Nodes/comfyui-fvmtools/K2 Post Upscale
ComfyUI Node

K2 Post Upscale

Exact output size, with or without a neural upscaler

By ping1979ping·Created 6 months ago·Updated about a month ago· 0
K2 Post Upscale
  • image
  • upscale_model
  • image
scale2.00
methodlanczos
tile512

K2 Post Upscale is the "get this image to exactly the size I asked for" node at the end of a K2 Lab pipeline in comfyui-FVMtools. It scales the finished image to an exact final size, and the bit that makes it more than a resize node is the upscale_model method: run a neural upscaler first, then resize to the exact requested factor. A 4× model can serve a 2.5× target without a second node, and the result is still exactly 2.5×.

The methods

  • lanczos (default) - deterministic, high-quality resize. No inventing, runs in milliseconds. The KB's upscaling essay would classify this as the "more pixels, source already sharp" job, and it's still the correct answer for a finished render you just need bigger.
  • bicubic, bilinear, nearest-exact - the classic interpolation set, for when you specifically want those characteristics (bilinear for softer output, nearest-exact for pixel-art).
  • upscale_model - runs the connected ComfyUI UPSCALE_MODEL (typically a 2× or 4× ESRGAN-style model) first, then resizes to the exact requested factor.

The inputs

  • image - the batch to enlarge.
  • scale - final width/height multiplier, 1..8. This is the number you actually want out; the node guarantees it.
  • method - above.
  • upscale_model - required only for upscale_model method.
  • tile - tile size for the neural upscaler, lower = less VRAM. If a neural upscale OOMs, this is the first knob to turn.

Output is a single image. That's it - this node is a terminal convenience, not a pipeline.

Honest guidance

For a Krea 2 render that's already sharp and you just need it bigger for a client or a print, lanczos is right and the KB backs it: interpolation can't hallucinate, so it can't hurt you. If your image is soft and you want the model to invent plausible detail, that's the upscale_model job - and the KB warns this is generation, it will rewrite details, including faces. Don't let a neural upscaler rewrite a face you just spent a whole K2 workflow getting right. If you do use one, tile it and check the faces afterward.

Install

Same as the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/ping1979ping/comfyui-FVMtools

Restart ComfyUI, find it under FVM Tools/K2. No extra dependencies; the upscale model (if any) comes from your normal models/upscale_models folder.

CategoryFVM Tools/K2

Inputs (5)

NameTypeDefaultDescription
imageIMAGEImage batch to enlarge.
scaleFLOAT2.001–8Final width and height multiplier.
methodCOMBOlanczoslanczos: deterministic high-quality resize. upscale_model: run the neural model first, then resize to the exact target.
upscale_modeloptUPSCALE_MODELRequired only for the 'upscale_model' method.
tileoptINT512128–2048Tile size for the neural upscaler (lower = less VRAM).

Outputs (1)

NameTypeDescription
imageIMAGE