ImageTransformResizeAbsolute

When you need an image to be exactly a specific width and height - no arguments - Allor's ImageTransformResizeAbsolute is the blunt instrument you want. You give it target width and height in pixels and it resizes every frame to precisely that size. It's the node for "the model wants 832×1216," "this tile must be 512×512," or "standardize all these images before batching." And unlike ComfyUI's core resize node, it gives you a real choice of resampling method, which matters more than people think.

It's part of the Allor Plugin (Nourepide/ComfyUI-Allor), the pack that keeps RGBA and batches intact through its transform family. The "Absolute" flavor is the mechanical one; there's a "ResizeClip" sibling that fits within bounds while preserving aspect, and this one is the "exact dimensions" option.

How it works

Per image in the batch, it runs PIL's .resize((width, height)) with the resample method you select from the method dropdown:

• lanczos (default) - high-quality interpolation, the standard pick for downsizing and general resizing. This is the one the upscaling community reaches for when they want real pixels and nothing invented.
• bicubic, bilinear - smoother, slightly softer; bilinear is faster and fine for drafts.
• hamming, box - niche; box is a fast average that can look chunky.
• nearest - hard pixel edges, the pixel-art/hard-pixelation choice.

The key thing to internalize: width and height are absolute targets, so unless they match the input's aspect ratio, the image gets stretched, not cropped or letterboxed. ResizeAbsolute obeys your numbers literally; it will not preserve aspect for you. If you want "fit within 1024×1024 without distortion," that's the ResizeClip node's job, not this one's.

Output is an IMAGE at exactly the requested size, alpha preserved, and it runs on the whole batch in one go.

Inputs and outputs

• images (IMAGE) - single frame or batch.
• width (INT, default 256, min 1) - exact target width.
• height (INT, default 256, min 1) - exact target height.
• method - lanczos, bicubic, hamming, bilinear, box, nearest.

Output: IMAGE, exactly width×height.

Installing it

It ships in the Allor Plugin: ComfyUI Manager → "Allor Plugin" → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nourepide/ComfyUI-Allor

No model downloads - this is pure PIL resize.

Where people get burned

The stretch trap is #1 by a mile: feed a 512×512 image to a 832×1216 target and everything goes wide/soft, because the node isn't preserving aspect. If you need to change size and keep composition, crop or pad first, or use ResizeClip. Second, resample method matters for quality but not for speed on stills - on batches of hundreds of frames, lanczos is noticeably slower than bilinear; use bilinear for drafts and reselect lanczos for the final pass. And the usual default trap applies: the default 256×256 silently squashes anything you forgot to wire up properly.