ControlNet support

ControlNet support

High quality large images with tidy details.

• Our Tiled Noise Inversion feature can cooperate with ControlNet v1.1 tile model (CN Tile, for short) to produce amazingly clear results with proper details. Example
  • CN Tile with large denoising strengths (i.e. >= 0.4) tends to produce overly sufficient details, making the image look dirty or messy.
  • MultiDiffusion Noise Inversion tends to produce tidy but overly retouched images without enough details.
• Combine the two, you get amazingly good results:
  • Clear lines, edges, and colors
  • Proper and reasonable details, no weird or dirty pieces.
• Recommended settings:
  • Denoising Strength >= 0.75
  • Method = Mixture of Diffusers, Overlap = 8
  • Noise Inversion Steps >= 30
  • Renoise strength = 0
  • CN Tile preprocessor = tile_resample, downsampling rate = 2
• If your result is blurry:
  • Try higher Noise Inversion Steps.
  • Try lower Denoising Strength.
  • Try another checkpoint.
• Compare with pure CN Tile.
  • Comparison1, 50 NI steps, denoise = 0.75
  • Comparison2, 200 NI steps, denoise = 1
• Note that high denoising strengths will change the image color. This is a known issue of CN Tile that cannot be fixed by us.