ComfyUI-PreviewWebM

ComfyUI-PreviewWebM

A scrubable inline video preview for image batches in ComfyUI — with a real seek bar, and without saving anything.

ComfyUI has Preview Image, but no preview for video. The built-in SaveWEBM and SaveVideo nodes write into output/, which means every time you queue a sampler to see how the motion looks, you leave a file behind to clean up later.

This node encodes to the temp directory instead and hands the frontend a preview descriptor, so the player appears in the node and nothing lands in output/. It also deletes its own previous file each run, so a long tuning session doesn't quietly fill the disk with discarded takes. Assembling and saving the real thing stays wherever you already do it.

Node

Preview Video (WebM) (category: image/video)

| Input | Default | Notes | |---|---|---| | images | — | The image batch to play back. | | fps | 24.0 | Playback frame rate, defaulting to the native rate of most current video models. Stored as an exact rational, so 23.976 stays 2997/125. | | audio | (optional) | Muxed as Opus, trimmed to the length of the image batch. | | quality (advanced) | near-lossless | Encoder speed against picture quality. See below. | | mute | false | Leave the audio out without unwiring it. The track is skipped at encode time, so the file has no audio stream at all. |

No outputs — wire images in from the same socket that feeds your save node. It is an output node, so it runs on its own.

The video's width × height is printed underneath it, matching the readout ComfyUI already draws under an image preview — including the Show width × height below the image preview setting, which turns both off.

Performance

libvpx-vp9's defaults suit neither end of what a preview is for — too slow for a glance, too lossy for a verdict. 60 frames of 1920x1080, smooth structured content:

| preset | time | per frame | size | |---|---:|---:|---:| | SaveWEBM defaults (crf 32), for comparison | 2.24 s | 37.4 ms | 228 KiB | | fast | 0.87 s | 14.5 ms | 674 KiB | | balanced | 1.12 s | 18.7 ms | 186 KiB | | high | 1.30 s | 21.7 ms | 323 KiB | | near-lossless (default) | 2.08 s | 34.7 ms | 1443 KiB |

fast is around 2.5x quicker than the built-in node's settings at the cost of a larger throwaway file. high spends its budget on picture quality rather than compression, which is why it is both slower and bigger.

near-lossless is the default because the common reason to look at a preview is to decide whether the model got it right, and at a streaming crf the codec's own artifacts are easy to mistake for the model's. It costs about 7x the bytes of balanced and none of that matters: the file is deleted on the next run, so the only real cost is time — and even then it lands slightly under what SaveWEBM takes at crf 32, because the thread and row-mt settings more than pay for the lower crf. Drop to fast when you only want to see whether the motion reads.

Installation

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

No extra dependencies; PyAV and torch ship with ComfyUI. torchaudio is only needed if you feed in audio at a sample rate Opus does not support, and ComfyUI already ships it in practice.

Notes

• Nothing is saved. Files go to ComfyUI's temp directory, which is wiped when the server starts and stops. Each preview node keeps exactly one file at a time and deletes its previous one on the next run.
• A preview does not survive a restart. The player will 404 until you queue again — the workflow remembers the result, but the temp file is gone.
• Scrubbing works because /view serves through a range-capable response. Keyframes are placed about once per second (once every two seconds on high), which bounds how far a drag has to seek back.
• Colour is converted and tagged explicitly as BT.709, full range. Left to its own devices the encoder converts with swscale's BT.601 defaults and writes no matrix tag at all, and browsers assume BT.709 for HD — a mismatch that rotates saturated hues by up to 15% (pure green arrives as (0, 216, 0)). Neutrals are unaffected, so it shows up as a tint on saturated colour rather than as a contrast shift.
• near-lossless uses 4:4:4 chroma, which is VP9 profile 1. Chrome and Firefox decode it in software; nothing decodes it in hardware. The other three presets stay 4:2:0 profile 0. An RGBA batch always stays 4:2:0, because libvpx-vp9 offers yuva420p and no 4:4:4 equivalent.
• 10-bit is not available from libvpx-vp9 unless it was built with --enable-vp9-highbitdepth, so banding in very smooth gradients is the one artifact a lower crf cannot fully remove.
• Alpha is preserved when the batch is RGBA: the stream is encoded as yuva420p and the container is tagged alpha_mode=1. Note that the alpha rides in a separate channel, so tools inspecting the file will report the main plane as yuv420p — that is normal, not a loss.
• Audio must be Opus, because WebM accepts only Vorbis and Opus, not the AAC that ComfyUI's MP4 path uses. Sample rates Opus does not support are resampled up to the nearest supported one, matching how core writes .opus files. Channel counts other than 1, 2 and 6 are narrowed to stereo.
• mute drops the track rather than silencing it, so a muted preview is a video-only file. It short-circuits before the resample, which makes it a way out if you have audio at a rate Opus cannot take and no torchaudio to convert it with.
• Odd dimensions are fine. Non-even widths and heights encode without any padding or cropping.
• The resolution line is the video's own, read off the decoded stream rather than off the batch, so it is what the player is actually showing you.
• One video per node. The frontend displays only the first result, so this deliberately produces a single file.

License

MIT