ComfyUI 27 Incredible HiRes with Kohya Deep Shrink (free workflows), Stable Diffusion

Video Statistics and Information

Video
Captions Word Cloud
Reddit Comments
Captions
in the previous video we had a look at the selfawareness guidance note which was hidden in coni testing folder in that same testing folder there's another note called coas Deep shrink and that is an incredible note that can create highr images without any upscaling it's a note you might easily have missed but it's really there and it is really wonderful have a look at this image which fills my 4K screen in its entirety and it is was it was rendered in one go there's no upscaling involved just look at the incredible amount of detail that we have here in the city and the skyscrapers and in the clouds this is another example a tusen landscape under uh white cloudy sky and it's just an incredible amount of detail without any upscaling it comes out in one pass straight out of the sampler and this is another example look at the incredible amount of detail that we found find in these images let's have a look how we can accomplish this here is a very basic default comi workflow and I deliberately created a gap here because that's where we are going to place the coas Deep shrink node it can be found in the ad node 4 testing folder that's why you could easily miss it uh it's there the COA deep shrink note and let's add it and it goes here uh between the sampler uh uh between the checkpoint loader and the sampler it sits in the model chain and the big bonus is if you you don't want to use it it can easily be bypassed just hit control Q with this very basic workflow uh the following images were rendered uh first a th pixel image that is the normal image that you get with sdxl checkpoints and on my system that rendered in close to 10 seconds with the Coya Noe enabled and the same th000 pixel image it rendered 15% faster and the image yeah it is a bit different but that is not a problem for me uh it has much more detail I like the way this image comes out so it's a nicer image and it is faster that can in itself already be enough reason to always keep it there and switch it on uh but what can we do we can also make a larger image even with the standard uh dream shaper turbo and probably any other checkpoint you can go beyond the th000 pixels this is 1280 pixels uh it's still a nice image and this is also with Goya enabled 1280 pixels again 15% faster and well personally I I see more detail in this this Coya image so I have no reason to switch it off but let's go even further yeah and now the default uh render checkpoint rendering starts to get into trouble what you often see if you try to render uh too large an image you get these twins or doublings or other problems but with the Coya note nothing is wrong everything is quite okay so let's go on and now we are at 1,800 pixels well the standard method cannot handle that Coya is still going strong and look at the incredible amount of detail that we see in that image and and well let's go now full scale 248 pixels that's almost the vertical screen size of my 4k monitor and uh yeah default that does not work but Coya is still going strong and this is an incredible amount of detail in the in the waves and in the rocks and in the sky it it's just wonderful and there's no upscaling involved this is all that very basic workflow that we just saw uh all right right is there nothing negative about it well yeah it does not work with any kind of image I found that with images if be it photographs or artwork of persons it can faster go wrong uh default nothing is the matter uh but if we grow the number of pixels well this still goes quite okay but this arm is already a bit long and this arm well maybe also already is a bit too long uh the next size we are already at500 pixels that goes wrong with the standard uh method with Coya it might still look okay on first glance but this arm really has a problem maybe with different seats you are lucky and it still goes okay but yeah then you can ask yourself why not just do a standard up scale and if we go a little bit larger again yeah this is too long a person and this also is not quite right here around the hips so things go wrong with certain images it does not work but with a whole lot of other images it really does work for instance with landscape images uh here are a couple of other examples this comes out of dream shaper turbo default at a size of 3070 pixels so that is really quite a large image but we see a lot of trees here at locations where you would not normally expect them and this tree here in the middle is much too large with Coya we get a perfect image at least a plausible scenery this could be a reality and also look at the incredible amount of detail here along the lake sides that that is just incredible and the the mountains they have a lot of detail all right and then we already have a quite a large image but but we can still upscale that and then you can get incredible sizes with Incredible detail and if you shrink them again they still look very very good and that's what I tried to do uh of course the note that we handled in the previous video the self attention guidance I want to include it so in my default workflow I added this Coya deep shrink note uh and I liked to uh combine it with the self attention guide guidence because that gave more pop to almost or all images but apparently they don't like each other I could not find a a working method of combining them it always gave me an error so I thought of this method the first sampler we use the self attentions guidance and then we go with the Coya uh deep shrink note to the second sampler that I gave an upscale scaling factor of 2.1 and if I start with 1,000 pixels now I have 2700 pixels and that gives me a total coverage of my 4K screen and this is the workflow that uh I'm now using this is a image 2160 pixels large uh it is up skilled and uh but in the default method this is the same IM image now with the self attention guidance switched on and yeah what we saw in the previous video it has just more pop this is off this is on that has more pop uh uh more contrast you might like it sometimes maybe not always but this now is the new image the Coya image and it might look uh it's in the details uh let's have switch back and forth between the uh self attach image and the Coya image uh have a look here at these wave these well look okay over here but now I switch look there's a lot more detail over there have a look at the rock this looks almost a bit cartoonish it is not a photo it's a painting so that's okay but now have a look there is a lot more detail also in the tower and also these waves over here at the right side have a look if I switch there's a lot more detail I I particularly like that the translucent and also this um reflection a lot more detail over there these clouds over here uh a lot more detail with the Coya note so it looks more like a photog yeah I I I personally I like it let's place them side by side this is the default on the left uh all these images are 2160 pixels uh this is the self attention guidance that image has a bit more pop which we which we saw in the previous video and this is the comparison with the new Coya note and that really has a lot more detail if you look closely than the image on the left and well now the self attention and Coya compared well the self attention still has a bit more contrast it looks brighter so you might like it sooner but if you look closely this Coya image has just has an incredible amount more detail you can make a choice you can switch them both on and off with control Q so do what you like the workflows are included in the download that goes with the text a landscape um this is a photo in the style of anel Adams a famous uh scenery photographer he already died uh so that's unlucky but yeah he made these style of photographs this is upskill to my 4k monitor uh in the default way this is with selfawareness on and if I switch back and forth you'll see that it has more pop uh but it is actually not really sharp let's now go to the Coya note man this is a crazy amount of detail look over here at the uh banks of the lake and the leaves of these trees over here if I switch back and forth and also the of detail here in the mountains look we go back this yeah suddenly it looks terrible and now the Coya note crazy amount of detail back and Coya well I would always choose the Coya note for photographic work because it looks more like a photograph oh and by the way I also use that very basic workflow so create a a Coya photo in one go and this is what came out so this is a photo without any upscaling straight out of the sampler okay here we have the city with the thunderclouds above it also straight out of the sampler but in the in the old way apparently Dream Ship was able to handle this which is already great in itself but now with the Coya note well that is just a a difference of day and night well on the picture it is night but look at the difference this is default this is with the COA note enabled incredible just a real photo now let's have a look at this same with upsampling um upscaling this is the default way of working with uh latent up skill followed by an image up skill this is with the selfawareness enabled and like always with the self-awareness that image this is the old one this is with sag on that has more pop uh so that can be nice in some cases but now let's have a look at the Coya image this is the Coya image this this is just if you look at these details here in these buildings that is just an incredible amount of detail much more than with this image and I was already happy with these kind of images uh yesterday but today well the Coya note that is incredible here are the jellyfish that we also saw in the self awareness video um this is the default upscaling and this is with the selfawareness on well that yeah that really has more pop this is off and this is on now let's go to the Coya note and man that is an incredible amount of detail look at these these threats and these tentacles the but specifically these threads they are so thin over here they may pop more but here they are much more detailed and also here at the underside there is so much more look this is course and this is fine man that is just incredible oh and just for comparison I also rendered it in one pass with the very basic default workflow and this is what came out no upscaling involved just one pass straight out out of the sampler with the Coya note enabled let's do a portrait this is is a portrait of a lady in the 50s I like how dream shaper renders skin that is so natural not not that that plastic and this is the default way of working this is with the selfawareness on that has a bit more pop off and on uh you may like it or you may not like it but now switch the Coan not on and this is a very subtle difference maybe but uh let's have a look look at the hair uh left and right well maybe you have a look here at the right that looks very finely detailed now I switch back look that is far more coarse the same with the straw hat look at the straw head the Coya note very very fine detail and this is quite coarse look at the eyes let's just zoom in on the eyes uh yeah they are not really bad there are reflections in it but the irises they are a bit yeah not human and also the hairs let's now go back and have a look at the Coya eyes look if I zoom in on the Coya ice they have a lot more yeah human detail and also the hairs are much much nicer uh it's it depends whether you would like this photo more for the pop but this photo looks more like a real photo this still looks a little bit Yeah AI rendered and this starts to look like real photography because of the ultra fine detail also the necklace look this is a just a plain necklace and here it is yeah it is very detailed I really like this Coya note however if you try to render a person uh in one go this is the one pass without upscaling that not always goes right uh we have two phases and also the single phase it is much too stretched uh so persons or people uh that is a bit of a problem with the Coya note anything else uh might pass your criticism this is a uh let's say a more fleshy photo and then the skin yeah is much too uh plastic for me but it is not about the skin it is about how it pops out this is the default way of working this is with uh selfawareness uh switched on yeah that image as usual has a bit more pop as we can see now the Coya on look again at the detail in the hair and in the eyes these eyes look beautiful they are they look like human eyes and the hair is very detailed I will switch back this is well plain eyes and coarser hair so this Coya looks much more like a real picture at least in my eyes and this is what came out uh without any upscaling this is an 2160 pixel image straight out of the simpler and luckily yeah you can be lucky sometimes this was a person but nothing went wrong it came out quite nice this is a Scandinavian girl the default upscaling is used uh it's a reasonable image this is with selfawareness switched on the image has a bit more pop but now look again at her hair and at the eyes when I switch to Coya right now man that has so much more detail and the eyes look so much more natural everything looks more like a real photograph again I also rendered it in one pass and again we were lucky uh besides the the thumbs over there and to much fingers which is quite usual uh everything went well this came out without any upscaling straight out of the sampler another Scandinavian girl this time with more pop more color uh the default way of working and this is with uh self-awareness enabled as usual it has more pop and you may like that off and on uh but now let's switch to Coya and again have a look at her eyes and her the details in her hair this already looks good we think but now look at that detail in the eyes and in the hair that is incredible and of course also the detail in our sweater let's go back this is already a poppy image this is more a photograph I like this Coya uh note and this is what came out in one Past a 2160 pixel without any upscaling well it's just incredible okay let's design a new kitchen this is the default workflow this is with selfawareness uh on and off on and off picture has as usual a bit more pop now we switch to Coya uh look at the detail here in the plant leaves and also the tabletop uh the things that are on it it almost looks a little bit cartoonish right now at least if you compare it we go to Coya 3 2 1 look at the details in those leaves of the plants and also the stuff that is on the countertop and also over there it looks more like a photograph this almost suddenly almost looks cartoonish and this looks yeah more real uh more lively okay this is the image rendered in one go not a bad image at all no upscaling involved in this image one more image the watercolor painting this is the default way of working this is with selfawareness switched on it has more pop uh off and on look at the strange thing that happens we have water and boats here and with self Awareness on WE suddenly have a street with cars yeah that can happen but the image has more pop now we switch to Coya and look at the incredible amount of detail here in the leaves of the trees and also yeah her dress in the car and particularly her head and her face if I switch back you will see the difference this almost suddenly it it looks cartoonish and this looks more real and now finally the image that came out in one go yeah here we have this problem that her arms are merged too long so you need to be lucky not always with persons specifically it goes right but the incredible thing of the Coya note is it creates more detail it creates uh large images in one go no more upscaling needed and it creates more photo realistic stuff I like it a lot thank you for watching maybe see you back in the next video in the meantime of course as always have fun
Info
Channel: Rudy's Hobby Channel
Views: 5,760
Rating: undefined out of 5
Keywords: stable diffusion, comfyui, kohya deep shrink, upscaling, photorealistic
Id: E3Ss9-QZ7Cw
Channel Id: undefined
Length: 21min 27sec (1287 seconds)
Published: Thu Feb 22 2024
Related Videos
Note
Please note that this website is currently a work in progress! Lots of interesting data and statistics to come.