Changes in version 0.1.3 (2026-03-23)                  

TIFF scaling

  - New tiff_scale parameter in av1r_options() for explicit control of
    TIFF/image sequence output resolution. Supports multiplier (e.g. 2
    for 2x) or bounding box (e.g. c(1920, 1080) to fit into 1080p).
    Aspect ratio is always preserved — no distortion.
  - Frames smaller than hardware minimum (VAAPI: 320x128) are
    automatically scaled up proportionally across all backends (CPU,
    VAAPI, Vulkan).
  - Example script: inst/examples/tiff_scale.R.

Batch conversion

  - New example script inst/examples/batch_convert.R for batch folder
    conversion with summary table (input/output size, compression
    ratio).
  - tiff_crf parameter (default 5) automatically applied to TIFF/image
    input for near-lossless microscopy encoding. Overrides crf.
  - verbose option controls informational messages.

Quality control architecture

  - Each backend now uses the approach best suited to its rate control:
      - CPU (SVT-AV1): CRF passed directly — SVT-AV1 adapts to content
      - Vulkan: CRF passed directly as QP — CQP mode (RADV limitation)
      - VAAPI: 55% of input video bitrate via ffprobe — adapts to
        content
  - Removed calibration table approach (0.1.3-dev): empirical testing
    showed fixed CRF→bitrate tables are content-dependent and break on
    different input formats (MP4 vs AVI). Adaptive approach is simpler
    and more robust.
  - Vulkan CQP limitation documented: RADV only supports constant
    quantizer (no VBR/CRF), resulting in ~0.93 SSIM vs ~0.96 for
    CPU/VAAPI at the same bitrate. This is a driver limitation, not a
    bug.

CPU encoding fixes

  - SVT-AV1 preset and thread count now passed via -svtav1-params
    instead of -preset and -threads which had no effect on libsvtav1.
  - Added -pix_fmt yuv420p for explicit pixel format.

Audio preservation

  - CPU path now preserves audio tracks from input files (-c:a copy)
    instead of stripping them (-an). TIFF inputs still use -an since
    they have no audio.

Benchmark improvements

  - Added "vs CPU" column showing speedup relative to CPU backend.
  - Added calibration scripts: calibrate_crf.R, calibrate_raw.R,
    calibrate_vaapi_bitrate.R for measuring backend response curves.

                 Changes in version 0.1.2 (2026-03-05)                  

Minimum coded extent handling

  - GPU encoder now queries minCodedExtent from the Vulkan driver before
    opening the ffmpeg decode pipe. Input frames smaller than the
    hardware minimum (e.g. 170×196 vs driver minimum 320×128) are
    automatically scaled up to the minimum supported resolution.
  - Vulkan context is initialized before the ffmpeg pipe so that the
    corrected dimensions are used for both decoding and encoding.
  - Fixed segfault in av1r_vulkan_query_min_extent(): the
    VkVideoEncodeAV1CapabilitiesKHR structure was missing from the pNext
    chain passed to vkGetPhysicalDeviceVideoCapabilitiesKHR, causing
    RADV to write to an unmapped address.

CRF / QIndex mapping

  - CRF-to-QIndex mapping changed from linear (crf * 4) to nonlinear LUT
    approximating aomenc --cq-level. Linear mapping gave qIndex 112 at
    CRF 28 which was too low for real compression on RADV. The new
    mapping gives qIndex 175 at CRF 28, matching expected AV1
    compression ratios.
  - On RADV, VBR mode is reported as supported but produces huge files;
    DISABLED mode (CQP) with constantQIndex is the only working path.

                        Changes in version 0.1.1                        

Vulkan AV1 GPU Encoding — fully working

  - GPU encoding now works end-to-end on AMD RDNA4 (RX 9070, RADV
    GFX1201) via Vulkan VK_KHR_VIDEO_ENCODE_AV1.
  - Two-queue architecture: transfer queue (buffer→image upload) +
    encode queue (video encoding) with semaphore synchronization. Fixes
    VK_ERROR_DEVICE_LOST on GPUs where the encode queue has no TRANSFER
    capability.
  - Correct NV12 upload: Y plane and UV plane copied as separate
    vkCmdCopyBufferToImage calls with proper bufferRowLength for each
    plane.
  - Sequence Header OBU automatically prepended via
    vkGetEncodedVideoSessionParametersKHR — output plays in all players.
  - Rate control: DISABLED mode (CQP) with constantQIndex derived from
    CRF (qIndex = crf * 4). VBR/CBR modes are reported as supported by
    RADV but have no effect; CQP is the only working mode.
  - Audio track from the original file is preserved during IVF→MP4
    muxing.
  - Dynamic loading of all Vulkan Video KHR extension functions via
    vkGetInstanceProcAddr/vkGetDeviceProcAddr for compatibility with SDK
    < 1.3.290.

Multi-page TIFF support

  - Multi-page TIFF stacks are now extracted to a temporary PNG sequence
    via the magick package before encoding. Both GPU and CPU paths
    support this.
  - magick added to Suggests (required only for multi-page TIFF input).
  - Image sequences (printf patterns like frame%06d.png) are passed to
    ffmpeg with -framerate for correct decoding.
  - Tested on 510-frame 16-bit grayscale confocal microscopy TIFF stack
    (32.7 MB → 3.8 MB at CRF 1, 8.6x compression with quality
    preservation).

CRAN Fixes

  - Removed redundant "A tool for" from Description.
  - Single-quoted software and API names in Title and Description per
    CRAN policy.
  - Replaced \dontrun{} with \donttest{} in all examples.
  - Examples now write to tempdir() instead of user's home filespace.

Bundled Headers

  - Bundled Khronos VK_KHR_video_encode_av1 headers (src/vk_video/) —
    package no longer requires Vulkan SDK >= 1.3.290 at build time;
    builds against any Vulkan SDK >= 1.3.275 with the bundled polyfill.
  - configure now detects bundled AV1 encode headers and enables GPU
    path automatically.

Initial Release

  - CPU encoding via FFmpeg (libsvtav1).
  - Automatic backend selection: GPU (Vulkan) → CPU (FFmpeg).
  - Supports H.264, H.265, AVI/MJPEG, TIFF stacks, and TIFF sequences as
    input.
  - convert_to_av1() for single-file and batch conversion.
  - detect_backend(), vulkan_available(), vulkan_devices() for
    diagnostics.