Thread by @JacobLeeSingh

🚀Unlock the secrets to realism and image upscaling in Stable Diffusion with this tutorial!

From low-resolution image from Midjourney to detailed and realistic image from Stable Diffusion!

Check how to do it!👇

Let's start with what you need to install
stable-diffusion-webui. This is the first step.


I won't go through the whole installation here because it will take too much time.github.com/AUTOMATIC1111/…

Once you have it installed, you need to download some models.

I personally use Realistic Vision V5.1 for this. But you can use any model you prefer.

We also need some Lora models.

Below are list of all the models I used here. 👇

List of models:
- Realistic Vision V5.1
- Detail Tweaker LoRA
- LowRA
- Skin & Hands (male/female) from Polyhedron
- polyhedron_new_skin_v1.1
- detailed_eye-10

Everything can be found on civitai or huggingface.

Once we have downloaded everything, we run the SD.

Now we must install ControlNet - in the Extensions tab - sd-webui-controlnet.

We also install ultimate-upscale-for-automatic1111.

This is one of the most important steps.

Then we download control_v11f1e_sd15_tile - which can be found on huggingface.

Place it in the ControlNet folder.
If you did everything correctly, you should see it here.

Once we've done all that, it's time to go to the img2img tab.

At the beginning, we upload a photo that we want to upscale or add details.

Below, depending on the power of the GPU you have, you can set its initial resolution. For me, it is 768x512.

Set the Sampling Method to DPM++ 2M Karras.
In my opinion, it works best.

Denoising strength is reduced to about 39-45 - this range works best for me.

But I recommend playing around with these settings and see what works best for you.

Then we go to the ControlNet options.

At the beginning, we select all the options, i.e.:
- Enable
- Low VRAM
- Pixel Perfect
- Allow Preview

And also we add the same photo as above.

The next step is to select "Tile" from the options below.

From Preprocessor, select the tile_resample option.
From the model we choose control_v11f1e_sd15_tile.

Click the icon between them - "run preprocessor"

It's almost over!

Now to do upscaling click on Script.
And we choose Ultimate SD upscale.

Change the Target size type to "Scale from image size"
And we choose how much we want to scale our image.

Also, we choose Upscaler, I prefer R-ESRGAN 4x+

The last thing we have left to do is go back up and add a prompt.
I'm posting the basic prompt I use. 👇

I'm not good with prompts in SD so I recommend you to try and play around with this prompt as well as lora model settings.

Prompt: RAW photo, full sharp, detailed face, 8k uhd, DSLR, soft lighting, high quality, film grain, Fujifilm XT3 <lora:add_detail:0.2> <lora:LowRA:0.2> <lora:polyhedron_skinny_all:0.1> <lora:skinny_new_skin:0.2> <lora:detailed_eye:0.1>

Negative prompt: BadDream, bad hands, nude, nipples, naked, NSFW, unrealistic skin, (ugly face), ugly mouth, drawing, painting, digital art, helmet, nude, large breasts, Asian, slanted eyes, Chinese

Now all we have to do is click Generate and wait for the result!

Everything can take up to 30 minutes depending on the upscaling and your GPU.

When testing the prompt/settings or anything, I recommend reducing the initial resolution of the image and setting the Scale to 1.

And this is what the result looks like! 🥳

Judge for yourself if it's worth it.

In my opinion, if someone has a 16GB GPU or more, it is one of the best ways for upscaling.

Finally, I would like to point out that I'm not very familiar with SD, I've been using it for several days, so if you have any comments or advice, please write.

I'm eager to learn more and improve the whole process.

Also, remember that this process works best for portraits or photos of faces.

It can be different for other images and you will definitely need to adjust the prompt.

✨If it is useful for you, leave a comment, share the result or leave a like.

And if you've made it this far, congratulations, and thank you! 🙏🧡