stdin is always flushed before yeilding. While this loses some potential
speedup by processing video output on cpu while the gpu is processing
the next batch of frames, it is likely incompatible with errors related
to metadata which could not be replicated. For now, stdin is flushed to
reduce the likelihood of regression.
Fixed a mistaken ordering in the batch_manager node causing the
unique_id to be overwritten.
Re-add the dummy count value to the batch manager node in js. Without
it, an error during the first batch step will result in teh batch
manager node not being marked as stale. Since the batch_manager node
exists only as an optional input, IS_CHANGED will never be called and
there must be some external change to ensure it is still marked stale
Processing video can have prohibitively high memory requirements.
This commit introduces a system for dividing the processing of a video
input into multiple workflow executions.
To accomplish this, the reference to the cv2 video_capture instance, and
the ffmpeg subprocess are kept open and the workflow is automatically
re-executed for each partial step.
Previously, Load Video (Upload) and Load Images (Upload) would swap to
the new file or folder, but would not update the preview. They have both
been fixed.
In switching to video previews to use DOMWidgets, the performance
optimization to hide elements when zoomed out was left enabled by
default. This hiding has been disabled to restore prior functionality.
Load Video and Load Audio both function with inputs not officially
listed as supported. Load Video can be used to load image files as a
substitute for Load Image and Load Audio can be passed video files.
While users have always had the option to manually type the path to a
filetype not ostensibly supported, this can be inconvenient.
A hotkey (Ctrl+g) has been added to temporarily disable the extension
filtering.
Migrates preview widgets to use the ComfyUI DOMWidgets implementation.
This resolves a number of issues with flickering, adds support for DOM
element clipping, and should have better compatibility with future
updates.
resolves#128
Fixes to logic for determining when the aspect ratio of a video preview
has changed.
Previously, any change to force_size, custom_width, or custom_height
would cause the video preview to reload. This has been changed to only
reload the preview if a change of values has occurred that could change
the aspect ratio.
Before, custom_width and custom_height widgets were virtual widgets that
were used to generate an equivalent force_size value on serialization.
This logic has been moved to the python side, but is largely unchanged
for simplicity.
This allows for custom_width and custom_height to be converted to an
input and set by output from python code which is not known before
execution.
resolves#119
After substantial rewriting, paths are declared as strings in the python
code, and converted to path widgets in the webui if they include an
option for vhs_path_extensions. resolves#104
VALIDATE_INPUT and IS_CHANGED for path nodes were rewritten and largely
consolidated in utils to accommodate validation of inputs not known before
execution.
Known issue: VALIDATE_INPUT is never called for load_audio which results
in error messages that aren't very helpful if a bad path is provided
An error is no longer immediately raised if ffmpeg is not found
Internal paths are now absolute for load images advanced previews
Instead of requiring VHS_UNSAFE_PATHS to disable checking paths on the
new endpoints, checking is disabled by default and VHS_STRICT_PATHS can
be used to re-enable it. This makes the added functionality opt out,
instead of opt in, as has been a primary goal, but the quality of life
has won me over, the difference in availability is minimal compared to
the functionality of the existent path nodes and I suspect setting would
cause unnecessary burden upon the average user.
Fix an improper check in the js code causing previews to break when
force_size is left on disabled.
Previously, force_size information was discarded when requesting
an advanced preview as the video was instead scaled to the size of the
node. Fixing this was deemed particularly difficult since the python
code does not have access to the actual video resolution before starting
the ffmpeg process.
This is solved with a more complex ffmpeg command that passes the
cropping logic to the ffmpeg subprocess.
Further migrations to video parameters had broken display of output
videos when an old workflow is loaded. This is fixed.
When the refresh button was clicked in ui, the custom size options of
Load Video were overwritten and lost. When the refresh button is
clicked, the custom size options are now readded properly.
Fix params being uninitialized on migrated workflows
Fix null filename when performing a right click -> save on a video
Fix edge case in paths of a specific length causing console spam
The toggle for saving metadata has been changed back to a boolean
instead of a combo. While I wished to provide for future compatibility,
the change broke existing workflows and there was no accommodating
implementation to load metadata saved under a different tag.
The log level for ffmpeg preview processes has been raised to fatal.
There's a good chance this will make debugging harder for users when
errors arise, but aborted requests were too common for the error
messages produced by them to be acceptable.
Format widgets for ints are now properly rounded.
Load Images had a typo prevent it's extension filter (of nothing) from
applying
Audio inputs and outputs now consistently use lowercase audio as a
description and have a format name of VHS_AUDIO. While other extensions
could share the audio type name, they would not pass functions.
Previously when using ffmpeg to load a video, care was needed to ensure
that the first frame was index 0, but the frame at 1s was the first to
appear after 1s. This should now be fixed.
The js code for load video nodes now passes the extension as part of the
format so that gifs and webps can be properly distinguished between when
deciding if an advanced preview should be loaded.
Fixed a mistake in serialization of custom height on load video nodes
Add "-b:v" flag to webm format, resolves#78
Fixes Video Combine to provide an output usable outside the ui,
a list of files produced, resolves#57
Fix error in file path safety checks across drives on windows
Fixes for preview generation when backing file has audio
Don't scale video previews above their base resolution
Implement select_every_nth for advanced previews
Fix regression with improper callback checking on Load Audio
Reorder Load Audio in node mappings to not be grouped with image nodes
Video Combine now returns a list of output files:
Partial fix for #57
Move safety checks to single shared function, added opt out by
environment variable. Easier methods are being considered
Modifications to viewvideo endpoint to work with paths instead of
standard methodology
Return an empty list of options instead of an error status for path
suggestions in an attempt to further reduce console spam.
Substantial js refactoring to simplify preview logic and
compartmentalize code that is shared between upload and path nodes
Enable path suggestions for Load Images
Add masks equal to image size if any image has masks
This causes images with masks to function as expected and those
without to remain fully opaque.
Special thanks to @miaoerduo
Fix counter regex to work with audio outputs
Fix extension not being set for animated image outputs
Reductions to extraneous preview loads
Prevent new preview code from applying to webp files
save_metadata is now a format flag
- This allows saving to different tag and hides for image formats
Implement high bit depth support
- Disabled by default, but works for images and has shown in testing to
produce objectively higher quality output than base Save Image
Fix target_size from load video not applying when 2 dimensions are set
When an input video is scaled to a new aspect ratio, it is now cropped
Removed try wrapping of load video nodes. This creates more confusion
now a system of fallbacks isn't used
A fix for parsing format_widgets when applied outside a '_pass'
Fixed missing import of logger in utils
Animated images are now included in the animated preview code
- This improves performance and allows for pausing
Cleaned up injection of code to load metadata
Switch to new path widget code on remaining nodes
Performance tuning on video preview code
Surround preview code in try block in case request is closed
- May warrant a future pass, but ffmpeg errors will still log to console
Swap lack of common path in get_path to status=403
- This was mostly made to ease future changes to common path behaviour
Simplify the value field of preview widget.
- Reduces complexity and lessens duplicate requests on first loa
Introduce ui setting for advanced previews
- Reliability is shakey, so it's better kept as opt-in for now
Previews are properly integrated and apply to Video Combine nodes
More advanced formatting code when a path is too long to display
Reduced console spam from path suggestions
A number of nodes take a path as input.
This commit gives a search box with a list of valid files and folders
It also provides support for quickly selecting from valid options by
clicking or pressing tab.
Completion options are currently only suggested for subfolders of
comfyui, but this is easily changeable.
While having the optional arguments passed is non-trivial, there's an
easy fix to get retrieve all params sent from the client.
Substantial changes to how formats are stored on the backend to ensure
that the extra data isn't displayed in the dropdown list or upon
clicking the "Refresh" button in the ui.
Update most existing existing formats to expose options as widgets
Combo widgets which use the first item instead of a default field are
now successfully reloaded if not defined.
Fixed a minor IOB from combo inputs having only a single element in
their definition
The old preview method stored the url instead of the params.
This makes it non trivial to parse whether it was to an image or video.
For safety, this information is instead discarded.
computeSize now calculates directly from the nodes width instead of
saving the width from the last draw operation. This fixes a race
condition when a video loads before the draw frame occurs.
the paused and hidden state of a preview widget are now serialized as
part of the vlaue. This means they survive reload or restoring a
workflow. Addionally, videos are paused before being hidden to improve
performance on some browsers.
A frustrating amount of digging with an awfully simple solution.
The current selected widget for mouse events is held in
app.canvas.node_widget.and clearing it prevents the continued
propagation of move events to parent widgets.
Also provides a minor fix for canceled file uploads
Austin Mroz