Preview Video (WebM)

Here's the thing ComfyUI has never bothered to give you: a Preview Video node to match Preview Image. So every time you queue a video sampler just to see whether the motion reads, you reach for SaveWEBM or SaveVideo - and both write a file into output/. Tune for an hour and you've buried a dozen throwaway clips you'll never assemble. Preview Video (WebM) is the one-node fix: it plays your image batch back as a scrubable WebM inside the node, encodes to the temp directory instead of output/, and deletes its own previous file on every run. Nothing to clean up. Ever.

It's a small, unusually well-made pack - one node, one opinionated author (dreevelle), and a README that includes actual per-frame benchmark tables. You don't need a pack reputation when the docs are this honest about tradeoffs.

How it works

The node is an output node (it has no outputs - you wire images in from the same socket that already feeds your save node, and it runs on its own). On each execution it encodes the batch with libvpx-vp9 via PyAV into ComfyUI's temp directory, hands the frontend a preview descriptor, and the player pops up in the node. Scrubbing works because ComfyUI's /view endpoint serves the file with range requests, and keyframes are placed about once a second so a drag never has to seek back far. The previous run's file is deleted the moment the next one lands, so a long tuning session stays at one file per node, not gigabytes.

Because it's a preview, it's tuned like one. Colour is explicitly tagged BT.709 full-range (left to its own devices, the encoder uses swscale's BT.601 defaults and writes no matrix tag - a mismatch that can rotate saturated hues by up to 15%). near-lossless even encodes 4:4:4 so the codec's own artifacts can't be mistaken for the model's.

The inputs that actually matter

• images - the batch to play. The one socket that matters.
• fps - defaults to 24, the native rate of most current video models. Stored as an exact rational, so 23.976 stays 2997/125 instead of drifting across a long batch.
• quality (advanced) - fast, balanced, high, near-lossless. The default is near-lossless, and honestly it's the right call: at crf 15 the only thing you're judging is the model. Drop to fast when you just want to see whether the motion reads - it's about 2.5x quicker than SaveWEBM's settings.
• audio (optional) - muxed as Opus, trimmed to the batch length. WebM only takes Vorbis or Opus, so AAC from the MP4 path won't fly; unsupported sample rates get resampled.
• mute - skips the audio stream entirely at encode time rather than silencing it.

Install

The pack is dreevelle/ComfyUI-PreviewWebM - search "PreviewWebM" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/dreevelle/ComfyUI-PreviewWebM

Then restart ComfyUI. No extra dependencies - PyAV and torch ship with ComfyUI. torchaudio is only needed if you feed audio at a rate Opus doesn't support and don't want to use mute.

Gotchas worth knowing

• A preview doesn't survive a restart. Temp files are wiped when the server starts and stops, so the player 404s until you queue again. The workflow remembers the result; the file doesn't. Don't panic.
• near-lossless is 4:4:4 (VP9 profile 1) - Chrome and Firefox decode it in software, nothing decodes it in hardware. The other presets stay 4:2:0. If your machine grinds on playback, that's why.
• One video per node. The frontend only shows the first result, so it deliberately makes a single file.
• Odd dimensions are fine, RGBA batches keep their alpha, and there's no 10-bit out of libvpx-vp9 unless it was built with the high-bit-depth flag - so banding in very smooth gradients is about the only artifact a lower crf can't fix.

The name's a lie in the best way: it doesn't save anything, and that's the whole point.