Ksampler Advanced, but it's acting crazy! Looking forward to Advanced Understanding Part 4.

[@latentvision]

depends on the checkpoint. Turbo/lightning/hyper need special attention because they have so few steps. You may also need to change the scheduler

[@Amithyst-MAL]

@@latentvision changing the schedule to Stocastic dpmpp_sde helped. Thank you! I’m working through the tile upscale now.

[@Amithyst-MAL]

I got way better results from using the Depth Anything V2 with SDXL. It got a bit tricky at the end to calculate the feathering with a 3:4 portrait ratio, but some trial and error got me through! Thanks again for making this.

[@latentvision]

@@Amithyst-MAL I updated the Essentials with an untile node that takes care of everything. that might help

[@nowhereman22]

Best Comfi tutorial ever. Thanks man! Please keep going - you vids are awesome. And your product too!

[@chillingisliving]

12:32 made me laugh haha, great video as always. It's funny, a lot of this stuff I already know how to use and do naturally after using AI for so long - but you're putting the knowledge of why things work for me.

[@banzai316]

Thanks Mateo, guys like you made me go from newbie to start coding my own node. Thanks a million 👍🔥

[@wildinsights5033]

please matteo keep posting basics video, pretty please!

[@joshlbiomechanic]

ComfyUI is the most Frankenstein machine to build ever! I love this interface so much, reminds me of Analogue patching or Max/MSP Snakes!

[@astronautpolar596]

I'm new to this, you are the only one that makes me understand what's going on. Thank you so much ❤

[@comfyuiadrian]

Thank you so much Matteo! for your hard work and explanation.

[@chornsokun]

2:12 blow my mind. I didn't know you can do that!

[@piezoelectric627]

I love these types of videos, I just meet your channel and I love it.

[@0A01amir]

Always learning new stuff from you and using your nods like everyday

[@luman1109]

Italian Chad making spaghetti. Just kidding, I'm no puritan when it comes to nodes. Thanks for your effort, the noise injection+tile workflow is very nice

[@CrlsSrz]

Great video and explanation. Keep making this kind of videos, please. Thank you very much for your time and effort.

[@Bosnianpowerhouse]

This is amazing thank you for this, is there a way to generate images keeping the same face? would love a detailed advanced understanding on that subject too.

[@unmutualchap6753]

Amazing job, thank you mate !

[@BuckleyandAugustin]

This was insane. Nice job man! And thank you. I learned a ton.

[@omarei]

You spinelli my head right round

[@latentvision]

lol 🤣

[@WhySoBroke]

BIG Braaaavooooo badaaaaassss!!! Maestro Latente with his eternal magical wisdom!! Much grateful, as always and will rewatch many times until the next one amigo!!! ❤️🇲🇽❤️

[@JimmyGhelani777]

I love your content as always. I always learn a lot. I think I have learned more from a few of your videos than I have watching a multitude of others. These videos are great and I for sure would appreciate more of these if you are able :) I can imagine how much time it would take so I appreciate it for sure. might I also make a request to one day make a series on the backed or programming aspect of comfy hehe :) I would love that ❤❤❤ 🙏

[@YING180]

The video is really what i need, thank you so much and looking forward for more videos can go basic.

[@gokayfem]

LETS GO MATTEO!

[@Kuraikari]

Good stuff. Really love the details. Thanks!

[@mixlee9965]

your video is awesome! As a beginner, it really helps me a lot! Thanks Matto!

[@jobigoud]

Question, at 2:15, when we send a batch of 4 images to the sampler, with a fixed seed, where is the variability coming from? I would have expected all the outputs to be the same.

[@latentvision]

the seed kickstarts the random number generation. it doesn't mean that the numbers are fixed, just that we changed sequence of random numbers

[@danieldilly]

How did you get the number 192 @ 28:05? If the image is currently 1024x1024 with a 25% overlap wouldn't it be 256 and not 192? Obviously 192 is right but I just don't know how you calculated that.

[@danieldilly]

1024 * 0.25 = 256 pixels. 256 / 2 = 128. 128 * 1.5 = 192 .. I think thats it

[@latentvision]

upgrade the essentials, there's a the "overlay" output to the tile node... and also an untile node

[@Gmorks]

@@latentvision I'm also intrigued how did you get 192... I'm trying to upscale a 960x640 image and I'm going insane trying to get the correct size

[@latentvision]

@@Gmorks like I said, use the new Tile node (just update the essentials). You'll get an "overlap_x/y" output you can use

[@jurgenmayr8164]

Danke!

[@latentvision]

Bitte!

[@jurgenmayr8164]

@@latentvision Sag mal, kann man dich eigentlich auch für einen Workflow beauftragen?

[@latentvision]

technically yes, but I'm always super busy

[@MrDinamate]

please keep making those! they are great!

[@Quantumspace23]

I cannot find the ControlNet version of canny anywhere. Do you have a pointer or recommended alternative? I'll also need a pointer to the Depth model.

[@AbsolutelyForward]

This is really interesting. So many little new details and insights. I think I'll have to watch it again later and make notes on the individual steps :) Would the same workflow look similar with SDXL, just a bit more reduced (because you might reach the desired level of detail earlier)?

[@latentvision]

yes it looks more or less the same with SDXL but you need to play with the controlnet strength because SDXL doesn't like strong conditioning. But the concept is the same. It's a bit more complicated with fast models (turbo, lightning, ...) because you have too few steps to converge if you add noise, but it's doable.

[@lukeovermind]

@@latentvision This is good to hear, because I use those models alot. Some Hyper models go up at least 10 steps so maybe use those? I guess itss just playing around with the start and end steps to get a sweet spot?

[@latentvision]

@@lukeovermind yeah I would say 10 steps is good. play with samplers and schedulers and start/end step

[@romanbulatnikov7157]

Just amazing

[@prabhuprasad2669]

Very helpful.love uour explanation.i understood everything except the part where you did (a-192)*2 as the final image width. How did you get that number 192 ???it would be great if you could also demistify it😅.

[@Grunacho]

Thank you very much for the informative video. As always, great explanation!

[@OP_Myrmidon]

I prefer the tile upscale method here as opposed to the Ultimate Upscaler. The control and customization is unparalleled. My only question is, when computing the math, you ran the numbers in the simple math node for 512 x 512. I'm having a challenging time calculating how it would work with 512 x 768 dimensions. Any advice?

[@latentvision]

update the essentials! The tile node outputs all the numbers you need. Then image batch together the tiles and use the Untile node to compose them back together.

[@OP_Myrmidon]

@@latentvision ah thank you for that. I'm close to getting it perfect. math is not my strong suit and I have the overlap_x = 192 and I think overlap y = 288. seems to work well and any leftover seams are fixed with another ksampler pass. ty man, you are an amazing instructor. best videos on YT I have ever found

[@morozig]

Hi! I think this video is mostly about stable diffusion and half of it can be done in any other web UI. Please make a video about ComfyUI without stable diffusion. For example batch processing of directory of images, color correction, resizing, auto retouching of photos. I'm curious on how do you find nodes that you need for the task? How to programm simple nodes yourself? How to save result in a target directory? I think that's the main appeal of ComfyUI - general pipelines.

[@codescv]

2:20 I wasn't able to reproduce this consistent result. Could you explain how the "initial latent image" is used and how the noise is added to the initial image? And could you share the "noise christmas decorations red.png"?

[@codescv]

I was only able to get more consistent result by setting the start step to 1-5.

[@latentvision]

@@codescv the image needs to be somewhat compatible with what you are prompting. be sure not to use too small of a conditioning image. small elements are basically ignored. also try to change the checkpoint, some are overtrained and don't follow gladly what you ask

[@tracingzimpossible]

awesome tut .. i think its the best place to learn SD... here but can we have a 4th module on detailer as mentioned at the end?

[@1982manga]

As always, your work is excellent. Could you include a recursive image in your selection that features spaces, architecture, or interiors? This would be important to demonstrate the capabilities of AI applied to more complex and diverse subjects and to assist professionals who use AI in their work and explore these themes. THX!

[@silas-dd5ll]

Thank you, best ComfyUI channel!

[@HTRO-EG]

Finally a new video ❤

[@marlus]

Another great video !!! One question about `Repeat Latent Batch` and `Batch Size`: are they part of the same seed? Why they have variations?

[@marlus]

Second question: why IPAdater refers to model and Controlnet to Positive and Negative CLIP?

[@DarioToledo]

Man, stable diffusion is huge. I have but one question: at 15:40 why upscaling the decoded image and re-encode it for the following ksampler instead of sending the latent directly?

[@latentvision]

I'm trying to get more details, if I take the image directly from the latent it would still be 512x512, some defects would be fixed but I want more pixels to work with. You can technically upscale the latent directly with NNLatentUpscale but it's usable only at high denoise and with low scale factor (usually between 1.5 and 2)

[@DarioToledo]

@@latentvision you need a high denoise because usually upscaling the latent results in a blotchy image?

[@latentvision]

@@DarioToledo exactly

[@bgtubber]

Spectacular once again! One of my main gripes with Ultimate SD Upscaler is that you can't prompt each tile individually so you've got to be more general with the prompt, which in turn doesn't produce the best and most detailed results. Prompting each tile separately seems to be better quality-wise, but as a downside it's a bit more complicated to set up. Is there a way to make masks with separate prompts assigned to each of them, so that the model doesn't confuse one object/surface for another? This way one could still use Ultimate SD Upscaler and it will get some help via the masks to create the correct object/surface. Is there such a thing in ComfyUI?

[@latentvision]

I'm thinking of automating part of the process but still maintaining full control, need to thing about it

[@mickmack8026]

Awesome work! One question: You always seem to use "dpmpp_2m_sde", so what's the differente to 3m_sde and the cpu variants?

[@latentvision]

they are just different samplers, some are better than others in certain scenarios. the GPU versions simply use the gpu for the random number generator

[@THbeto8a]

I seem to be missing a couple of nodes from the ComfyUI Essentials: ImageColorMatch+ and ImageTile+, does anybody knows a way to fix this or replace them with other similar ones? thanks

[@latentvision]

you just need to upgrade the essentials

[@THbeto8a]

@@latentvision thanks for the advice. couldn't update either, ended up deleting the whole essentials folder and installing it again, that solve it for me. thank you for your great work

[@PixelArt_YW]

Thank you, I learned a lot here.

[@noelbachinimiliti1760]

Wow, this is amazing thanks !!!! I have a question, as I’m working on a workflow for look dev character. I want to apply loras but I’m not sure when is better to do it. Do you recommend to do it before any upscaling process? Or better to do it afters upscale. As far I know loras work better when image has bigger resolution, so it has more pixels to work on. Thank you so much

[@latentvision]

depends on the lora and the effect you want. generally speaking I would start adding the lora at the beginning

[@noelbachinimiliti1760]

@@latentvision Thank you so much, have a good day

[@proxi6904]

What's the advantage of using this over ultimate upscaler? Is it the fact that you can have specific prompts for each tile?

[@latentvision]

I'm trying show you what the "upscalers" are actually doing behind the scene. As I say in the video I don't suggest actually using this workflow, but technically you can build your very own upscaler out of this concept

[@proxi6904]

@@latentvisionmakes sense, thank you for all your hard work !