↔ S42 CutFlow Batch Resize

Every AI video pipeline hits the moment where the output resolution is wrong for what's next: too small to deliver, too big to process cheaply, or just inconsistent between clips that need to composite together. S42CF Batch Resize is the pack's answer - one node that resizes the whole IMAGE batch, with a mode dropdown covering the five ways people actually think about size, and honest control over interpolation.

The modes, because picking the right one is half the job:

• exact (default) - force width × height. The blunt instrument; it distorts if your aspect ratio changes. Fine for standardizing dimensions before compositing.
• fit_width - set the width, height scales to keep aspect. For "I need 1080 wide, whatever height that makes."
• fit_height - the vertical twin. This is your 9:16 output mode: set height to 1920 and let width follow.
• percentage - scale by percentage (50 = half, 200 = double). The fast-iteration mode: drop your whole clip to 50% for testing, flip to 100 for the final pass. Pairs naturally with the pack's Quick Preview philosophy.
• max_dimension - scale so the longest side equals max_dim. The safe normalizer: "make everything fit under 1024 on its long edge" is the classic input for models with strict resolution limits.

Then interpolation, which matters more than people think:

• bilinear (default) - smooth, the everyday choice.
• nearest - the pixel-art safe option; no smoothing, so edges stay hard. Wrong for video, right for sprites and masks.
• bicubic / lanczos - the high-quality pair, routed through PIL. Lanczos is the sharpest and is what you want when the resize is the final step before delivery; bicubic is nearly as good and a hair faster.

Outputs are clip (resized IMAGE batch), frame_count, and info - the same output shape as most S42-CutFlow utilities, so it drops into the same spots in a graph.

The mechanism is a straightforward tensor resize per frame - PIL-backed for the high-quality modes, torch for the fast ones. Deterministic, CPU-cheap, no model. Like everything in the pack, it runs on standard IMAGE batches, so it slots between any loader and any downstream node.

The practical take: don't default to exact unless you mean it. Most of the time you want fit_height (vertical delivery), max_dimension (feeding a model), or percentage (iterating). And when the resize is the last thing before the final render, spend the tiny bit of extra time on lanczos - for upscaling it's visibly better than bilinear, and the whole point of post-processing is that the output is what the viewer sees.

Installing it

S42-CutFlow install: ComfyUI Manager → search "S42 CutFlow" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/GeekyGhost/S42-CutFlow.git
pip install -r S42-CutFlow/requirements.txt

One-line requirements (opencv-python-headless); torch and Pillow (both bundled) do the resizing. "[S42 CutFlow] Loaded ..." on restart confirms the pack.