🔥 Winnougan LTX Image Resize

Any LTX-2.3 image-to-video workflow needs your reference image resized to dimensions the model will accept - width and height divisible by 32 - and it needs that size to match the latent you're generating into. In a plain ComfyUI graph that's two resize nodes, a math expression or two, and a prayer that the numbers line up. Winnougan LTX Image Resize collapses that into one node whose width and height inputs plug straight into the pack's Resolution & Frame Calculator, so the whole graph agrees on the same numbers without you doing arithmetic.

It's a purpose-built video-workflow resize rather than a general-purpose one, and the differences matter: the width and height inputs step by 32 and auto-snap, and the resize modes are the five aspect-ratio behaviors you actually want for video, not a menu of vague presets.

How it works

The node takes your image plus target width/height, snaps the targets to multiples of 32 (never below 32), and resizes with ComfyUI's own common_upscale using your chosen interpolation. Because it accepts plain integer inputs, you wire width and height outputs from the calculator directly - no text-to-int conversion nodes, no expression strings.

The mode input is where the real decisions live:

• fit_inside (default) - scale to fit within the target, preserving aspect ratio. Letterboxes/pillarboxes. Right call when your source aspect differs from the target, e.g. a portrait image going to landscape video.
• fill_and_crop - scale to fill the target entirely and crop the excess from center. No black bars, but you lose the edges of your frame.
• stretch - ignore aspect ratio entirely. Only sane when the ratios already match.
• fit_width / fit_height - lock one dimension exactly and scale the other proportionally.

interpolation defaults to lanczos, which the tooltip flags as best for downscaling; bicubic is the pick when you're upscaling.

Inputs and outputs

• image - your reference frame (IMAGE).
• width / height - defaults 768×512, wired from the calculator in practice; both snap to ÷32.
• Outputs: image (the resized result), plus width and height reporting the actual snapped size - useful downstream, and the honest way to discover your input got nudged.

Install

Part of the ComfyUI_WLTX_nodes pack, under Winnougan LTX:

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/ComfyUI_WLTX_nodes

Restart ComfyUI, or search ComfyUI_WLTX_nodes in ComfyUI Manager. No extra dependencies - it's pure torch and ComfyUI's own upscale helper.

Common issues

The recurring failure is a mismatch between the resized image and the latent dimensions downstream: resize to 1280×736 but generate into a 1920×1088 latent and you get a composited mess. The clean habit is to wire this node's width/height from the same calculator outputs that feed EmptyLTXVLatentVideo, so the two can't drift apart. And if you're resizing a portrait image for a landscape video and hate the black bars, fill_and_crop exists precisely for that - just budget for the center crop chopping your subject's edges.