Whirlpool Upscaler

Whirlpool Upscaler

This is a modified implementation of impact-pack's iterative upscaler. It leans in on the idea that giving too much attention to computation at high resolutions isn't a good idea.

Node settings

• upscale_by: Final resolution multiplier (e.g., 2.0 = double width and height)
• upscale_curve: Progression curve for all parameters (resolution, CFG, steps, denoise). 1.0 = linear progression, >1.0 = exponential progression
• iterations: Number of complete sampling cycles to perform (4 iterations would mean doing four sets of Steps)
• steps_start: Number of sampling steps for the first iteration
• steps_end: Number of sampling steps for the last iteration
• cfg_start: CFG scale for the first iteration
• cfg_end: CFG scale for the last iteration
• denoise_start: Denoise strength for the first iteration
• denoise_end: Denoise strength for the last iteration
• resize_filter: Image resizing filter algorithm - "lanczos", "nearest-exact", "bilinear", "area", or "bicubic"
• tile_size: Tile size for VAE operations to manage memory usage (if you get lag due to low VRAM then set this lower)

How It Works

Each iteration upscales the image to a progressively larger resolution. Steps, CFG, and denoise values evolve from start to end values across iterations. The upscale_curve determines how linear or non-linear this progression is.

Upscale Curve Examples (4 iterations)

• upscale_curve = 1.0: Linear progression → 1.25x → 1.50x → 1.75x → 2.00x

• upscale_curve = 2.0: More exponential → 1.13x → 1.42x → 1.69x → 2.00x

• Higher upscale_curve values:

  • Faster, spends more time sampling at lower resolutions
  • Smarter, reduces body horror
  • Less detail, resembles base image more

• Lower upscale_curve values:

  • Slower, spends more time sampling at higher resolutions
  • Dumber, more body horror
  • More detail, different to base image

So try to strike a balance. And if you change upscale_by then you'll definitely want to change upscale_curve as well.

Tips

• Reducing Artifacts: Either decrease CFG, increase steps, or connect the model input to a Skimmed CFG node.
• Too Many Fingers/Body Horror: Either reduce denoise, reduce the base resolution of the image you're feeding the upscaler, or increase the upscale_curve.
• Better Images: If you connect the model input to a Skimmed CFG node and set cfg_start really high, it'll usually result in better images.

Known Issues

• Cancelling doesn't instantly stop the generation process. You have to wait for the current iteration to finish before the process will terminate. Please share if you have the solution.