Watch the picture form one sampling step at a time and stop a run that is going wrong. Step Preview needs no wires and shows whichever sampler is working, with Stop and Stop all. Save Image and Save Video get a steps switch, off by default, that shows the steps where the saved picture goes, with a Stop button. A queued workflow that shows its steps asks ComfyUI for the auto preview method for that run, only when no method was chosen in the frontend or on the command line.
6.7 KiB
Save Video
Saves your frames as a video file in the output folder and plays the result on the node. The workflow is saved inside the file, so an mp4 dragged back into ComfyUI restores the graph that made it.
The player on the node shows the saved video's size, fps, frame count and
duration, shows a tall video whole, and has a LOOP control. gif and
webp are shown as pictures with just the file name, since a video player
cannot play either.
STEPS (step_preview, off by default) shows the clip's first frame in
the player while it is being made, one step at a time, with a STOP
button beside it. Stop a clip that is going wrong after the first steps
instead of waiting for the end; a stopped run saves nothing. The saved
video replaces the steps. The Step Preview
page has the details.
Controls
- frames: BHWC frame batch to encode. For the subsampled formats, odd
dimensions are cropped down one pixel to the even sizes they require.
Optional only because a connected
videobrings its own frames — connect one or the other. - FPS (
fps, starts at16): Playback frame rate. Wire Load Video'sfpsoutput to keep the source timing. It steps in whole frames but stays a float, because the real broadcast rates are 23.976, 29.97 and 59.94 — those can be typed in, and an integer input would refuse Load Video'sfpslink entirely. - Prefix (
filename_prefix, starts asvideo): Path under the output folder; subfolders are created automatically. Files are named<prefix>_00001_.<ext>, counting up. Date tokens such as%date:yyyy-MM-dd%work, as in ComfyUI's own save nodes. A path outside the output folder is refused. - CRF (
crf): Quality — lower is better and larger.19is visually lossless for h264. What it means locally depends on the codec; see the format table. - Format (
format): Container and codec, see below. - Ping-pong (
pingpong): Plays the clip forward then backward so it loops seamlessly. Roughly doubles the frame count and the encode time. The first and last frames are not repeated at the turnaround, so the loop does not hitch. - Metadata (
save_metadata): On by default. Turn it off to share a file without your prompt and node graph embedded in it. If ComfyUI was started with--disable-metadata, nothing is embedded whatever this says. - audio (optional): A track to mux into the file, e.g. Load Video's
audiooutput. - video (optional): A core
VIDEOhandle, e.g. Load Video'svideooutput. Connecting it supersedes theframesandaudioinputs — the video's own frames and track are encoded instead. Core decodes it in one uninterruptible step; see The one phase Cancel cannot reach below.
Frame rate precedence
A connected video brings the rate its frames were cut at, and that rate wins
over the fps widget: replaying those frames at a stale widget rate would
drift the picture away from the audio muxed beside it. When the two differ,
one line naming the rate actually used is written to the ComfyUI console. With
nothing connected to video, the widget is the only source and the console
stays quiet.
Formats
| format | what it is | audio | crf means |
|---|---|---|---|
mp4 h264 |
the default; plays everywhere | AAC | crf |
mp4 h265 |
about half the size at the same quality, where supported | AAC | crf |
mp4 h264 nvenc |
h264 encoded on an NVIDIA GPU — much faster, slightly larger | AAC | cq |
mp4 h265 nvenc |
h265 on the GPU | AAC | cq |
webm vp9 |
web delivery | Opus | crf |
webm av1 |
newer, smaller than vp9 | Opus | crf |
mov prores |
ProRes HQ editing master; large | PCM | ignored |
mkv ffv1 |
bit-exact lossless; very large | FLAC | ignored |
gif |
short silent loop, per-frame adaptive palette | none | ignored |
webp |
short silent loop, much smaller than gif | none | quality |
The nvenc formats need an NVIDIA GPU with a working driver, and refuse frames below roughly 145x49. If the encoder cannot start, the node says so and names a software format to use instead rather than failing with a driver stack trace.
gif and webp have no streaming encoder, so every frame is held in memory
before the first byte is written — keep those to short clips. Neither carries
metadata, so an embedded workflow is not available for them.
The node's face follows the table: widgets the chosen format ignores are
hidden rather than left to look meaningful. The CRF field leaves the
FPS · CRF row for mov prores, mkv ffv1 and gif (their rows read
ignored above), and the Metadata row disappears for gif and webp,
which cannot carry it. Hidden is not removed:
the widget keeps its position and its value, saved workflows store and restore
it exactly as before, and switching the format back brings it back with the
number it had.
Output
One file per run. Audio and video land in a single file — there is no separate audio-less copy. The embedded workflow survives in mp4, mov, webm and mkv; matroska uppercases the tag names, which is only visible to tools that read them directly.
The node also has a file_path output: the absolute path of the file it just wrote, for wiring into downstream nodes that want the file itself. Leaving it unconnected changes nothing — the node saves either way.
While it runs
Encoding happens off ComfyUI's execution thread, so the server keeps answering while a long batch is written. The node fills its progress bar frame by frame, and Cancel stops the encode within a frame instead of at the end of the batch. Muxed audio is encoded after the picture, so the bar restarts and counts that track's audio frames; Cancel stops there within one of them.
The one phase Cancel cannot reach
Connecting video adds a step this node does not control. ComfyUI core reads
the whole file — every frame, decoded to float32 — before the first frame
reaches this pack, and the core call that does it, VideoInput.get_components(),
takes no progress or interrupt argument to pass one in. So that phase shows no
progress and does not stop on Cancel: the cancel is registered right away
(the decode runs off the execution thread, so the server still answers) and
takes effect at the first frame of the encode, once control returns here.
How long that lasts scales with pixels, not with file size. Measured on one
desktop: ten seconds of 832x480 at 16 fps came back in about 0.6 s, while ten
seconds of 1080p at 30 fps took about 7 s and held 7.5 GB of frames in memory.
Minutes of high-resolution footage are the case to avoid — RAM runs out before
the wait does. Wiring Load Video's frames output instead of video keeps the
whole run interruptible and progress-tracked, since that decode is this pack's
own loop.