Upscale Method JK🐉

Before you get excited: this node does not upscale anything. It's a settings dispatcher. You pick an image upscale method and a latent upscale method from two dropdowns, and it outputs both as strings so you can wire one choice into every upscale node in your workflow. If you've ever rebuilt a big hi-res-fix graph and had to set "lanczos" in four different places, you know the itch this scratches.

That's not a knock. In a graph with a latent upscale between sampling passes and an image upscale before the VAE, the two steps want different methods - and forgetting to change one is exactly how you get a pixelated intermediate. Centralizing the choice in one node means you flip a dropdown instead of hunting through the graph.

The two dropdowns

• image_upscale_method (default lanczos) - nearest-exact, bilinear, area, bicubic, or lanczos. This feeds ComfyUI's image resampling, e.g. before a VAE encode or after a decode.
• latent_upscale_method (default bilinear) - nearest-exact, bilinear, area, bicubic, or bislerp. This feeds latent-space resizing between sampling passes.

The defaults are the ones you actually want. For the image side, lanczos is the right answer when your source is already sharp and you just need more pixels - it's an interpolation, it can't invent content, and it's effectively free. For the latent side, bilinear is the standard companion to a hi-res-fix second pass, and bislerp (Batched Islerp) is the fancier option that reduces the blocky/pixelated look you can get when scaling latents aggressively - worth trying if your second pass comes out splotchy.

Outputs are the two same-named strings: image_upscale_method and latent_upscale_method. Wire them into the matching input of your resize/upscale nodes.

Install

It's one node in the JakeUpgrade pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jakechai/ComfyUI-JakeUpgrade
cd ComfyUI-JakeUpgrade
pip install -r requirements.txt   # Windows standalone: install.bat

Or ComfyUI Manager → search "JakeUpgrade". No models needed.

Where people get burned

The biggest misunderstanding is expecting more than pixels. This node's image methods are all interpolation - none of them add detail that wasn't there. If you want a generative upscaler (a model that invents pores and texture, SeedVR2 or SUPIR territory), that's a whole different workflow with its own model files, and this node won't help. The other classic mistake is wiring the same string into both the image and latent upscalers - the dropdowns exist separately because they need to be. Keep image on lanczos and latent on bilinear and you're in the sweet spot 90% of the time.