🍒Image_Resize📐图像尺寸缩放

ComfyUI's stock resize node works, but it has a habit of handing back odd dimensions that then fight with latent-size requirements. Image_Resize from KimNodes is a friendlier take: give it a target size, and it scales your image so the long side lands on that value, rounded up to a multiple of 8 - the resolution sweet spot most diffusion pipelines want. Aspect ratio is preserved, no stretching.

It also carries a little anti-aliasing trick you don't see in every resize node: the supersample option. Turn it on and the image is first upscaled 8× and then downscaled to the final size. It's a cheap, well-known way to get smoother edges when downscaling - the same idea behind supersampling in game rendering, and exactly the "interpolation, not generation" philosophy the upscaling crowd keeps coming back to.

How it works

It takes image, target_size, supersample (true/false), and resampling (lanczos, nearest, bilinear, bicubic). The algorithm looks at the current width vs height: it scales the long side to target_size (rounded to the nearest multiple of 8), computes the other side proportionally, and does the same 8-multiple rounding there. Then it applies the chosen filter - LANCZOS is the default recommendation for quality, and if supersample is true it resizes to 8× first, then down to final with the same filter.

Output: a single IMAGE. One honest limitation: it processes the first frame of a batch and drops the rest. If you need to resize a whole batch, loop it or use a batch-capable resize node.

The inputs that matter

• target_size - the desired long-side length (256–15360, step 4). Default 1024.
• resampling - filter choice. For downscaling with quality, keep lanczos; nearest is only for pixel-art where you want hard edges.
• supersample - on = 8×-then-down anti-alias pass; off = single pass. Costs a little time, buys cleaner edges.

Installation

Standard KimNodes pack install - Manager → search "ComfyUI_KimNodes" → Install → Restart, or git clone https://github.com/wjl0313/ComfyUI_KimNodes into custom_nodes. No extra dependencies (pure PIL + torch).

Common issues

• Wrong aspect / stretched output - shouldn't happen here; the node preserves aspect by design. If you see distortion, check what fed it - an unusual tensor layout (e.g. CHW) can confuse the width/height read.
• Batch only produces one resized image - the first-frame-only limitation. For full batches, use it per-frame or reach for a batch node.
• Odd-looking dimensions - it rounds to 8-multiples, which is usually a feature, but if you need exact dimensions for a strict controlnet alignment, the rounding can surprise you. Check the output size and account for it.

It's not a headline node, but it's a solid default resize - the kind that quietly prevents "why won't my KSampler accept this size" errors.