Upscaling without mush: hires fix, ESRGAN and tiled diffusion

Rendering big directly breaks composition (see the resolution note in the prompt-structure thread). The working route is: render at native size, then upscale in one of three ways. 1. Hires fix (a second img2img pass) Render 512x768 (SD 1.5), upscale 1.5-2x, then re-sample at denoise 0.35-0.55. The second pass adds real detail at the new size. In ComfyUI it is simply Upscale Latent or Upscale Image, then a second KSampler with denoise below 1. 2. A pure upscaler model (ESRGAN family) Real-ESRGAN and community models such as 4x-UltraSharp enlarge 4x with learned sharpening, no diffusion. Fast, faithful, no new detail. Good for a final print size, or as the upscale step inside hires fix (use "Upscale Image (using Model)" in ComfyUI). Real-ESRGAN: https://github.com/xinntao/Real-ESRGAN 3. Tiled diffusion (Ultimate SD Upscale and similar) Upscale with ESRGAN, then cut the image into tiles (512 or 768 square with overlap) and re-sample each tile at low denoise, 0.2-0.35. This reaches 4K and beyond on modest hardware because only one tile is in memory at a time. What goes wrong - Tiles each grow a face: denoise too high, or the prompt describes a subject. Use a texture-only prompt for tiled passes ("detailed, sharp focus") and keep denoise low. - Visible tile seams: increase overlap or enable seam fixing. - Plastic skin: the ESRGAN model is over-sharpening; try a gentler one or blend 50% with a plain Lanczos resize. On a CPU Tiled passes multiply time by the number of tiles. On a CPU engine, ESRGAN alone is the practical choice.