Comfyui Advanced Tutorial 01 - Highresfix

0:00 Welcome to a new video. Today I will present the Comfy UI advanced course part one as promised. Note that this is not a beginner course. So let's get to the first topic. Which model to use? This question has a simple answer, Flux. But there are many reasons behind that. I would like to say Hydream, but it just doesn't have Loras control net and the generation time is worse, but it has better prompt adherence. There are also fine-tuned models for Flux. I prefer flux sigma vision for generating human-based images. The fine-tuned models and luras are available on Civot AI. In this video, I will use the Fluxdev QX quantiz model and also the quantized text encoder model. All links

0:45 will be in the description. Adjust the quantization model to your PC settings. You must use the Comfy UI manager to install the custom nodes I used in this workflow. The main problem while generating images from Flux is background blur. There is a way of getting rid of that. It is called hires fix. This method is upscaling the latent image from sampler one and connect it to the second sampler. This way sampler add more details without changing the image completely. There are some critical values while doing it. The D noiseis level of sampler 2. The flux guidance value 50 works. Default value of 3.5 on general workflows doesn't work well and Lauras, especially the anti-blur Laura on Civet AI. You can check the website.

1:32 Strongly recommended. Now it's time to see the results of workflow. The second image is more detailed. Look at background. The blur amount is significantly less. But we are not done yet. What if we increase the anti-blur Laura strength to two? You will be surprised. The blur is completely gone and it even fixed the hands. Now I need to tell you important tip. The prompt is also very significant parameter. If you describe more detail about the background, the blur has become less. But you also need to tell the main subject of the focus. I mean you need to be balanced while prompting it. If you less describe the main

2:18 subject then background is more on focus. So make more detailed long prompts but be balanced. [Music] Now I have big workflow that works really well. I will explain it part by part. The load flux section this is where the models will be loaded and then set parameter section. You can use your own Laura. After that there is first pass sampler. This is where the first image created. Now here comes interesting part. There is unsampler node here. What it makes is mixing the empty latent with the latent of the image we first created. I need to warn you that this unsampler process also takes time. Same as the first

3:03 sampler. This process is not short. So I'm not telling you that this image created in seconds. After the second sampler, the outputs will be down there, the first and the second image. So you can see the massive difference for the quality. Lastly, there is third pass node here. I will bypass this as default. It also increases the quality, but it takes really long time. So if you want to try it, I need to warn you that it takes much time. The whole workflow is not created by me. I just found it. So credits goes to

3:50 owner. You can check it. It is time to see the results. Let's test it with one image. This is the prompt I am using. Also, I use flux fine-tuned. You can use normal flux dev model. You can check the civet AI for model I am using. I will give the link. I test it with house interior rather than easy single subject. It is because there are lots of details. So the goal is to minimizing errors and test the limits. You can easily see it is really bad without hyra fix. Then we made it better with hyres. But we done yet? No. There is one more thing the upscaling. But it is for next video. However, I will show you the results of upscaling. It is called TTP upscaling method. The

4:37 result is way better. Minimize the errors. Here is the final image we get. Thank you for watching. Hopefully see you next video.