Comfyui Advanced Tutorial 02 - Upscaler

0:00 Welcome to another tutorial video. Today's topic is upscaling. I mean smart upscaling and how to do face detailed images for commercial usage. Let's get into the workflow. I will explain it part by part. There is load image node and after that there is a functional group which adjust max resolution. This method is smart because it uses flux to upscale the images and the Laura loader. There is a flux turbo lura here. It is important and must be downloaded. I will leave the link and down there there is load upscale model. I will prefer the real Esseran 4X. The upside you will see the Florence 2 and a string node which is our main prompt. Florence 2 group is when you

0:46 want to make prompt automatically from the image but I will make it manual because it is better. If you scroll down you will see the float node which adjusts the X value of upscaling. If you make it 2, you will get 2x of the your image. For example, 1024 to 248. If you make it 4x, you will get 4 A96 resolution. I advise you to always make it two. We will talk the reason later. On the right side, we have TTP tile image size, width, and height values. Now, I should explain how this whole workflow works. This TTP method divide the image into 2 * 3 the width and height factor which is six equal tiles. After that the sampler looks each

1:32 tile individually and goes to your prompt section searching for a meaningful word so that it understand that there is object here and it says like okay the user said that there is a street in the background so I will be aware of. This is why it is the best method. So the prompt is again more important than ever before. I am using Gemini to describe the image with 150 words. But I am also adding details manually. If I want a detail to be occur, a text or something, you must tell it. Now the final and most important parameter is coming. It is the basiculer steps and d noiseise. So the steps is how long each tile be investigated. If you increase the steps,

2:17 your image will be more detailed, but it takes too much time to make it. depends on your PC of course. Time to talk about D noiseis level. It is the value of how much your image will be consistent or detailed. The 030 is a balanced level. You can also try 040 if you want more details and changes. I do not recommend more than 040. If you go 020, it will be more consistent and less details. Okay, the first one is 8step 030D noise which works very well. I added the prompt that there is Nissan logo and GTR icon and surprisingly workflow understood it and made it. Normally all upscalers are ruining the text. This one is adding it. Very impressive.

3:03 Now time to see 12 steps. Same d noiseise. It removes the blur more but it can add more inconsistent details. Also the upscaling time is too much. I don't think you need 12 steps but depends on the image of course. and another one. This time four steps but 4x upscale of the float value. This is why you should not do 4x if you want consistency more. The logo is ruined as you can see but if you just want details you can try 4x. Now the interior photo which is very hard to make it real for AI. Eight steps 030 den noiseise is my favorite value. It also adds some details. Some of them are consistent with original. Only

3:49 little errors come in handy. We will fix it with the workflow of inpaint in the future. And now four steps and 30 D noiseis default workflow value. I think you should always start with this value to test it. Of course, if you don't find it enough, then you should try eight steps. Lastly, what if we make D noiseis too much like 0.60? I don't recommend you over 0.40. Time to see why it completely broke the image and make like abstract art. So do not do over 0.40. Okay, the upscaling course is done. Now it is time to generate hyperrealistic human images for like e-commerce, your client, maybe even your own brand. Now I will show you

4:34 the workflow. I will tell only changes I did. Rest is same as previous video. Here we use flux sigma vision and triple loader. Also the clips are different. I will leave all the links. Left side we have Damon detailer. Laura loader is also off because Sigma vision came from Flux Chanel. Do not try flux dev lauras. It doesn't work. We have cfg guider and negative prompt. You can add negative words. Rest is hireer's fix. On the right side we have upscale group. I am using nmkd upscale model. Also we have keywords down there. you must add at the end of your prompt. It details the image and face. In the future, we will replace any

5:20 object with our brand object. Take it to the commercial level. Next video probably will be about flux context. If you're wondering, I also have made workflow for context if you want to try early. Thank you for watching.