hi today we will see a comparison between the two versions of the flux model between the flux Dev version and the flux chel version we see that there are very big differences both in the understanding of the prompt and in the quality of the images beyond that we will see how we can do in painting with flux Dev in addition we will try to create the same character with different facial expressions and we will see how we can reach a result quite similar to this thing as usual we will start at the beginning a small
update regarding the models today we have the ability to download these models half of their original volume thanks to Kaji this thing can make it easier for computers that have a little less Ram in order to work with these models you need to download them and put them in the model / uned folder the other things like the t5x XL and the VA are exactly the same as we saw in the previous video so if you have already installed the previous version of flux you only need the models and of course I will leave in
the description of the video the link to this page and also the link to the previous video for those who want to see how to install all the other things after you have downloaded the models and placed them in the unet folder you can use the original flux workflow and simply select this model all other things remain the same as we saw in the previous video before we start it is important to remember that flux Dev unlike flux Chanel does not allow commercial use and this is something that should be taken into account another thing
to note is that the flux Dev is adjusted to 20 steps compared to the flux chel which is adjusted to four steps in the tests I did I kept all the data the same with the exception of two data the type of model and the number of steps flux Snell with four steps and flux Dev with 20 steps these are the only differences the images we see have the same seed the same prompt and all other things remain the same so let's start with the text for image as you can see there is a very
big difference between the Chanel and the dev throughout this comparison the images of the dev will be on the right and the images of the schel will be on the left here we can see the Dune that is obtained also in the result of the Chanel the result is excellent in my opinion but there is something a little more dramatic and interesting in the image of the dev but I guess it has to do with Personal Taste let's look at the image of the eye all I took was the logo of the pixel easel Channel
and I got this prompt based on the logo and this is the image I got with Schnell exactly the same prompt and Dev and you see that the qualities are completely different a completely different reference also in the atmosphere of the landscape and in my eyes also in the combination of these binary numbers in the picture more interesting throughout the comparison we are making I will also leave the prompt below so that you can read it and compare to the result I will not read the prompts so as not to make the video too long
if you need to just pause and read the prompts and that way you can compare the understanding of the models in relation to the prompts we wrote in this picture you can clearly see the photographic quality we get in Dev compared to chel we have almost identical characters in terms of clothing and in terms of style but in terms of the quality of the image it feels to me that Dev is much more realistic and much closer to photography and this is something I have seen throughout many of the tests I have done also regarding
text I felt that the dev was more accurate in the previous video we tested this prompt a food stall with a pixel easel sign and we saw that Chanel does a good job of referencing the text but comparing the same prompt in the dev version we get the text in a much more logical place and in my opinion there is some kind of atmosphere here as well that feels more right to me note that we still have all kinds of artifacts even in 20 steps under the dev model so you have to pay attention and
check the results this image that came out is is also relatively realistic in chel but if we compare it to the image we now have in dev then there is no doubt that Dev gives us a more photographic more cinematic quality Also regarding the understanding of the fro should be inside a train and in chel we get him as if he is sitting in the station and not really inside the train another interesting comparison to the problem we had last time with the schel the fingers in this picture for example we see that we still
have a problem with the fingers in Dev as well this is a little more correct but it is important to remember that there is no perfect model yet probably in this case working with a different seed or maybe a different number of steps the problem would have gone away in this photo an atmospheric photo for body cream you can see that the dev's photo is much more harmonious with a very small depth of field which gives a very photographic feeling and also quite realistic in my opinion the body cream text which we see on the
jar that came out in Chanel is a mistake because according to our prompt we shouldn't have text on the jar once again it seems that dev's ability to refer to the text is better and also the general direction is much more realistic and photographic here we already get something that takes the eyes even out of focus because in our prompt it says shallow depth of field which is probably a little too much but in terms of the texture and in terms of the qualities here as well there is no doubt that in Dev we get
much much more realistic results I also checked illustration and in my opinion here too the feeling is that Dev gives more depth more richness in Shades if you pay attention here to the area of the fingers we still have some kind of disorder with the fingers here I checked what is happening with a lower resolution this image is 512 * 768 pixels and we still manag to get a very highquality image and of course more realistic schell's result is also good but if we look at the shape of the shoulder and the shape of the
chest then there is something more harmonious and more accurate in the composition in the dev version one more example of resolution before we move on to the image to image comparison in this image that I did my test I was accidentally on a resolution of 1216x 1216 and as you can see the result was very blurry probably with another seed the result would be fine but with the same seed and the same prompt and changing the resolution to 1024x 1024 we get this picture so remember that resolution and the ratio of length and width are
still very significant from the point of view of the model and this is true for both the Chanel and the dev regarding the image to image here too we have very significant differences here too I tried to make the test as accurate as possible exactly the same data the same denoise although the denoise works a little differently with a small number of steps and from what I saw working in image to image with flux Dev there is much more freedom with the Deno you can clearly see that there are very big differences between the images
made in image to image with the Chanel and between what was done with the dev as we saw earlier in general the dev is much more realistic and much closer to photography of course there are photos here and there that I liked more than the result that came out with the Shel but in the vast majority of the results the photos created in flux Dev are on a different level than the results of the chel I also checked the difference between the versions of the model and the most basic in painting for those who don't
know the idea of imp painting is to change only a certain part of the image I will include in the description a link to a video that explains the principle of imp painting and in the next lessons we will dedicate an entire video to imp painting with flux so you should subscribe to the channel and stay updated when I tried in painting with the chel I couldn't get even one result that came close to something that could be worked with no matter what I did on the other hand as soon as I switch switch to
the dev model I suddenly got very good results that fit really well with the other parts of the image as I mentioned before this is a very basic in painting you can see here just a marking of a mask on the face and the addition of the word smile in the prompt and we get our girls smiling and the change seems to fit and seamlessly in this example you can see a relatively extreme example that the model manages to work with very well take a pict picture of a woman with closed eyes and just open
her eyes and get a quite harmonious result and another little experiment I did I thought that because flux is very good at understanding prompts it would be interesting to test a very detailed prompt if only one or two words are changed in it maybe we will be able to get the same character with a different facial expression I used GPT all in all I asked the chat to write me a description of a woman's face face that the description be relatively detailed I defined to Him in Advance that I wanted to change her facial expression
and mood between the different promats and keep the same character and this is the result that the chat gave me this is the format there are two words here that change the first word that describes the character's expression and another word that describes a general atmosphere I very simply copied and pasted these prompts that the chat gave me and as you can see just changing these two words allows you to get a relatively consistent character with different facial expressions and mood once we have the same character with different Expressions it is also possible to animate
the character quite easily in our case I use Luma with two images with different Expressions as a reference there are many tools that allow this today so I hope you learned and we will meet in the next lessons you are more than welcome to ask questions comment and like if you liked it and most importantly have fun bye