Models and target size are optimized for 16Gb VRAM cards.

Key distinctions of my workflow:

• Only 4 nodes to deal with (prompt, settings, and two preview windows)
• Generating small image as fast as possible to assess the composition and skip it if you don't like it
• Upscalled image always have target pixel count, so once you found the balance between speed and quality, it will stay the same no matter how big the 1st step image, or it's aspect ratio
• Improved textures on upscaled image thanks to little hack with adding noise before second step (if textures and skin is too plasticky, add more noise (last setting in Everything node)
• No tiled upscale, no seams. Upscalling relies on you VRAM

For lower VRAM you just need to set lower target size, but to get higher resolution pictures you probably have to add tiled upscaling.