Vewd

The problem it actually solves

Vewd is the answer to a very specific kind of mess: the "I generated 400 images and now I have to figure out which five are good" problem. Normally you either save everything with SaveImage and cull the output folder later, or you run PreviewImage and watch perfectly good generations get wiped with the temp dir. Vewd sits between those two - it's a media viewer node that auto-captures everything your workflow produces into a grid, lets you heart the keepers with the spacebar, and exports only those to a selects/ subfolder. Fewer files saved, zero time spent sorting.

It's from the spiritform/vewd pack, and it's a real niche rather than a clone: the node part is just one half. The same code ships a standalone app (python viewer.py path/to/images) that points a browser-based viewer at any folder, no ComfyUI required. If that's all you want - reviewing a folder of images - the app is arguably the better half of the deal, and it's worth saying so.

How it works

The node is deceptively inert, and that's the trick. At execution time process() mostly returns an empty ui message and passes along whatever IMAGE tensor it was handed - or a black 512×512 fallback if nothing's wired. The real machinery lives in web/vewd.js, which listens to ComfyUI's executed event for every node in the graph. Any node that outputs images, videos, audio, GLB/OBJ models, or Gaussian splats lands in the grid, no wiring needed. That's why the README can claim "Vewd captures everything automatically" - it's not reading your output folder, it's watching node outputs live.

Export is a separate HTTP route (/vewd/export). Hearted files are copied to your folder, and PNG exports keep the source image's metadata and get a seed chunk stamped in - so the workflow-in-the-image convention from ComfyUI's SaveImage survives the trip. Deleting from the viewer is non-destructive; it just removes the thumbnail, not the file.

The inputs that matter

Most of them you can ignore, honestly. The info_schema has:

• input (IMAGE, optional) - wire a specific image here if you want the node to output that exact one. It's optional because Vewd captures everything anyway.
• folder - where hearted media gets exported, under a selects/ subfolder. Set this. The default is a hardcoded Windows path (C:/AI/comfy/ComfyUI/output/vewd), so on Mac/Linux your exports will fail or land somewhere you won't find.
• filename_prefix - prefix for exported files, default vewd.
• max_frames - max video frames to extract when you output video as frames (0 = all). Only matters if you're feeding video out.
• selected_media - an internal passthrough for multi-select and cloud support. Leave it alone.

The output is output (IMAGE): your selected image loaded from disk like a LoadImage, or a batch tensor of everything you multi-selected. That batch mode is the interesting part - wire it to a Video Combine node and you can turn a bunch of selected frames into a GIF or MP4 on the spot.

Installing it

No models, no heavy dependencies, no pip installs - pyproject.toml lists an empty dependency list, which is refreshing. Via ComfyUI Manager, search "Vewd"; it's on the ComfyUI Registry. Or the old-fashioned way:

cd ComfyUI/custom_nodes
git clone https://github.com/spiritform/vewd.git

Then restart ComfyUI. Two optional extras: opencv-python enables real video frame extraction (without it, video output falls back to a screenshot), and ffmpeg is only needed if you export audio files and want FLAC/WAV converted to MP3.

Gotchas and the honest verdict

The community reception when this launched was politely skeptical - the top comment on the announcement was "Why a node instead of a sidebar or modal dialog?" - and that's a fair question. If you're only reviewing images, the standalone app (or any image viewer) does it with less machinery. Where Vewd earns its place is the zero-wiring auto-capture across all media types, including 3D and splats, which nothing else in the default UI really does.

The two things that bite people: that default Windows folder path, and the fact that it re-executes every run (IS_CHANGED returns NaN), so it's a review layer, not something to put in a hot path. Set your folder once, and it's a genuinely useful culling workflow - especially for video batches, where eyeballing frames beats guessing.