Upscaling images by IMG2IMG, my basic guide

Intro

Modern models know not so many human body poses. Some know more than others - like ponyDiffusion, but its knowledge quite... specific. And if you have some pose in mind - you can't always describe it in all the details, including view angles and the such. And even if you can describe them the way humans will understand you - models won't always do too. Of cause there's LoRAs, but even they're not all-powerful. So there's a use for IMG2IMG tool and reference images for it.

My "data preparation"

I personaly have storage of snapshots that I take from everything you can take screenshot of - videogames, movies, tv-shows, youtube videos, tik-toks, camgirl streams and of cause classic porn. In addition to just pics you can get from any corner of internet.

I take pictures of battle-scenes, where you can see people in interesting superheroic poses shot from unusual view-angles, or pictures of CGI-effects, or landscapes - basicaly everything interesting. Images don't take space at all, on modern drives.

This snapshots I store sorted and resized, so when I cut person from frame, resulting size isn't very large, but thankfuly SDv1.5 models can nicely take a grip of even 400*300px image.

Workflow

So my pipeline is really simple and easy.

1. I cut person from frame. I try always use some standard aspect-ratios: like 0.666-2:3 or 0.750-3:4 or 1-1:1. When you use something strange like 0.5-1:2 you even can have some artifacts on resulting picture.

2. I go to IMG2IMG tab in Automatic1111 and scale original image to comfortable for SDv1.5 resolution: its about 768px per bigger side, or kinda 600*800px or close. Sometimes you have to upscale source, sometimes downscale it. I use random seed and go from 0.30 denoising strength (DS) up to 0.60-0.70 with a step of 0.05. With DS over 0.55 neural models usually add too much own fantasy to image - they change phisique, pose and shot-angle, destroying original idea. And this is not always bad. Normaly best for me is 0.42-0.54 DS.

I don't write sophisticated prompt, I go with basic scene description - model can take from source image enough data.

Common way is to make a lot of variations and choose what you see as best. Even on starter-level modern GPUs, time to generate one not-upscaled image by SDv1.5 models is very quick. On my RTX 4060ti 16GB its 2-6sec per image with DPM++ 2M Karras 34 steps, I usually use this preset.

3. When I've done enough variations, I take one of them and replace source image at IMG2IMG tab with it. I don't change prompt. I just choose an upscale to x3-x4 times - it's maximum I can do with 16+16GB of shared VRAM using SDv1.5. Resulting size is about 2300*2300px or 2000*3000px. I set DS to 0.32-0.38 - its quite important, I think at anything over 0.40 DS you too often get too many artifacts. I use standard R-ESRGAN upscaler, I can't see difference from -anime6B option or 4x-UltraSharp.

4. Sometimes I iterate that workflow more than one time - downscale resulting upscaled image to 768px by bigger side, to get more style on new source picture, and get better upscale from it.