Comfyui Advanced Tutorial 05 - Depth Map

0:00 Welcome to episode 5. Today's discussion is depth map. The main reason to use this is getting reference from an image on the net and making a version or a copy without getting copyright. I will give you three workflows. After that, I will show you the results for each one of them. First, we have load reference image here. The image which we will get depth from. After that, we have a group called depth. It includes depth anything pre-processor. If you don't have depth v2, it will be downloaded automatically about one or two gigs. And the depth laura, you must download this. It doesn't work without this depth laura. I will give all the links.

0:46 Optional Laura load section. You can put your own loras. It is for sampler one. Do not forget to put your keywords. After that, we have Laura load again for pass two. I recommend same loras as pass one. The rest is you know from first episodes the high-risisk fix which always has to be in all our workflows. But one parameter is now more critical to be explained here. I call it the bridge which has two options the steps and d noiseis. It adjusts between consistency and creativity. Close to value one makes creative close to zero makes consistent. I chose the 0.65 65D noise and 30 steps. I suggest you should read the notes. Now, let me show you one example of this

1:33 workflow. First, I load this image. A classroom with students and after that I asked Gemini to make me prompt as if this image was made by Lego style. Now, results. It is impressive. Without changing the position of anything, we just changed the style. Of course, there are errors like text, details, etc. Because of this, it shows us a need. and inpainting workflow will be our topic later but overall it is not bad. Now time to get into the second workflow which is absolutely same just one little difference. Now this second workflow's only difference is control net although it looks small difference there is significant change on images it makes more consistent details but lacks creativity but do not try to use at

2:20 strength value 1. I suggest 025 or 050. The value of one completely broken and similar image to original. You should use this one. Now an example with controlNet plus the same style and prompt is describing the original image. Here we use and try to replicate the original image. You see it is very similar but one of them is AI. Although I have not upscaled this one yet. It is too good. Before we get into the third workflow, I want to show you the output of it. Since we can't use depth with Sigma vision, there is a need to make realistic human images with the position we want to replicate. I made a custom workflow for human images. With Laura's and detailer, we get this result. Now,

3:06 let's see how we can get this quality. I will only tell the extras different from the other workflows. First difference is we have extra Laura load for human. I will leave the notes you should read carefully. We have special luras for skin detailer. On the right side of the workflow, we have two detailers and math expressions which adjust the tile size according to the input. We have upscaler flux dev and sigma vision upscaler. I prefer the sigma vision but if you want like more creativity you can use flux. These are detailers. You should read the notes I have given you on the workflow. I made the basiculer steps set to two to make faster, but if you want more details, you can set it to four. It increases the SD upscaler steps, but it increases the

3:53 time. So to be more efficient, I put turbola while upscaling. Also, this workflow can be optimized with turbola and its steps you can adjust according to your settings. Time to see the result of upscaler. You can easily see Sigma vision detailer is best on human images. Realistic details have been added. For example, this person can be our t-shirt model. You know, I gave you the object migration workflow. We can easily change the t-shirt and add logo with that. This is why we are doing this. Now, I will show you some different examples of these workflows result. So, the goal of these workflows is to get reference we like on the net and after that we tell Gemini to give the prompt for the exact same which is the replicate or make different version of it. For the human

4:39 images, our goal is to make realistic character we can use for our branding for like Instagram model showing off. I am telling you these to give ideas. Also, I am planning to make episode for this commercial walkthrough after showing in painting and some tools we need. Hope this tutorial helps to improve your designs. See you next video.