v7.42.0

What's Changed

Framework

• Added list-based image-conversion callback dispatch on mlt_frame.
  Use mlt_frame_append_convert_image() to register a converter instead of
  setting frame->convert_image directly. Multiple converters are tried in
  registration order; the first that succeeds wins. New public API:
  • mlt_frame_prepend_convert_image()
  • mlt_frame_append_convert_image()
  • mlt_frame_convert_image()
  • mlt_frame_next_convert_image()
  • mlt_frame_has_convert_image()
  • mlt_frame_copy_convert_image()
• API behavior change: mlt_frame_s::convert_image is now read-only after
  init and is permanently set to the dispatcher mlt_frame_convert_image().
  External code that set this field directly to a custom function will no
  longer have that function called. If you really need that (never heard of
  someone who does), clear the frame's "_convert_image_callbacks" property
  and use mlt_frame_append_convert_image() instead.
• fx_cut (Adjustment Clip) now applies its filters only to tracks
  below it. Higher tracks composite on top of the filtered result.
  New image-stack helpers used by tractor stacking and this path:
  • mlt_frame_prepend_image_from_service()
  • mlt_frame_push_image_with_fx_cut()
• Fixed crash applying a Movit (OpenGL) filter to an fx_cut (Adjustment Clip).
  Cross-producer image sharing (including the reverse fx_cut share) finalizes
  an in-progress Movit chain to rgba64 and does not copy producer-keyed
  _movit chain state. Same-producer prepends keep the GPU chain. Playlist
  fx_cut frames now set _producer so Movit can key the chain.
• Copy meta. and set. properties from the producer onto fx_cut frames
  so producer annotations reach tractor output. New public API:
  • mlt_producer_pass_frame_properties()
  • Mlt::Producer::pass_frame_properties()
• Changed the default audio sample format to float32 (mlt_audio_f32le).
  Added mlt_audio_format_id() and an mlt_audio_format property on the
  multi consumer. Integer/float conversion is lossless and symmetric
  (s16 and u8 round-trips are exact).
• Deprecated audio services that only support integer samples.
• Propagate per-track audio-level metadata to output frames (meta.audio.*).
• Improved mlt_property_is_anim() so strings that merely contain =
  (for example URLs) are not treated as animation.
• mlt_properties_get_data() no longer returns a data pointer for rect
  properties.
• Fixed signed integer overflow in mlt_luma_map_from_pgm.
• Fixed fx_cut in-point handling in mlt_transition.
• Fixed Android mlt++ builds by disabling ELF symbol versioning.
• Fixed a macOS crash in numeric property conversion caused by setlocale.
  Parsing and formatting no longer change the process LC_NUMERIC.

Modules

• Image converters (movit.convert, avcolor_space, imageconvert) are
  now attached data-driven via loader.ini key image_convert.
• Movit rgba64 is now a valid input as well as output: FlatInput is
  created with GL_UNSIGNED_SHORT and set_pixel_data() uses the 16-bit
  overload. Previously rgba64 was uploaded as 8-bit, which distorted the
  image when an OpenGL filter ran on an fx_cut.
• Added placebo GPU module (placebo.render, placebo.shader) using
  libplacebo for scaling, debanding, dithering, tonemapping, and mpv-style
  .hook shaders (D3D11 / Vulkan / OpenGL).
• Fixed deadlock in the decklink consumer when pausing, seeking, and
  resuming playback.
• Added ducking to the mix transition (duck_threshold, duck_attenuation,
  duck_level, duck_fade_in, duck_fade_out, and an optional prefix for
  duck_level on the frame).
• Added #frame+1# keyword in dynamictext (1-based frame number).
• Added more audio formats to the audiolevel filter.
• FFmpeg 9 support in avformat.
• Fixed 10-bit hevc_vaapi encoding.
• Log when the avformat consumer finishes.
• Fixed rnnoise only working with 48 kHz audio.
• frei0r.kaleid0sc0pe and frei0r.squigglevision report internal threading.
• Fixed frei0r.camerashake (and its blur) with preview scaling.
• Skip duplicate OpenFX plugins.
• USE_LV2 and USE_VST2 require USE_LADSPA; install LADSPA
  blacklist.txt without JACK; expand LV2_PATH correctly on Windows.
• Fixed affine and qtblend inside mask_apply.
• qtblend: flag to avoid double PAR normalization.
• Fixed luma transition alpha scaling, alpha when using a luma file, and
  documented the invert property.
• Fixed shape filter alpha subtraction for rgba.
• Fixed motion-tracker shapes clipping outside the frame.
• Fixed colorspace and range of cached images in qimage and pixbuf.
• Fixed crop with hardware decode+scale plus preview scale.
• Fixed subtitle filter rendering one frame past the end time, and
  subtitle_feed end-frame calculation.
• Fixed file-name buffer overflows in gpsgraphic and gpstext.
• Fixed the sdl2 consumer freezing on Windows after a few seconds.
• Fixed sdl2_audio hanging in consumer_stop() when stopped while paused.
• Added s32, float, s32le, f32le, and u8 to the noise producer,
  and s16, s32, s32le, f32le, and u8 to tone. Fixed tone
  integer conversion to match audioconvert.
• Fixed a double-free in the xml producer when a filter, transition,
  link, or producer fails to load.
• Fixed movit.convert leaking a Movit input on every frame when an
  effect chain is reused.
• Fixed the heart-rate keyword in gpstext (#gps_hr#).
• gpsgraphic no longer sets time_offset when a GPS file is first loaded.

Other

• melt prints progress at the end and exits non-zero when interrupted
  by a signal.
• Fixed melt on Windows ignoring window resize and keyboard transport.
• Fixed melt -jack option not working with PipeWire.

New Contributors

• @Magnussmari made their first contribution in #1270
• @doruphin made their first contribution in #1279
• @michaeljtbrooks made their first contribution in #1286
• @zhuowei made their first contribution in #1293
• @klg90 made their first contribution in #1307
• @darkdi made their first contribution in #1309
• @iDoMeteor made their first contribution in #1323

Full Changelog: v7.40.0...v7.42.0