Fix temp files generated during the upload process not being cleaned up quickly enough. This included
downloaded files, generated preview images, and Ugoira video conversions.
Before we relied on `Tempfile` cleaning up files automatically. But this only happened when the
Tempfile object was garbage collected, which could take a long time. In the meantime we could have
hundreds of megabytes of temp files hanging around.
The fix is to explicitly close temp files when we're done with them. But the standard `Tempfile`
class doesn't immediately delete the file when it's closed. So we also have to introduce a
Danbooru::Tempfile wrapper that deletes the tempfile as soon as it's closed.
For videos with sound, save information about audio volume levels in the
media asset's metadata. These values are stored:
* FFmpeg:AudioPeakLoudness The peak loudness of the audio track, from 0.0 (silent) to 1.0 (max volume)
* FFmpeg:AudioAverageLoudness The average loudness of the audio track, from 0.0 (silent) to 1.0 (max volume).
* FFmpeg:AudioLoudnessRange The difference between the quietest and loudest sounds in the audio track (in decibels).
* FFmpeg:AudioSilencePercentage The percentage of the video that is silent (1.0 is completely silent, 0.5 is 50% silence, 0.0 is no silence).
These values are calculated based on the EBU R 128 standard, using the ffmpeg command below:
ffmpeg -i file.mp4 -af silencedetect=duration=0.05:noise=0.0001,ebur128=metadata=1:peak=true:dualmono=true -f null /dev/null
See the links below for details:
* https://en.wikipedia.org/wiki/EBU_R_128
* https://www.ffmpeg.org/ffmpeg-filters.html#ebur128-1
* https://tech.ebu.ch/loudness
* https://tech.ebu.ch/docs/tech/tech3341.pdf
Fix .webm files not including the `FFmpeg:VideoBitRate` and `FFmpeg:AudioBitRate`
fields in the media_metadata table. This was because the .webm format
doesn't include the video or audio bit rates in the metadata, and
ffprobe doesn't calculate them either, so we have to calculate them
ourselves by hand.
Fixup for 523d7afdd.
Include the following metadata tags for videos:
* FFmpeg:MajorBrand (e.g. mp42)
* FFmpeg:PixFmt (e.g. yuv420p. Indicates bit depth and color subsampling mode)
* FFmpeg:FrameCount (e.g. total number of frames in the video)
* FFmpeg:VideoCodec (e.g. h264)
* FFmpeg:VideoProfile (e.g. Main)
* FFmpeg:AudioCodec (e.g. AAC)
* FFmpeg:AudioProfile (e.g. LC)
* FFmpeg:AudioLayout (e.g. stereo)
* FFmpeg:AudioBitRate (e.g. 128kb/s)
If a media asset is corrupt, include the error message from libvips or
ffmpeg in the "Vips:Error" or "FFmpeg:Error" fields in the media
metadata table.
Corrupt files can't be uploaded nowadays, but they could be in the past,
so we have some old corrupted files that we can't generate thumbnails
for. This lets us mark these files in the metadata so they're findable
with the tag search `exif:Vips:Error`.
Known bug: Vips has a single global error buffer that is shared between
threads and that isn't cleared between operations. So we can't reliably
get the actual error message because it may pick up errors from other
threads, or from previous operations in the same thread.
Use ExifTool to get the dimensions of Flash files instead of calculating
it ourselves. Avoids copying third-party code.
Fixes a bug where Flash files with fractional dimensions (e.g. 607.6 x 756.6)
had their dimensions rounded down instead of rounded up.
Fixes another bug where Flash files could return negative dimensions.
This happened for two files:
* https://danbooru.donmai.us/media_assets/228662 (-179.2 x -339.2)
* https://danbooru.donmai.us/media_assets/228664 (-179.2 x -339.2)
Now we round these up to 1x1. This is still wrong, but it's less wrong than before.
Fix certain corrupt GIFs returning dimensions of 0x0. This happened
when the GIF was too corrupt for libvips to read. Fixed by using
ExifTool to read the dimensions instead.
Also add validations to ensure that it's not possible to have media
assets with a width or height of 0.
Don't allow uploading videos with unsupported video codecs.
The only video codecs we allow for MP4 files are H.264 and VP9. Other
codecs, including H.265 (aka HEVC), MPEG-4 part 2, and AV1, are
disallowed because they're not universally supported by browsers.
Firefox doesn't support H.265 or MPEG-4 part 2, and Safari doesn't
support AV1.
Additionally, don't allow videos with multiple video tracks, multiple
audio tracks, or no video tracks. Multiple video and audio tracks are
disallowed because they're rare and for moderation purposes, we don't
want people hiding content in extra tracks.
These restrictions really only apply to MP4 videos, since WebM files
don't support multiple video or audio tracks and only support a limited
number of codecs (VP8 and VP9 for videos, Vorbis and Opus for audio).
There are currently 22 posts with unsupported video codecs:
* https://danbooru.donmai.us/posts?tags=video+is:mp4+-exif:Track1:CompressorID=avc1+-exif:Track2:CompressorID=avc1+-exif:Track1:CompressorID=vp09+-exif:Track2:CompressorID=vp09 # AVC1 is H.264
There is one post that has multiple audio tracks:
* https://danbooru.donmai.us/posts/2382057
Fix StatementInvalid exception when uploading https://files.catbox.moe/vxoe2p.mp4.
This was a result of multiple bugs:
* First, generating thumbnails for the video failed. This was because
the video uses the AV1 codec, which FFmpeg failed to decode. It failed
because our version of FFmpeg was built without the `--enable-libdav1d`
flag, so it uses the builtin AV1 decoder, which apparently can't
handle this particular video (it spews a bunch of errors about "Failed
to get pixel format" and "missing sequence header" and "failed to get
reference frame").
* Because generating the thumbnails failed, an exception was raised. We
tried to save the error message in the upload_media_assets.error
field. However, this also failed because the error message was 77kb
long (it contained the entire output of the ffmpeg command), but the
`upload_media_assets` table had a btree index on the `error` column,
which meant the maximum length of the error column was limited to
~2.7kb. This lead to a StatementInvalid exception being raised.
* Because the StatementInvalid exception was raised while we were trying
to set the upload media asset's status to `failed`, the upload was
left stuck in the `processing` state rather than being set to the
`failed` state.
* Because the upload was stuck in the `processing` state, the upload
page would hang forever waiting for the upload to complete.
The fixes are to:
* Build FFmpeg with `--enable-libdav1d` to use libdav1d for decoding AV1
videos instead of the builtin AV1 decoder.
* Remove the index on the `upload_media_assets.error` column so that
setting overly long error messages won't fail.
* Catch unexpected exceptions in ProcessUploadMediaAssetJob so we can
mark uploads as failed, even if `process_upload!` itself fails because
it raises an unexpected exception inside its own exception handler.
* Check that the video is playable with `MediaFile::Video#is_corrupt?` before
allowing it to be uploaded. This way we can return a better error
message if we can't generate thumbnails because the video isn't
playable. This requires decoding the entire video, so it means uploads
may take several seconds longer for long videos. It's also a security
risk in case ffmpeg has any bugs.
* Define `MediaAsset#preview!` as raising an exception on error, so
it's clear that generating thumbnails can fail. Define `MediaAsset#preview`
as returning nil on error for when we don't care about the cause of
the error.
Add ability to upload .webp images.
Animated WebP images aren't supported. This is because they aren't
supported by FFmpeg yet[1], so generating thumbnails and samples for
them would be more complicated than for other formats.
[1]: https://trac.ffmpeg.org/ticket/4907
Features of AVIF include:
* Lossless and lossy compression.
* High dynamic range (HDR) images
* Wide color gamut images (i.e. 10- and 12-bit color depths)
* Transparency (through alpha planes).
* Animations (with an optional cover image).
* Auxiliary image sequences, where the file contains a single primary
image and a short secondary video, like Apple's Live Photos.
* Metadata rotation, mirroring, and cropping.
The AVIF format is still relatively new and some of these features aren't well
supported by browsers or other software:
* Animated AVIFs aren't supported by Firefox or by libvips.
* HDR images aren't supported by Firefox.
* Rotated, mirrored, and cropped AVIFs aren't supported by Firefox or Chrome.
* Image grids, where the file contains multiple images that are tiled
together into one big image, aren't supported by Firefox.
* AVIF as a whole has only been supported for a year or two by Chrome
and Firefox, and less than a year by Safari.
For these reasons, only basic AVIFs that don't use animation, rotation,
cropping, or image grids can be uploaded.
Remove the last remaining uses of the PixivUgoiraFrameData model. As of
32bfb8407, Ugoira frame data is now stored in the MediaMetadata model,
under the `Ugoira:FrameDelays` EXIF field.
The pixiv_ugoira_frame_data table still exists, but it can be removed
after this commit is deployed.
Fixes#5264: Error when replacing with ugoira.
Store Ugoira frame delays in the MediaMetadata model as a fake EXIF
field instead of in the PixivUgoiraFrameData model. This way we can get
rid of the PixivUgoiraFrameData model completely. This is a step towards
fixing #5264.
Fix `Cannot write log file 'ffmpeg2pass-0.log' for pass-1 encoding: Permission denied` error
when uploading ugoira files. Caused by the fact that 2-pass encoding tries to write a log file in
the current directory by default, which fails in production because the default working directory in
the Docker image is /danbooru, which is read-only.
Fix certain ugoiras having very low quality webm samples. This was
because we had a target bitrate of 5 Mbps, but this wasn't enough for
videos that were high resolution or that had choppy, hard-to-compress
motion, such as post 5081776 (nsfw).
This is used to provide higher resolution thumbnails for high pixel
density displays, such as phones or laptops. If your screen has a 2x
pixel density ratio, then 360x360 thumbnails will be rendered at 720x720
resolution.
We use WebP here because it's about 15% smaller than the equivalent
JPEG, and because if a device has a high enough pixel density to use
this, then it probably supports WebP.
720x720 thumbnails average about 36kb in size, compared to 20.35kb for
360x360 thumbnails and 7.55kb for 180x180 thumbnails.
These other formats aren't actually generated during upload, but support
for creating them is there.
Also tune the parameters for generating JPEGs:
* Use Q=85 instead of Q=90 because Q=85 enables 4:2:0 chroma
subsampling, while Q=90 doesn't use subsampling. Subsampling reduces
filesize by ~30% in most cases. It does reduce quality for certain
images, particularly for images with lots of bright red, but in most
cases the quality difference isn't noticeable.
* Enable several MozJPEG-specific options, including trellis
quantization and scan optimization. These reduce filesize without
reducing quality, at the cost of slower encoding times.
Calculate the dimensions of thumbnails ourselves instead of letting
libvips calculate them for us. This way we know the exact size of
thumbnails, so we can set the right width and height for <img> tags. If
we let libvips calculate thumbnail sizes for us, then we can't predict
the exact size of thumbnails, because sometimes libvips rounds numbers
differently than us.
The problem was that we were stripping color profiles from thumbnails,
but we weren't setting `export_profile: "srgb"` to convert images to
sRGB first. This resulted in wrong colors for images with non-sRGB color
profiles, such as Adobe RGB.
The fix is to convert images to sRGB when possible, while leaving CMYK
and greyscale images alone. We leave CMYK images alone because we can't
convert CMYK to sRGB without losing color. We leave greyscale images
alone if they don't have a color profile, that way they stay as
one-channel greyscale (or two-channel greyscale, in case of alpha)
instead of being converted to three-channel sRGB. However, if a
greyscale image has a color profile, then we have to convert to sRGB,
otherwise the colors would be wrong when we strip the profile.
We also have to set the import profile, otherwise images with broken
embedded color profiles won't have a fallback profile and may get
incorrect colors. In this case we also have to be careful, because we
can't specify an sRGB fallback for greyscale or CMYK images.
* Make it so replacing a post doesn't generate a dummy upload as a side effect.
* Make it so you can't replace a post with itself (the post should be regenerated instead).
* Refactor uploads and replacements to save the ugoira frame data when
the MediaAsset is created, not when the post is created. This way it's
possible to view the ugoira before the post is created.
* Make `download_file!` in the Pixiv source strategy return a MediaFile
with the ugoira frame data already attached to it, instead of returning it
in the `data` field then passing it around separately in the `context`
field of the upload.
Fix how the duration of videos and animated GIFs / PNGs is calculated.
If we can't determine the duration from the file metadata, then play the
entire video or animation back using FFmpeg and scrape the duration and
frame count.
This is necessary for things like WebM files where the duration metadata
is optional, or animated GIFs and PNGs that don't have a duration field
in the metadata, only a frame count and a sequence of frame delays.
Fix certain animated PNGs returning NaN as the duration because the
frame rate was being reported as "0/0" by FFMpeg. This happens when the
animation has zero delay between frames. This is supposed to mean a PNG
with an infinitely fast frame rate, but in practice browsers limit it to
around 10FPS. The exact frame rate browsers will use is unknown and
implementation defined.
Add methods to MediaFile to calculate the duration, frame count, and
frame rate of animated GIFs, PNGs, Ugoiras, and videos.
Some considerations:
* It's possible to have a GIF or PNG that's technically animated but
just has one frame. These are treated as non-animated images.
* It's possible to have an animated GIF that has an unspecified
frame rate. In this case we assume the frame rate is 10 FPS; this is
browser dependent and may not be correct.
* Animated GIFs, PNGs, and Ugoiras all support variable frame rates.
Technically, each frame has a separate delay, and the delays can be
different frame-to-frame. We report only the average frame rate.
* Getting the duration of an APNG is surprisingly hard. Most tools don't
have good support for APNGs since it's a rare and non-standardized
format. The best we can do is get the frame count using ExifTool and the
frame rate using ffprobe, then calculate the duration from that.
Fix a bug where where PNG images could be incorrectly detected as
exif-rotated. This would happen when a PNG contained the
IFD0:Orientation flag. It's technically possible for a PNG to contain
this flag, but it's ignored by libvips and by browsers.
post #3762340 (nsfw) is an example of a PNG like this.
The fix is to use `autorot` to let libvips apply the rotation instead of
trying to interpret the exif data ourselves. Note that libvips-8.9 has a
bug where it doesn't strip the orientation flag after applying
`autorot`, which leads to the image being incorrectly rotated a second
time when generating the thumbnail. Use libvips-8.11 instead.
Rotate the image based on the EXIF orientation flag when generating
thumbnails and samples.
Also fix the width and height to be calculated correctly for rotated
images. Vips gives us the unrotated width and height of the image; we
have to detect whether the image is rotated and swap the width and
height manually to correct them. For example, if an image with the
"Rotate 90 CW" flag is 100x500 before rotation, then after rotation it's
500x100. This should fix#4883 (Exif rotation breaks Javascript fit-to-window)
We also have to fix it so that regenerating a post updates the width and
height of the post, in the event that it's a rotated image.
Finally we set `image-orientation: from-image;` even though it's
probably not necessary.
Remove code for working with older versions of libvips. This makes
libvips 8.10+ a hard requirement. Older versions were already broken and
failed certain tests in the test suite.
Fix a bug where generating thumbnails failed for certain images when
using libvips 8.10. Specifically, it failed for single-channel greyscale
images and four-channel CMYK images without an embedded color profile.
In these cases we specified an sRGB fallback profile, but under libvips
8.10 this failed because the sRGB profile was incompatible with
single-channel and four-channel images. Before libvips 8.10 this worked,
but as of 8.10 it's a hard error.
The way libvips handles fallback color profiles differs across versions,
so we have to use different arguments for different versions. In 8.7,
vips doesn't have builtin color profiles, so we have to specify our own
manually. In 8.9, it has builtin profiles, so we can omit the import
profile, but we're still required to set the export profile to sRGB,
otherwise it will leave CMYK images as CMYK when generating thumbnails.
In 8.10, we have to _not_ to set the import or export profile to sRGB,
otherwise it will fail with an incompatible profile error when it tries
to convert CMYK images to RGB.
The builtin sRGB profile used by libvips[1] is different than the one we
used previously[2]. The builtin one comes from LCMS[3], whereas ours
came from ArgyllCMS.[4] Not all sRGB profiles are created the same[5],
so this may result in some imperceptible differences in thumbnail
output. The ArgyllCMS profile was used before because it seemed to be
the best one[6], but realistically it probably doesn't matter.
1: https://github.com/libvips/libvips/blob/v8.10.6/libvips/colour/profiles/sRGB.icm
2: 906eec190d/config/sRGB.icm
3: https://www.littlecms.com/
4: https://www.argyllcms.com/
5: https://ninedegreesbelow.com/photography/srgb-profile-comparison.html
6: https://ninedegreesbelow.com/photography/srgb-profile-comparison.html#addendum
Fix regression in ef2857667 that caused animated GIFs and PNGs to
generate thumbnails that were larger than 150x150.
Also fix a bug with cropped previews not being generated for animated
GIFs and PNGs.
flash files can be quite big (the biggest on danbooru.donmai.us being
68.6MB atm). Reading it and applying complex transformations twice seems
unnecessary.
MediaFile#dimensions is called twice - in #width and in #height but
it only works on the first call because the file is read to the end and
consumed the first time so when #read is called the second time it only
returns the empty string
