Geekatplay HyperTile Resize Image

Let's not oversell it: HyperTileResizeImage is a resize node. One image in, one image out, four interpolation methods. The reason it exists inside this pack is consistency. The HyperTilePlanner computes a target size snapped to a multiple, and this node is the one that applies it exactly, so the stage-1 neural upscale result lands on precisely the geometry the tiled pass expects. You could do the same job with a stock resize node; this one just keeps the plan in one place so you never have to retype the number.

How it works

If the image is already at target_width/target_height, it passes through untouched - no pointless resampling. Otherwise:

• bilinear, bicubic, and nearest go through torch's interpolate on the GPU.
• lanczos - which torch can't do - falls back to PIL per-frame. Slower, but sharper, and it's the default.

Output is clamped to the 0–1 range ComfyUI expects.

Inputs

• image - the stage-1 upscale result.
• target_width / target_height - default 4096. Wire these from the Planner's target_width/target_height outputs rather than typing them; that's the whole point.
• method - lanczos (default), bicubic, bilinear, nearest.

Output

image (IMAGE). Drop it on the VAE encode for the tiled pass or a preview node.

Install

It's part of GeekatplayStudio/ComfyUI-UltimateUpsacaler: Manager → search "ComfyUI-UltimateUpsacaler", or:

cd ComfyUI/custom_nodes && git clone https://github.com/GeekatplayStudio/ComfyUI-UltimateUpsacaler

Restart afterwards. The node itself only needs numpy, Pillow, and torch - the model downloads (SDXL base, Tile ControlNet, the ESRGAN upscalers) come with the pack's installer, not with this node.

Common issues

Two things worth saying, and neither is a bug:

• It adds pixels, not detail. This is the "more pixels" rung of upscaling, not generation. Our upscaling knowledge base splits the job into three - pixels, detail, video - and this node is firmly in the pixels bucket. If you feed it a soft image you get a bigger soft image; the generative detail comes from the tiled denoise pass downstream.
• nearest gets blocky at 4x. It's there for masks and pixel art, not photos. Leave it on lanczos for general use, and reach for bicubic when you're resizing a batch - the GPU path is far faster than the PIL loop.

One small gotcha: the Planner snaps its target to multiples of 64, so a "4000px" plan becomes 4032. This node doesn't question it, it just does it - which is exactly the point of keeping the plan in one place.