Skip to content

Upscaling#

The Upscaling tab (3) enlarges an image in two stages: an upscale model (a super-resolution network) multiplies the pixels, then a tiled generative pass with a tile control layer adds the detail the network cannot invent. The generative pass works with SD 1.5 and SDXL main models.

The upscaling launchpad

Steps#

  1. Drop or pick the source image (JPG, PNG or WebP up to 100 MB), or send one from the gallery or the Edit tab (Send to Enhance).
  2. Pick the upscale model. The catalogue ships several: RealESRGAN x4 and x2 for photos, ESRGAN SRx4 for general images, SwinIR for fine texture, AnimeSharp for drawings and line art.
  3. Set the scale (2x or 4x). The output size is the input multiplied by the scale; the tab warns when it exceeds the maximum dimension.
  4. Pick the main model (SD 1.5 or SDXL) and its tile ControlNet; the tab lists what is missing and links to the model manager.
  5. Optional prompt: describe the image, not the change. A short, accurate caption helps the detail pass; a style word shifts the look.
  6. Set creativity and structure, or pick a preset: Conservative, Balanced, Creative, Artistic.
  7. Press Generate. The result lands in the gallery next to the original.

Creativity and structure#

  • Creativity is how much freedom the model has when adding detail. Low stays close to the source; high invents, and with a prompt it increases the prompt's influence.
  • Structure is how strictly the layout is kept. High keeps every edge where it was; low allows the composition to shift.

Photos and renders: conservative or balanced. Rough drafts and small generations you want to "finish": creative. Concept art you are willing to let change: artistic.

Tiles#

The generative pass runs on tiles so any size fits in memory. Tile size defaults to 768 for SD 1.5 and 1024 for SDXL; lower it if you run out of memory. Tile overlap (default 128) hides the seams; raise it if you see grid artefacts on flat areas, at the cost of time.

When to upscale#

Upscaling a 1024 generation is the right way to reach print or 4K sizes; generating directly at those sizes repeats subjects, except with Flux models and DyPE. For a single region of a large image, the canvas recipe in Inpainting is more precise than a whole-image pass.