Video Crop Upload (Purz)

Video diffusion models are picky about aspect ratio and resolution - feed Wan or LTX a 1920×1080 clip when the model wants 1024×1024 and you'll fight the pipeline the whole way down. PurzVideoCropUpload is the "fit this video to those dimensions" node: upload a clip, pick a target size and a fit mode, and it hands you back a VIDEO at exactly that resolution, audio intact. No ffmpeg, no leaving ComfyUI.

It's one of the two crop flavors in purzbeats' ComfyUI-Purz pack. The Upload suffix means your source file goes into ComfyUI's input/ directory and persists across browser reloads - reload the page and the video, the crop settings, everything is still there. (The plain Video Crop node does the same job client-side, source never touching the server, but loses state on refresh.)

How it works

On queue, the backend reads the source from input/ and resizes it frame by frame with OpenCV - INTER_AREA when scaling down, INTER_LANCZOS4 when scaling up, roughly 10× faster than the old per-frame PIL path. The fit modes are the classic four:

• stretch - squish the source to fill the target, aspect ratio be damned.
• cover - scale to fill, then crop the overflow using the anchor (center, top, bottom, left, right, corners).
• contain - scale to fit inside, letterboxing the leftover with background_color.
• crop - like cover, plus manual zoom and offset_x / offset_y so you can punch in and reposition by hand.

Audio passes through. The output is a ComfyUI VIDEO, so it plugs straight into a video preview node or anything downstream that eats the VIDEO type.

The inputs that matter

• video - a dropdown of video files in input/; upload one via the node's button. It'll read (no video) until you do.
• width / height - 16–16384, default 1024. The target canvas.
• fit_mode - stretch / cover / contain / crop. cover is the sane default.
• anchor - where the crop window sits when you're cropping; center is right most of the time.
• background_color - the letterbox color for contain.
• zoom / offset_x / offset_y - only active in crop mode; offsets are normalized −1…1.

Output is a single video (VIDEO).

Install

Standard for this pack:

cd ComfyUI/custom_nodes
git clone https://github.com/purzbeats/ComfyUI-Purz.git
cd ComfyUI-Purz
pip install -r requirements.txt

then restart ComfyUI (or install "ComfyUI-Purz" from ComfyUI Manager). Requirements are OpenCV, Pillow, NumPy, and torch - no models to download.

Where people get burned

• It needs a modern ComfyUI. The node outputs the VIDEO type, which depends on the newer comfy_api video support. On an old install the node still loads but hands back a tiny 64×64 black fallback frame and prints diagnostic lines in the console. Update ComfyUI first.
• This is a re-encode, not a passthrough. It's fast, but the clip is resized frame-by-frame and muxed back; don't expect a byte-identical stream.
• Big targets get slow and hungry. The max is 16384 on a side - nobody actually wants that. 1024–2048 is where this node stays comfortable.
• Pick the right flavor for the job. If you need persistence across reloads and don't mind the source sitting in input/, use this one. If you'd rather the source never leave your browser, use the client-side Video Crop node instead.

If you're normalizing source footage before a video pass, cover + center at the model's native size is the setup you'll use 90% of the time.