RUDRA v1.0.0-rc.2

RUDRA 0.9.0 beta 10

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the tenth public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG,
  ProRes 422 HQ or ProRes 4444 (the finishing master), with the audio,
  checked before it is published, through a queue you can stop and resume
  (ProRes 4444 with alpha from the command line:
  rudra-native video --format prores4444). The ProRes files come graded in
  PQ, or scene-referred in ACEScct (AP1) or ARRI LogC4 (AWG4): the
  master before peak and knee, to grade like camera footage. A log file is
  tagged unknown primaries and transfer (there are no codes for these
  curves); set the clip's input to the curve in Resolve. Its sidecar names
  the curve and gamut.
• Masters as ACES 2065-1 EXR sequences.
• The model (optional, off by default: Reconstruct > Use the model):
  sdr2hdr_image_v8, level with the inverse on real SDR, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 10 (for the next beta)

• Getting started, for 1.0. The welcome says the model is optional, and
  the five steps now cover the Source picker (BT.1886 by default), exposure,
  Fill clipped highlights, the HDR10+ and Dolby Vision metadata and Hold
  steady. They show once more after updating, then from Help > Getting started.

• The anchor follows the Source. With Anchor on (the default), the master
  kept mid-tones on the SDR read as an sRGB display at 203 nits, whatever the
  Source picker, calibration or reference said, so below the anchor knee a
  BT.1886 master came out as an sRGB one and a calibrated grey went back to
  where it was. The anchor now reads the SDR through the picked curve with its
  calibration or reference: BT.1886 masters are what was scored on real SDR,
  and a calibrated point keeps its nits. On the viewer too. Unknown and ACES
  with no calibration are unchanged, bit for bit.

• Exposure. Reconstruct has an Exposure slider, -3 to +3 stops: the whole
  master brighter or darker, live on the viewer, in the master, movie exports
  and projects. It is part of the grade: after the anchor (so mid-tones keep
  the offset), before the Region EV bands (which see the exposed nits).
  Ctrl+Shift+Up / Down move it by a third of a stop, double-click sets it
  back to 0; it is undone like any other control.

• A new default for SDR: BT.1886. Reconstruct > Source has a BT.1886 entry,
  the SDR read the way a display shows it (each BT.709 code to the power 2.4,
  white at 203 nits), and new sessions start on it. On real SDR video with its
  graded HDR it scored far above the old default, the ACES inverse: +14 dB and
  +2.3 JOD on Netflix's Meridian, +4.2 dB and +2.1 JOD on HDRTV1K (docs/RESULTS.md).
  CG renders and games made with a known tone curve still want ACES, filmic
  or AgX in the picker. Projects saved before keep the curve they were saved
  with, and the command line keeps the ACES inverse unless a request names
  "source_curve": "bt1886".

• Dolby Vision and HDR10+ metadata. HDR10 and ProRes (PQ) movie exports
  now write their dynamic HDR metadata per shot (the export's own cuts) beside
  the master: <file>_hdr10plus.json (HDR10+, in the format x265 and
  hdr10plus_tool read), <file>_dovi_generate.json (Dolby Vision L1 and L6,
  ready for dovi_tool generate), <file>_hdr_analysis.json (every value in
  nits and PQ) and <file>_hdr10plus_scenes.json. An HDR10 export carries
  HDR10+ inside the HEVC when the ffmpeg's x265 was built with HDR10+, and the
  quality check confirms it on the first frame; otherwise the report says so
  and the file is ready for hdr10plus_tool inject. RUDRA writes no Dolby
  Vision RPU itself: a licensed finishing tool, or dovi_tool, makes it from
  the file. --dynamic-metadata off turns it off. The dovi_generate file of
  rudra-native deliver and rudra metadata now uses dovi_tool's block format
  ({"Level1": {...}}), clamped as dovi_tool clamps; it was not accepted by
  dovi_tool before.

• Steadier highlights. Inside highlights the source has clipped (in this
  frame and the one before) the static hold now follows the HDR's own change:
  a few hundredths of an EV of model jitter settles, while a light that moves
  (a quarter of an EV or more) is each frame's own, so it leaves no trail.
  Before, nothing was held there.

• Playback no longer blends a late anchor measurement from before a loop or
  a cut into the frames after it.

• rudra batch (Python) runs the queues RUDRA writes: a job with the app's
  settings (grade) is handed to rudra-native video (found on PATH, or set
  RUDRA_NATIVE), and its report is checked as any other job's.

• Still areas play steady too. With Clip > Hold steady within a shot,
  playback now runs the static hold that movie exports run (on the GPU, after
  the composite): where the source did not change from the frame before, how
  much RUDRA brightened it is held across the shot, so a still shot no longer
  shimmers more while playing than in its export. It starts again at each cut,
  on a step back and when playback loops; a stopped frame shows its own.

• Measured on real footage (docs/TEMPORAL_RESULTS.md): with the model, a
  campfire flickered 0.03 to 0.05 EV a frame beyond the source (a visible
  0.19 EV at worst); with the holds, 0.006 to 0.013 EV. rudra-native video
  takes --grade <settings.json> (the app's export settings; - in place of
  the package for an analytic one), and tools/measure_temporal.py and
  tools/make_temporal_clips.py repeat the measurement.

Changes in beta 10 (since beta 9)

• Still areas stop shimmering. In a movie export, where the source pixel
  did not change from the frame before, RUDRA holds how much it brightened
  that pixel (the HDR over the source) across the shot, so codec noise lifted
  by the expansion and the model's jitter in recovered highlights settle;
  where anything moved, at a cut, and wherever the source is clipped or
  crushed, each frame is its own. The source's own change always passes
  through: only the brightening is held, never the pixel. On a still grainy
  ramp the worst frame-to-frame shimmer fell to under a fifth.
  --static-hold off turns it off (Clip > Hold steady within a shot off turns
  off both holds); it costs about 40 ms a 1080p frame on one CPU core.
• Movie reports add local_flicker_ev: the flicker measure taken area by
  area (a 36 x 64 grid) and averaged, which sees shimmer the frame average
  hides.
• The beta 9 notes below describe steady within a shot, which the beta 9
  build already had; they were written after it was tagged.

Changes in beta 9 (since beta 8)

• Movie exports carry your grade. HDR10, HLG and ProRes exports now go
  through the same chain as Master EXR, frame by frame: the reconstruction
  (analytic, or the model), recovery mode and strength, source curve,
  calibration or reference, painted masks, anchor, chroma, grain and Region
  EV. They used to run the model with recovery "all" and none of the grade.
  The settings travel beside the export as <file>.settings.json, and the
  report says which reconstruction made it. An analytic export needs no model
  package at all.
• Playback with the model. While playing, the model runs ahead of the
  playhead on a smaller copy of each frame (Clip > While playing: 960 px by
  default, 640 px, or full size) and its result is brought back to the frame's
  size; where you stop, the frame is reconstructed again at full size. A movie
  is decoded in order by one ffmpeg instead of frame by frame. Clip > While
  playing > Keep time skips frames the model has not reached, so playback runs
  at the clip's rate; off, every frame is shown. The status bar shows the rate
  reached. On the CPU the model is 4.2 times faster at 960 px than at 1080p
  and 7.6 times at 640 px; from the 154 ms per 1080p frame measured on an RTX
  4080 that is about 37 ms (27 fps) and 20 ms (50 fps), estimated, not yet
  measured on the GPU.
• Steady within a shot. The few values RUDRA works out for a whole frame
  (the model's shadow weight, residual scale and curve, and the anchor's hold)
  are now held steady across a shot instead of following each frame, so a
  grainy or busy clip no longer pumps in brightness. They follow the picture
  with a quarter-second time constant and start again at every cut. Movie
  ex...

RUDRA v1.0.0-rc.1

RUDRA 0.9.0 beta 10

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the tenth public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG,
  ProRes 422 HQ or ProRes 4444 (the finishing master), with the audio,
  checked before it is published, through a queue you can stop and resume
  (ProRes 4444 with alpha from the command line:
  rudra-native video --format prores4444). The ProRes files come graded in
  PQ, or scene-referred in ACEScct (AP1) or ARRI LogC4 (AWG4): the
  master before peak and knee, to grade like camera footage. A log file is
  tagged unknown primaries and transfer (there are no codes for these
  curves); set the clip's input to the curve in Resolve. Its sidecar names
  the curve and gamut.
• Masters as ACES 2065-1 EXR sequences.
• The model (optional, off by default: Reconstruct > Use the model):
  sdr2hdr_image_v8, level with the inverse on real SDR, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 10 (for the next beta)

• The anchor follows the Source. With Anchor on (the default), the master
  kept mid-tones on the SDR read as an sRGB display at 203 nits, whatever the
  Source picker, calibration or reference said, so below the anchor knee a
  BT.1886 master came out as an sRGB one and a calibrated grey went back to
  where it was. The anchor now reads the SDR through the picked curve with its
  calibration or reference: BT.1886 masters are what was scored on real SDR,
  and a calibrated point keeps its nits. On the viewer too. Unknown and ACES
  with no calibration are unchanged, bit for bit.

• Exposure. Reconstruct has an Exposure slider, -3 to +3 stops: the whole
  master brighter or darker, live on the viewer, in the master, movie exports
  and projects. It is part of the grade: after the anchor (so mid-tones keep
  the offset), before the Region EV bands (which see the exposed nits).
  Ctrl+Shift+Up / Down move it by a third of a stop, double-click sets it
  back to 0; it is undone like any other control.

• A new default for SDR: BT.1886. Reconstruct > Source has a BT.1886 entry,
  the SDR read the way a display shows it (each BT.709 code to the power 2.4,
  white at 203 nits), and new sessions start on it. On real SDR video with its
  graded HDR it scored far above the old default, the ACES inverse: +14 dB and
  +2.3 JOD on Netflix's Meridian, +4.2 dB and +2.1 JOD on HDRTV1K (docs/RESULTS.md).
  CG renders and games made with a known tone curve still want ACES, filmic
  or AgX in the picker. Projects saved before keep the curve they were saved
  with, and the command line keeps the ACES inverse unless a request names
  "source_curve": "bt1886".

• Dolby Vision and HDR10+ metadata. HDR10 and ProRes (PQ) movie exports
  now write their dynamic HDR metadata per shot (the export's own cuts) beside
  the master: <file>_hdr10plus.json (HDR10+, in the format x265 and
  hdr10plus_tool read), <file>_dovi_generate.json (Dolby Vision L1 and L6,
  ready for dovi_tool generate), <file>_hdr_analysis.json (every value in
  nits and PQ) and <file>_hdr10plus_scenes.json. An HDR10 export carries
  HDR10+ inside the HEVC when the ffmpeg's x265 was built with HDR10+, and the
  quality check confirms it on the first frame; otherwise the report says so
  and the file is ready for hdr10plus_tool inject. RUDRA writes no Dolby
  Vision RPU itself: a licensed finishing tool, or dovi_tool, makes it from
  the file. --dynamic-metadata off turns it off. The dovi_generate file of
  rudra-native deliver and rudra metadata now uses dovi_tool's block format
  ({"Level1": {...}}), clamped as dovi_tool clamps; it was not accepted by
  dovi_tool before.

• Steadier highlights. Inside highlights the source has clipped (in this
  frame and the one before) the static hold now follows the HDR's own change:
  a few hundredths of an EV of model jitter settles, while a light that moves
  (a quarter of an EV or more) is each frame's own, so it leaves no trail.
  Before, nothing was held there.

• Playback no longer blends a late anchor measurement from before a loop or
  a cut into the frames after it.

• rudra batch (Python) runs the queues RUDRA writes: a job with the app's
  settings (grade) is handed to rudra-native video (found on PATH, or set
  RUDRA_NATIVE), and its report is checked as any other job's.

• Still areas play steady too. With Clip > Hold steady within a shot,
  playback now runs the static hold that movie exports run (on the GPU, after
  the composite): where the source did not change from the frame before, how
  much RUDRA brightened it is held across the shot, so a still shot no longer
  shimmers more while playing than in its export. It starts again at each cut,
  on a step back and when playback loops; a stopped frame shows its own.

• Measured on real footage (docs/TEMPORAL_RESULTS.md): with the model, a
  campfire flickered 0.03 to 0.05 EV a frame beyond the source (a visible
  0.19 EV at worst); with the holds, 0.006 to 0.013 EV. rudra-native video
  takes --grade <settings.json> (the app's export settings; - in place of
  the package for an analytic one), and tools/measure_temporal.py and
  tools/make_temporal_clips.py repeat the measurement.

Changes in beta 10 (since beta 9)

• Still areas stop shimmering. In a movie export, where the source pixel
  did not change from the frame before, RUDRA holds how much it brightened
  that pixel (the HDR over the source) across the shot, so codec noise lifted
  by the expansion and the model's jitter in recovered highlights settle;
  where anything moved, at a cut, and wherever the source is clipped or
  crushed, each frame is its own. The source's own change always passes
  through: only the brightening is held, never the pixel. On a still grainy
  ramp the worst frame-to-frame shimmer fell to under a fifth.
  --static-hold off turns it off (Clip > Hold steady within a shot off turns
  off both holds); it costs about 40 ms a 1080p frame on one CPU core.
• Movie reports add local_flicker_ev: the flicker measure taken area by
  area (a 36 x 64 grid) and averaged, which sees shimmer the frame average
  hides.
• The beta 9 notes below describe steady within a shot, which the beta 9
  build already had; they were written after it was tagged.

Changes in beta 9 (since beta 8)

• Movie exports carry your grade. HDR10, HLG and ProRes exports now go
  through the same chain as Master EXR, frame by frame: the reconstruction
  (analytic, or the model), recovery mode and strength, source curve,
  calibration or reference, painted masks, anchor, chroma, grain and Region
  EV. They used to run the model with recovery "all" and none of the grade.
  The settings travel beside the export as <file>.settings.json, and the
  report says which reconstruction made it. An analytic export needs no model
  package at all.
• Playback with the model. While playing, the model runs ahead of the
  playhead on a smaller copy of each frame (Clip > While playing: 960 px by
  default, 640 px, or full size) and its result is brought back to the frame's
  size; where you stop, the frame is reconstructed again at full size. A movie
  is decoded in order by one ffmpeg instead of frame by frame. Clip > While
  playing > Keep time skips frames the model has not reached, so playback runs
  at the clip's rate; off, every frame is shown. The status bar shows the rate
  reached. On the CPU the model is 4.2 times faster at 960 px than at 1080p
  and 7.6 times at 640 px; from the 154 ms per 1080p frame measured on an RTX
  4080 that is about 37 ms (27 fps) and 20 ms (50 fps), estimated, not yet
  measured on the GPU.
• Steady within a shot. The few values RUDRA works out for a whole frame
  (the model's shadow weight, residual scale and curve, and the anchor's hold)
  are now held steady across a shot instead of following each frame, so a
  grainy or busy clip no longer pumps in brightness. They follow the picture
  with a quarter-second time constant and start again at every cut. Movie
  exports and playback do it; a stopped frame and Master EXR show the frame's
  own values. Clip > Hold steady within a shot turns it off;
  rudra-native video and rudra video take --temporal-smoothing SECONDS
  (0 is off). Every movie report now measures flicker: flicker_ev per frame
  and a `temporal...

RUDRA v0.9.0-beta.10

RUDRA 0.9.0 beta 10

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the tenth public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG,
  ProRes 422 HQ or ProRes 4444 (the finishing master), with the audio,
  checked before it is published, through a queue you can stop and resume
  (ProRes 4444 with alpha from the command line:
  rudra-native video --format prores4444). The ProRes files come graded in
  PQ, or scene-referred in ACEScct (AP1) or ARRI LogC4 (AWG4): the
  master before peak and knee, to grade like camera footage. A log file is
  tagged unknown primaries and transfer (there are no codes for these
  curves); set the clip's input to the curve in Resolve. Its sidecar names
  the curve and gamut.
• Masters as ACES 2065-1 EXR sequences.
• The model (optional, off by default: Reconstruct > Use the model):
  sdr2hdr_image_v8, level with the inverse on real SDR, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 9

• Still areas stop shimmering. In a movie export, where the source pixel
  did not change from the frame before, RUDRA holds how much it brightened
  that pixel (the HDR over the source) across the shot, so codec noise lifted
  by the expansion and the model's jitter in recovered highlights settle;
  where anything moved, at a cut, and wherever the source is clipped or
  crushed, each frame is its own. The source's own change always passes
  through: only the brightening is held, never the pixel. On a still grainy
  ramp the worst frame-to-frame shimmer fell to under a fifth.
  --static-hold off turns it off (Clip > Hold steady within a shot off turns
  off both holds); it costs about 40 ms a 1080p frame on one CPU core.
• Movie reports add local_flicker_ev: the flicker measure taken area by
  area (a 36 x 64 grid) and averaged, which sees shimmer the frame average
  hides.
• The beta 9 notes below describe steady within a shot, which the beta 9
  build already had; they were written after it was tagged.

Changes in beta 9 (since beta 8)

• Movie exports carry your grade. HDR10, HLG and ProRes exports now go
  through the same chain as Master EXR, frame by frame: the reconstruction
  (analytic, or the model), recovery mode and strength, source curve,
  calibration or reference, painted masks, anchor, chroma, grain and Region
  EV. They used to run the model with recovery "all" and none of the grade.
  The settings travel beside the export as <file>.settings.json, and the
  report says which reconstruction made it. An analytic export needs no model
  package at all.
• Playback with the model. While playing, the model runs ahead of the
  playhead on a smaller copy of each frame (Clip > While playing: 960 px by
  default, 640 px, or full size) and its result is brought back to the frame's
  size; where you stop, the frame is reconstructed again at full size. A movie
  is decoded in order by one ffmpeg instead of frame by frame. Clip > While
  playing > Keep time skips frames the model has not reached, so playback runs
  at the clip's rate; off, every frame is shown. The status bar shows the rate
  reached. On the CPU the model is 4.2 times faster at 960 px than at 1080p
  and 7.6 times at 640 px; from the 154 ms per 1080p frame measured on an RTX
  4080 that is about 37 ms (27 fps) and 20 ms (50 fps), estimated, not yet
  measured on the GPU.
• Steady within a shot. The few values RUDRA works out for a whole frame
  (the model's shadow weight, residual scale and curve, and the anchor's hold)
  are now held steady across a shot instead of following each frame, so a
  grainy or busy clip no longer pumps in brightness. They follow the picture
  with a quarter-second time constant and start again at every cut. Movie
  exports and playback do it; a stopped frame and Master EXR show the frame's
  own values. Clip > Hold steady within a shot turns it off;
  rudra-native video and rudra video take --temporal-smoothing SECONDS
  (0 is off). Every movie report now measures flicker: flicker_ev per frame
  and a temporal summary, the EV the HDR's average moved from one frame to
  the next beyond what the source did.

Changes in beta 8 (since beta 7)

• The analytic reconstruction is the default, and says so. Until now the
  app ran the model on every frame and every master although these notes
  called it optional. Now the view and Master EXR use the analytic inverse
  (your source curve, calibration, reference and grade) and run no model at
  all; RUDRA opens, views and masters with no model package installed.
  Reconstruct > Use the model (AI-assisted) turns the model on; the status
  bar says which is in use, and a master's sidecar records it
  ("reconstruction": "analytic (no model)", or the model's name under
  checkpoint). The choice is kept between runs and in a project.

Changes in beta 7 (since beta 6)

Fixes that bring the native app in line with the 9 Oct review of the
Studio's master path:

• Region EV lands in the master with Anchor on. The anchor (on by
  default) used to divide a Region EV push below its knee straight back out,
  in the viewer and in the master, while the sidecar said the grade was
  applied. The master now grades after the anchor, chroma and grain stages,
  and the viewer shows the same.
• The linear EXR container is really Rec.2020. It carried the plate's
  primaries (Rec.709 for most footage) under a "scene-linear Rec.2020" label
  with no chromaticities; it is now converted and tagged.
• rudra-native deliver, HLG: the BT.2100 system gamma (log10; it was
  right at 1,000 nits only) and saturated colours scaled together instead of
  clipping per channel (no hue shift). Footage in Rec.709 or P3 is converted
  to Rec.2020 before the roll-off, so MaxCLL/MaxFALL describe what is encoded.
• No half-written movies: deliver encodes under a hidden partial name
  and renames it into place, so a failed encode leaves nothing that looks like
  a delivery.

Changes in beta 6 (since beta 5)

• Projects. File > Save project (Ctrl+S) keeps the shot, the frame, the
  model and the whole grade (mode, strength, Region EV, source curve,
  calibration, reference, painted masks, view peak, Deliver settings) in a
  .rudra file; File > Open project (Ctrl+O), a drop, or on Windows a
  double-click opens it. Footage moved together with the project is found
  again; anything missing is named and the rest still opens.
• Autosave. The session is saved a few seconds after each change and on
  close; File > Reopen last session brings back the one the last run left.
• ffmpeg in the Windows package (the gyan.dev "full" build, GPL v3, in
  ffmpeg\ with its licence and source note): movies work without
  installing anything. RUDRA_FFMPEG_DIR picks another build.
• Update check. Once a day RUDRA asks GitHub whether a newer release is
  out and says so in the status bar; Help > Check for updates asks now.
  Nothing is downloaded. Off with RUDRA_NO_UPDATE_CHECK=1.
• Help for a first session: Help > Getting started (the five steps,
  shown once after the first-run check), Help > User guide (F1, also in the package
  as "User guide.md"), Help > Open log folder, and tooltips on every control
  of the Reconstruct, Grade and Deliver tabs.
• Signing hooks: the Windows build signs RUDRA.exe, rudra-native.exe and
  the installer when a code-signing certificate is supplied (the release
  workflow's WINDOWS_CERT_PFX secret).
• Open recent opens a movie as a shot again (it was added as a still).

Changes in beta 5 (since beta 4)

• Starts the same on every machine. Qt settings left in the environment by
  other software (a conda env with PyQt, a Qt SDK) no longer stop RUDRA with
  "no Qt platform plugin could be initialized": the package uses its own
  plugins and the log says what it ignored.
• A log file, rotated at 5 MB: on Windows
  %LOCALAPPDATA%\FXTD Studios\RUDRA\logs\rudra.log, on macOS
  ~/Library/Application Support/FXTD Studios/RUDRA/logs/rudra.log. Attach it
  to a report.
• Failed writes say why: "The EXR file could not be written (the disk is
  full: 1.2 GB free on D:, 3.4 GB needed)", access denied (with a pointer to
  Windows Security's Controlled folder access), or a missing folder. A master
  cut short by a full disk is removed rather than left truncated.

New in...

RUDRA v0.9.0-beta.9

RUDRA 0.9.0 beta 9

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the ninth public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG,
  ProRes 422 HQ or ProRes 4444 (the finishing master), with the audio,
  checked before it is published, through a queue you can stop and resume
  (ProRes 4444 with alpha from the command line:
  rudra-native video --format prores4444). The ProRes files come graded in
  PQ, or scene-referred in ACEScct (AP1) or ARRI LogC4 (AWG4): the
  master before peak and knee, to grade like camera footage. A log file is
  tagged unknown primaries and transfer (there are no codes for these
  curves); set the clip's input to the curve in Resolve. Its sidecar names
  the curve and gamut.
• Masters as ACES 2065-1 EXR sequences.
• The model (optional, off by default: Reconstruct > Use the model):
  sdr2hdr_image_v8, level with the inverse on real SDR, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 8

• Movie exports carry your grade. HDR10, HLG and ProRes exports now go
  through the same chain as Master EXR, frame by frame: the reconstruction
  (analytic, or the model), recovery mode and strength, source curve,
  calibration or reference, painted masks, anchor, chroma, grain and Region
  EV. They used to run the model with recovery "all" and none of the grade.
  The settings travel beside the export as <file>.settings.json, and the
  report says which reconstruction made it. An analytic export needs no model
  package at all.
• Playback with the model. While playing, the model runs ahead of the
  playhead on a smaller copy of each frame (Clip > While playing: 960 px by
  default, 640 px, or full size) and its result is brought back to the frame's
  size; where you stop, the frame is reconstructed again at full size. A movie
  is decoded in order by one ffmpeg instead of frame by frame. Clip > While
  playing > Keep time skips frames the model has not reached, so playback runs
  at the clip's rate; off, every frame is shown. The status bar shows the rate
  reached. On the CPU the model is 4.2 times faster at 960 px than at 1080p
  and 7.6 times at 640 px; from the 154 ms per 1080p frame measured on an RTX
  4080 that is about 37 ms (27 fps) and 20 ms (50 fps), estimated, not yet
  measured on the GPU.

Changes in beta 8 (since beta 7)

• The analytic reconstruction is the default, and says so. Until now the
  app ran the model on every frame and every master although these notes
  called it optional. Now the view and Master EXR use the analytic inverse
  (your source curve, calibration, reference and grade) and run no model at
  all; RUDRA opens, views and masters with no model package installed.
  Reconstruct > Use the model (AI-assisted) turns the model on; the status
  bar says which is in use, and a master's sidecar records it
  ("reconstruction": "analytic (no model)", or the model's name under
  checkpoint). The choice is kept between runs and in a project.

Changes in beta 7 (since beta 6)

Fixes that bring the native app in line with the 9 Oct review of the
Studio's master path:

• Region EV lands in the master with Anchor on. The anchor (on by
  default) used to divide a Region EV push below its knee straight back out,
  in the viewer and in the master, while the sidecar said the grade was
  applied. The master now grades after the anchor, chroma and grain stages,
  and the viewer shows the same.
• The linear EXR container is really Rec.2020. It carried the plate's
  primaries (Rec.709 for most footage) under a "scene-linear Rec.2020" label
  with no chromaticities; it is now converted and tagged.
• rudra-native deliver, HLG: the BT.2100 system gamma (log10; it was
  right at 1,000 nits only) and saturated colours scaled together instead of
  clipping per channel (no hue shift). Footage in Rec.709 or P3 is converted
  to Rec.2020 before the roll-off, so MaxCLL/MaxFALL describe what is encoded.
• No half-written movies: deliver encodes under a hidden partial name
  and renames it into place, so a failed encode leaves nothing that looks like
  a delivery.

Changes in beta 6 (since beta 5)

• Projects. File > Save project (Ctrl+S) keeps the shot, the frame, the
  model and the whole grade (mode, strength, Region EV, source curve,
  calibration, reference, painted masks, view peak, Deliver settings) in a
  .rudra file; File > Open project (Ctrl+O), a drop, or on Windows a
  double-click opens it. Footage moved together with the project is found
  again; anything missing is named and the rest still opens.
• Autosave. The session is saved a few seconds after each change and on
  close; File > Reopen last session brings back the one the last run left.
• ffmpeg in the Windows package (the gyan.dev "full" build, GPL v3, in
  ffmpeg\ with its licence and source note): movies work without
  installing anything. RUDRA_FFMPEG_DIR picks another build.
• Update check. Once a day RUDRA asks GitHub whether a newer release is
  out and says so in the status bar; Help > Check for updates asks now.
  Nothing is downloaded. Off with RUDRA_NO_UPDATE_CHECK=1.
• Help for a first session: Help > Getting started (the five steps,
  shown once after the first-run check), Help > User guide (F1, also in the package
  as "User guide.md"), Help > Open log folder, and tooltips on every control
  of the Reconstruct, Grade and Deliver tabs.
• Signing hooks: the Windows build signs RUDRA.exe, rudra-native.exe and
  the installer when a code-signing certificate is supplied (the release
  workflow's WINDOWS_CERT_PFX secret).
• Open recent opens a movie as a shot again (it was added as a still).

Changes in beta 5 (since beta 4)

• Starts the same on every machine. Qt settings left in the environment by
  other software (a conda env with PyQt, a Qt SDK) no longer stop RUDRA with
  "no Qt platform plugin could be initialized": the package uses its own
  plugins and the log says what it ignored.
• A log file, rotated at 5 MB: on Windows
  %LOCALAPPDATA%\FXTD Studios\RUDRA\logs\rudra.log, on macOS
  ~/Library/Application Support/FXTD Studios/RUDRA/logs/rudra.log. Attach it
  to a report.
• Failed writes say why: "The EXR file could not be written (the disk is
  full: 1.2 GB free on D:, 3.4 GB needed)", access denied (with a pointer to
  Windows Security's Controlled folder access), or a missing folder. A master
  cut short by a full disk is removed rather than left truncated.

New in beta 4 (since beta 3)

• Source panel. Name the curve that made the SDR (unknown, ACES,
  filmic/Hable, AgX, camera log, plain Rec.709 clip) and the baseline becomes
  its inverse, drawn against the ACES inverse with readouts in nits.
• Three-click calibration. Click black, 18% grey or a highlight and type
  the nits it should be; one to three anchors are fitted over the picked curve.
• Reference match. Load the graded HDR of the frame (EXR, or 16-bit PQ
  PNG/TIFF) and the curve is fitted to it at every code, with the residual in
  stops.
• Painted masks for the first four Region EV bands (left drag paints, Alt
  erases, the wheel sizes the brush); they travel with the master as
  <master>.masks.png.
• Anchor live on the viewer, with an Anchor knee slider: the master's
  anchor stage is what you see while Anchor is on.
• ProRes 4444 beside 422 HQ, and on both an Encoding row: graded PQ,
  or scene-referred ACEScct (AP1) or ARRI LogC4 (AWG4).
• The model is sdr2hdr_image_v8 (optional; masters default to the
  analytic inverse). shadow_v1 lost to the inverse on every real Meridian
  frame (−4.11 dB) and is retired. A bare start now follows the default when
  it changes, so an install upgraded from an earlier beta opens v8 instead of
  the package it used last; a model you picked yourself stays picked.
• Model parity on real 1080p frames: the ONNX graphs reduce GroupNorm in
  stages, so ONNX Runtime matches PyTorch to 9e-6 on all 287 Meridian test
  frames (it was 5.6e-3).
• Half precision: packages carry an fp16 tile graph, held to a 1% tolerance;
  --precision fp16 on rudra-native diff and bench, and a TensorRT device
  (--device tensorrt) where ONNX Runtime has it.
• Every live control (peak, region EV, strength, mode, preserve, source curve,
  calibration) is one com...

RUDRA v0.9.0-beta.8

RUDRA 0.9.0 beta 8

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the eighth public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG,
  ProRes 422 HQ or ProRes 4444 (the finishing master), with the audio,
  checked before it is published, through a queue you can stop and resume
  (ProRes 4444 with alpha from the command line:
  rudra-native video --format prores4444). The ProRes files come graded in
  PQ, or scene-referred in ACEScct (AP1) or ARRI LogC4 (AWG4): the
  master before peak and knee, to grade like camera footage. A log file is
  tagged unknown primaries and transfer (there are no codes for these
  curves); set the clip's input to the curve in Resolve. Its sidecar names
  the curve and gamut.
• Masters as ACES 2065-1 EXR sequences.
• The model (optional, off by default: Reconstruct > Use the model):
  sdr2hdr_image_v8, level with the inverse on real SDR, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 7

• The analytic reconstruction is the default, and says so. Until now the
  app ran the model on every frame and every master although these notes
  called it optional. Now the view and Master EXR use the analytic inverse
  (your source curve, calibration, reference and grade) and run no model at
  all; RUDRA opens, views and masters with no model package installed.
  Reconstruct > Use the model (AI-assisted) turns the model on; the status
  bar says which is in use, and a master's sidecar records it
  ("reconstruction": "analytic (no model)", or the model's name under
  checkpoint). The choice is kept between runs and in a project.
  Movie exports (HDR10, HLG, ProRes) still run the model package, whatever the
  switch says, and do not yet carry the grade: see Known in this beta.

Changes in beta 7 (since beta 6)

Fixes that bring the native app in line with the 9 Oct review of the
Studio's master path:

• Region EV lands in the master with Anchor on. The anchor (on by
  default) used to divide a Region EV push below its knee straight back out,
  in the viewer and in the master, while the sidecar said the grade was
  applied. The master now grades after the anchor, chroma and grain stages,
  and the viewer shows the same.
• The linear EXR container is really Rec.2020. It carried the plate's
  primaries (Rec.709 for most footage) under a "scene-linear Rec.2020" label
  with no chromaticities; it is now converted and tagged.
• rudra-native deliver, HLG: the BT.2100 system gamma (log10; it was
  right at 1,000 nits only) and saturated colours scaled together instead of
  clipping per channel (no hue shift). Footage in Rec.709 or P3 is converted
  to Rec.2020 before the roll-off, so MaxCLL/MaxFALL describe what is encoded.
• No half-written movies: deliver encodes under a hidden partial name
  and renames it into place, so a failed encode leaves nothing that looks like
  a delivery.

Changes in beta 6 (since beta 5)

• Projects. File > Save project (Ctrl+S) keeps the shot, the frame, the
  model and the whole grade (mode, strength, Region EV, source curve,
  calibration, reference, painted masks, view peak, Deliver settings) in a
  .rudra file; File > Open project (Ctrl+O), a drop, or on Windows a
  double-click opens it. Footage moved together with the project is found
  again; anything missing is named and the rest still opens.
• Autosave. The session is saved a few seconds after each change and on
  close; File > Reopen last session brings back the one the last run left.
• ffmpeg in the Windows package (the gyan.dev "full" build, GPL v3, in
  ffmpeg\ with its licence and source note): movies work without
  installing anything. RUDRA_FFMPEG_DIR picks another build.
• Update check. Once a day RUDRA asks GitHub whether a newer release is
  out and says so in the status bar; Help > Check for updates asks now.
  Nothing is downloaded. Off with RUDRA_NO_UPDATE_CHECK=1.
• Help for a first session: Help > Getting started (the five steps,
  shown once after the first-run check), Help > User guide (F1, also in the package
  as "User guide.md"), Help > Open log folder, and tooltips on every control
  of the Reconstruct, Grade and Deliver tabs.
• Signing hooks: the Windows build signs RUDRA.exe, rudra-native.exe and
  the installer when a code-signing certificate is supplied (the release
  workflow's WINDOWS_CERT_PFX secret).
• Open recent opens a movie as a shot again (it was added as a still).

Changes in beta 5 (since beta 4)

• Starts the same on every machine. Qt settings left in the environment by
  other software (a conda env with PyQt, a Qt SDK) no longer stop RUDRA with
  "no Qt platform plugin could be initialized": the package uses its own
  plugins and the log says what it ignored.
• A log file, rotated at 5 MB: on Windows
  %LOCALAPPDATA%\FXTD Studios\RUDRA\logs\rudra.log, on macOS
  ~/Library/Application Support/FXTD Studios/RUDRA/logs/rudra.log. Attach it
  to a report.
• Failed writes say why: "The EXR file could not be written (the disk is
  full: 1.2 GB free on D:, 3.4 GB needed)", access denied (with a pointer to
  Windows Security's Controlled folder access), or a missing folder. A master
  cut short by a full disk is removed rather than left truncated.

New in beta 4 (since beta 3)

• Source panel. Name the curve that made the SDR (unknown, ACES,
  filmic/Hable, AgX, camera log, plain Rec.709 clip) and the baseline becomes
  its inverse, drawn against the ACES inverse with readouts in nits.
• Three-click calibration. Click black, 18% grey or a highlight and type
  the nits it should be; one to three anchors are fitted over the picked curve.
• Reference match. Load the graded HDR of the frame (EXR, or 16-bit PQ
  PNG/TIFF) and the curve is fitted to it at every code, with the residual in
  stops.
• Painted masks for the first four Region EV bands (left drag paints, Alt
  erases, the wheel sizes the brush); they travel with the master as
  <master>.masks.png.
• Anchor live on the viewer, with an Anchor knee slider: the master's
  anchor stage is what you see while Anchor is on.
• ProRes 4444 beside 422 HQ, and on both an Encoding row: graded PQ,
  or scene-referred ACEScct (AP1) or ARRI LogC4 (AWG4).
• The model is sdr2hdr_image_v8 (optional; masters default to the
  analytic inverse). shadow_v1 lost to the inverse on every real Meridian
  frame (−4.11 dB) and is retired. A bare start now follows the default when
  it changes, so an install upgraded from an earlier beta opens v8 instead of
  the package it used last; a model you picked yourself stays picked.
• Model parity on real 1080p frames: the ONNX graphs reduce GroupNorm in
  stages, so ONNX Runtime matches PyTorch to 9e-6 on all 287 Meridian test
  frames (it was 5.6e-3).
• Half precision: packages carry an fp16 tile graph, held to a 1% tolerance;
  --precision fp16 on rudra-native diff and bench, and a TensorRT device
  (--device tensorrt) where ONNX Runtime has it.
• Every live control (peak, region EV, strength, mode, preserve, source curve,
  calibration) is one composite and display pass: 1.1 to 1.3 ms at 4K on an
  RTX 4080, under a tenth of a 60 Hz frame. Inference time is separate:
  rudra-native bench reports and budgets it per machine.

Install

macOS (Apple silicon, macOS 13 or later). Open the DMG and drag RUDRA to
Applications. The DMG is signed with RUDRA's Developer ID and notarized, so it
opens with no warning. A DMG built without the certificate (the build log says
"ad hoc") makes macOS warn on first open: System Settings > Privacy &
Security > Open Anyway, or xattr -dr com.apple.quarantine /Applications/RUDRA.app.

Windows (x64, Windows 10 or 11). Run RUDRA-0.9.0-beta.8-windows-x64-setup.exe:
it installs for your user by default (no administrator prompt), adds RUDRA to
the Start menu and can be removed from Settings > Apps. Or take the portable
ZIP: unzip anywhere and run RUDRA.exe. A GPU with DirectX 12 is used when
there is one; the CPU otherwise.

Movies. RUDRA needs an ffmpeg with libx265, prores_ks and zscale.
Windows: the package carries one, nothing to install. macOS: brew install ffmpeg@6, with $(brew --prefix ffmpeg@6)/bin on PATH (the DMG does not carry
ffmpeg yet). RUDRA_FFMPEG_DIR points either at another ...

RUDRA v0.9.0-beta.7

RUDRA 0.9.0 beta 7

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the seventh public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG,
  ProRes 422 HQ or ProRes 4444 (the finishing master), with the audio,
  checked before it is published, through a queue you can stop and resume
  (ProRes 4444 with alpha from the command line:
  rudra-native video --format prores4444). The ProRes files come graded in
  PQ, or scene-referred in ACEScct (AP1) or ARRI LogC4 (AWG4): the
  master before peak and knee, to grade like camera footage. A log file is
  tagged unknown primaries and transfer (there are no codes for these
  curves); set the clip's input to the curve in Resolve. Its sidecar names
  the curve and gamut.
• Masters as ACES 2065-1 EXR sequences.
• The model (optional; masters default to the analytic inverse):
  sdr2hdr_image_v8, level with the inverse on real SDR, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 6

Fixes that bring the native app in line with the 9 Oct review of the
Studio's master path:

• Region EV lands in the master with Anchor on. The anchor (on by
  default) used to divide a Region EV push below its knee straight back out,
  in the viewer and in the master, while the sidecar said the grade was
  applied. The master now grades after the anchor, chroma and grain stages,
  and the viewer shows the same.
• The linear EXR container is really Rec.2020. It carried the plate's
  primaries (Rec.709 for most footage) under a "scene-linear Rec.2020" label
  with no chromaticities; it is now converted and tagged.
• rudra-native deliver, HLG: the BT.2100 system gamma (log10; it was
  right at 1,000 nits only) and saturated colours scaled together instead of
  clipping per channel (no hue shift). Footage in Rec.709 or P3 is converted
  to Rec.2020 before the roll-off, so MaxCLL/MaxFALL describe what is encoded.
• No half-written movies: deliver encodes under a hidden partial name
  and renames it into place, so a failed encode leaves nothing that looks like
  a delivery.

Changes in beta 6 (since beta 5)

• Projects. File > Save project (Ctrl+S) keeps the shot, the frame, the
  model and the whole grade (mode, strength, Region EV, source curve,
  calibration, reference, painted masks, view peak, Deliver settings) in a
  .rudra file; File > Open project (Ctrl+O), a drop, or on Windows a
  double-click opens it. Footage moved together with the project is found
  again; anything missing is named and the rest still opens.
• Autosave. The session is saved a few seconds after each change and on
  close; File > Reopen last session brings back the one the last run left.
• ffmpeg in the Windows package (the gyan.dev "full" build, GPL v3, in
  ffmpeg\ with its licence and source note): movies work without
  installing anything. RUDRA_FFMPEG_DIR picks another build.
• Update check. Once a day RUDRA asks GitHub whether a newer release is
  out and says so in the status bar; Help > Check for updates asks now.
  Nothing is downloaded. Off with RUDRA_NO_UPDATE_CHECK=1.
• Help for a first session: Help > Getting started (the five steps,
  shown once after the first-run check), Help > User guide (F1, also in the package
  as "User guide.md"), Help > Open log folder, and tooltips on every control
  of the Reconstruct, Grade and Deliver tabs.
• Signing hooks: the Windows build signs RUDRA.exe, rudra-native.exe and
  the installer when a code-signing certificate is supplied (the release
  workflow's WINDOWS_CERT_PFX secret).
• Open recent opens a movie as a shot again (it was added as a still).

Changes in beta 5 (since beta 4)

• Starts the same on every machine. Qt settings left in the environment by
  other software (a conda env with PyQt, a Qt SDK) no longer stop RUDRA with
  "no Qt platform plugin could be initialized": the package uses its own
  plugins and the log says what it ignored.
• A log file, rotated at 5 MB: on Windows
  %LOCALAPPDATA%\FXTD Studios\RUDRA\logs\rudra.log, on macOS
  ~/Library/Application Support/FXTD Studios/RUDRA/logs/rudra.log. Attach it
  to a report.
• Failed writes say why: "The EXR file could not be written (the disk is
  full: 1.2 GB free on D:, 3.4 GB needed)", access denied (with a pointer to
  Windows Security's Controlled folder access), or a missing folder. A master
  cut short by a full disk is removed rather than left truncated.

New in beta 4 (since beta 3)

• Source panel. Name the curve that made the SDR (unknown, ACES,
  filmic/Hable, AgX, camera log, plain Rec.709 clip) and the baseline becomes
  its inverse, drawn against the ACES inverse with readouts in nits.
• Three-click calibration. Click black, 18% grey or a highlight and type
  the nits it should be; one to three anchors are fitted over the picked curve.
• Reference match. Load the graded HDR of the frame (EXR, or 16-bit PQ
  PNG/TIFF) and the curve is fitted to it at every code, with the residual in
  stops.
• Painted masks for the first four Region EV bands (left drag paints, Alt
  erases, the wheel sizes the brush); they travel with the master as
  <master>.masks.png.
• Anchor live on the viewer, with an Anchor knee slider: the master's
  anchor stage is what you see while Anchor is on.
• ProRes 4444 beside 422 HQ, and on both an Encoding row: graded PQ,
  or scene-referred ACEScct (AP1) or ARRI LogC4 (AWG4).
• The model is sdr2hdr_image_v8 (optional; masters default to the
  analytic inverse). shadow_v1 lost to the inverse on every real Meridian
  frame (−4.11 dB) and is retired. A bare start now follows the default when
  it changes, so an install upgraded from an earlier beta opens v8 instead of
  the package it used last; a model you picked yourself stays picked.
• Model parity on real 1080p frames: the ONNX graphs reduce GroupNorm in
  stages, so ONNX Runtime matches PyTorch to 9e-6 on all 287 Meridian test
  frames (it was 5.6e-3).
• Half precision: packages carry an fp16 tile graph, held to a 1% tolerance;
  --precision fp16 on rudra-native diff and bench, and a TensorRT device
  (--device tensorrt) where ONNX Runtime has it.
• Every live control (peak, region EV, strength, mode, preserve, source curve,
  calibration) is one composite and display pass: 1.1 to 1.3 ms at 4K on an
  RTX 4080, under a tenth of a 60 Hz frame. Inference time is separate:
  rudra-native bench reports and budgets it per machine.

Install

macOS (Apple silicon, macOS 13 or later). Open the DMG and drag RUDRA to
Applications. The DMG is signed with RUDRA's Developer ID and notarized, so it
opens with no warning. A DMG built without the certificate (the build log says
"ad hoc") makes macOS warn on first open: System Settings > Privacy &
Security > Open Anyway, or xattr -dr com.apple.quarantine /Applications/RUDRA.app.

Windows (x64, Windows 10 or 11). Run RUDRA-0.9.0-beta.7-windows-x64-setup.exe:
it installs for your user by default (no administrator prompt), adds RUDRA to
the Start menu and can be removed from Settings > Apps. Or take the portable
ZIP: unzip anywhere and run RUDRA.exe. A GPU with DirectX 12 is used when
there is one; the CPU otherwise.

Movies. RUDRA needs an ffmpeg with libx265, prores_ks and zscale.
Windows: the package carries one, nothing to install. macOS: brew install ffmpeg@6, with $(brew --prefix ffmpeg@6)/bin on PATH (the DMG does not carry
ffmpeg yet). RUDRA_FFMPEG_DIR points either at another build;
rudra-native ffmpeg-check says whether one has everything. Stills need
nothing extra.

Known in this beta

• macOS: HDR output on Metal EDR, and inference on Core ML, are being
  measured on hardware; until then, check the first-run card (Help > Check the
  display and the model), which reports the display's real peak.
• The model does not yet beat the analytic baseline on every condition of
  the bench (see STATUS.md in the repository); the reconstruction is shown
  where it applied, so you can judge it frame by frame.
• Windows: signed only when the release was built with a code-signing
  certificate (RUDRA.exe > Properties > Digital Signatures shows it); without
  one SmartScreen may ask once.
• macOS: ffmpeg is not in the DMG yet (see Install).
• Linux builds are made from source (scripts/native_app.sh); there...

RUDRA v0.9.0-beta.6

RUDRA 0.9.0 beta 6

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the sixth public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG,
  ProRes 422 HQ or ProRes 4444 (the finishing master), with the audio,
  checked before it is published, through a queue you can stop and resume
  (ProRes 4444 with alpha from the command line:
  rudra-native video --format prores4444). The ProRes files come graded in
  PQ, or scene-referred in ACEScct (AP1) or ARRI LogC4 (AWG4): the
  master before peak and knee, to grade like camera footage. A log file is
  tagged unknown primaries and transfer (there are no codes for these
  curves); set the clip's input to the curve in Resolve. Its sidecar names
  the curve and gamut.
• Masters as ACES 2065-1 EXR sequences.
• The model (optional; masters default to the analytic inverse):
  sdr2hdr_image_v8, level with the inverse on real SDR, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 5

• Projects. File > Save project (Ctrl+S) keeps the shot, the frame, the
  model and the whole grade (mode, strength, Region EV, source curve,
  calibration, reference, painted masks, view peak, Deliver settings) in a
  .rudra file; File > Open project (Ctrl+O), a drop, or on Windows a
  double-click opens it. Footage moved together with the project is found
  again; anything missing is named and the rest still opens.
• Autosave. The session is saved a few seconds after each change and on
  close; File > Reopen last session brings back the one the last run left.
• ffmpeg in the Windows package (the gyan.dev "full" build, GPL v3, in
  ffmpeg\ with its licence and source note): movies work without
  installing anything. RUDRA_FFMPEG_DIR picks another build.
• Update check. Once a day RUDRA asks GitHub whether a newer release is
  out and says so in the status bar; Help > Check for updates asks now.
  Nothing is downloaded. Off with RUDRA_NO_UPDATE_CHECK=1.
• Help for a first session: Help > Getting started (the five steps,
  shown once after the first-run check), Help > User guide (F1, also in the package
  as "User guide.md"), Help > Open log folder, and tooltips on every control
  of the Reconstruct, Grade and Deliver tabs.
• Signing hooks: the Windows build signs RUDRA.exe, rudra-native.exe and
  the installer when a code-signing certificate is supplied (the release
  workflow's WINDOWS_CERT_PFX secret).
• Open recent opens a movie as a shot again (it was added as a still).

Changes in beta 5 (since beta 4)

• Starts the same on every machine. Qt settings left in the environment by
  other software (a conda env with PyQt, a Qt SDK) no longer stop RUDRA with
  "no Qt platform plugin could be initialized": the package uses its own
  plugins and the log says what it ignored.
• A log file, rotated at 5 MB: on Windows
  %LOCALAPPDATA%\FXTD Studios\RUDRA\logs\rudra.log, on macOS
  ~/Library/Application Support/FXTD Studios/RUDRA/logs/rudra.log. Attach it
  to a report.
• Failed writes say why: "The EXR file could not be written (the disk is
  full: 1.2 GB free on D:, 3.4 GB needed)", access denied (with a pointer to
  Windows Security's Controlled folder access), or a missing folder. A master
  cut short by a full disk is removed rather than left truncated.

New in beta 4 (since beta 3)

• Source panel. Name the curve that made the SDR (unknown, ACES,
  filmic/Hable, AgX, camera log, plain Rec.709 clip) and the baseline becomes
  its inverse, drawn against the ACES inverse with readouts in nits.
• Three-click calibration. Click black, 18% grey or a highlight and type
  the nits it should be; one to three anchors are fitted over the picked curve.
• Reference match. Load the graded HDR of the frame (EXR, or 16-bit PQ
  PNG/TIFF) and the curve is fitted to it at every code, with the residual in
  stops.
• Painted masks for the first four Region EV bands (left drag paints, Alt
  erases, the wheel sizes the brush); they travel with the master as
  <master>.masks.png.
• Anchor live on the viewer, with an Anchor knee slider: the master's
  anchor stage is what you see while Anchor is on.
• ProRes 4444 beside 422 HQ, and on both an Encoding row: graded PQ,
  or scene-referred ACEScct (AP1) or ARRI LogC4 (AWG4).
• The model is sdr2hdr_image_v8 (optional; masters default to the
  analytic inverse). shadow_v1 lost to the inverse on every real Meridian
  frame (−4.11 dB) and is retired. A bare start now follows the default when
  it changes, so an install upgraded from an earlier beta opens v8 instead of
  the package it used last; a model you picked yourself stays picked.
• Model parity on real 1080p frames: the ONNX graphs reduce GroupNorm in
  stages, so ONNX Runtime matches PyTorch to 9e-6 on all 287 Meridian test
  frames (it was 5.6e-3).
• Half precision: packages carry an fp16 tile graph, held to a 1% tolerance;
  --precision fp16 on rudra-native diff and bench, and a TensorRT device
  (--device tensorrt) where ONNX Runtime has it.
• Every live control (peak, region EV, strength, mode, preserve, source curve,
  calibration) is one composite and display pass: 1.1 to 1.3 ms at 4K on an
  RTX 4080, under a tenth of a 60 Hz frame. Inference time is separate:
  rudra-native bench reports and budgets it per machine.

Install

macOS (Apple silicon, macOS 13 or later). Open the DMG and drag RUDRA to
Applications. The DMG is signed with RUDRA's Developer ID and notarized, so it
opens with no warning. A DMG built without the certificate (the build log says
"ad hoc") makes macOS warn on first open: System Settings > Privacy &
Security > Open Anyway, or xattr -dr com.apple.quarantine /Applications/RUDRA.app.

Windows (x64, Windows 10 or 11). Run RUDRA-0.9.0-beta.6-windows-x64-setup.exe:
it installs for your user by default (no administrator prompt), adds RUDRA to
the Start menu and can be removed from Settings > Apps. Or take the portable
ZIP: unzip anywhere and run RUDRA.exe. A GPU with DirectX 12 is used when
there is one; the CPU otherwise.

Movies. RUDRA needs an ffmpeg with libx265, prores_ks and zscale.
Windows: the package carries one, nothing to install. macOS: brew install ffmpeg@6, with $(brew --prefix ffmpeg@6)/bin on PATH (the DMG does not carry
ffmpeg yet). RUDRA_FFMPEG_DIR points either at another build;
rudra-native ffmpeg-check says whether one has everything. Stills need
nothing extra.

Known in this beta

• macOS: HDR output on Metal EDR, and inference on Core ML, are being
  measured on hardware; until then, check the first-run card (Help > Check the
  display and the model), which reports the display's real peak.
• The model does not yet beat the analytic baseline on every condition of
  the bench (see STATUS.md in the repository); the reconstruction is shown
  where it applied, so you can judge it frame by frame.
• Windows: signed only when the release was built with a code-signing
  certificate (RUDRA.exe > Properties > Digital Signatures shows it); without
  one SmartScreen may ask once.
• macOS: ffmpeg is not in the DMG yet (see Install).
• Linux builds are made from source (scripts/native_app.sh); there is no
  Linux package in this beta.

Reporting

Issues and results: https://github.com/FXTD-Studios/RUDRA/issues. Include the
version (RUDRA.app > Get Info, or RUDRA.exe > Properties > Details), what
Help > About shows, and for a crash the report macOS or Windows offers.

RUDRA v0.9.0-beta.5

RUDRA 0.9.0 beta 5

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the fifth public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG,
  ProRes 422 HQ or ProRes 4444 (the finishing master), with the audio,
  checked before it is published, through a queue you can stop and resume
  (ProRes 4444 with alpha from the command line:
  rudra-native video --format prores4444). The ProRes files come graded in
  PQ, or scene-referred in ACEScct (AP1) or ARRI LogC4 (AWG4): the
  master before peak and knee, to grade like camera footage. A log file is
  tagged unknown primaries and transfer (there are no codes for these
  curves); set the clip's input to the curve in Resolve. Its sidecar names
  the curve and gamut.
• Masters as ACES 2065-1 EXR sequences.
• The model (optional; masters default to the analytic inverse):
  sdr2hdr_image_v8, level with the inverse on real SDR, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 4

• Starts the same on every machine. Qt settings left in the environment by
  other software (a conda env with PyQt, a Qt SDK) no longer stop RUDRA with
  "no Qt platform plugin could be initialized": the package uses its own
  plugins and the log says what it ignored.
• A log file, rotated at 5 MB: on Windows
  %LOCALAPPDATA%\FXTD Studios\RUDRA\logs\rudra.log, on macOS
  ~/Library/Application Support/FXTD Studios/RUDRA/logs/rudra.log. Attach it
  to a report.
• Failed writes say why: "The EXR file could not be written (the disk is
  full: 1.2 GB free on D:, 3.4 GB needed)", access denied (with a pointer to
  Windows Security's Controlled folder access), or a missing folder. A master
  cut short by a full disk is removed rather than left truncated.

New in beta 4 (since beta 3)

• Source panel. Name the curve that made the SDR (unknown, ACES,
  filmic/Hable, AgX, camera log, plain Rec.709 clip) and the baseline becomes
  its inverse, drawn against the ACES inverse with readouts in nits.
• Three-click calibration. Click black, 18% grey or a highlight and type
  the nits it should be; one to three anchors are fitted over the picked curve.
• Reference match. Load the graded HDR of the frame (EXR, or 16-bit PQ
  PNG/TIFF) and the curve is fitted to it at every code, with the residual in
  stops.
• Painted masks for the first four Region EV bands (left drag paints, Alt
  erases, the wheel sizes the brush); they travel with the master as
  <master>.masks.png.
• Anchor live on the viewer, with an Anchor knee slider: the master's
  anchor stage is what you see while Anchor is on.
• ProRes 4444 beside 422 HQ, and on both an Encoding row: graded PQ,
  or scene-referred ACEScct (AP1) or ARRI LogC4 (AWG4).
• The model is sdr2hdr_image_v8 (optional; masters default to the
  analytic inverse). shadow_v1 lost to the inverse on every real Meridian
  frame (−4.11 dB) and is retired. A bare start now follows the default when
  it changes, so an install upgraded from an earlier beta opens v8 instead of
  the package it used last; a model you picked yourself stays picked.
• Model parity on real 1080p frames: the ONNX graphs reduce GroupNorm in
  stages, so ONNX Runtime matches PyTorch to 9e-6 on all 287 Meridian test
  frames (it was 5.6e-3).
• Half precision: packages carry an fp16 tile graph, held to a 1% tolerance;
  --precision fp16 on rudra-native diff and bench, and a TensorRT device
  (--device tensorrt) where ONNX Runtime has it.
• Every live control (peak, region EV, strength, mode, preserve, source curve,
  calibration) is one composite and display pass: 1.1 to 1.3 ms at 4K on an
  RTX 4080, under a tenth of a 60 Hz frame. Inference time is separate:
  rudra-native bench reports and budgets it per machine.

Install

macOS (Apple silicon, macOS 13 or later). Open the DMG and drag RUDRA to
Applications. The DMG is signed with RUDRA's Developer ID and notarized, so it
opens with no warning. A DMG built without the certificate (the build log says
"ad hoc") makes macOS warn on first open: System Settings > Privacy &
Security > Open Anyway, or xattr -dr com.apple.quarantine /Applications/RUDRA.app.

Windows (x64, Windows 10 or 11). Run RUDRA-0.9.0-beta.5-windows-x64-setup.exe:
it installs for your user by default (no administrator prompt), adds RUDRA to
the Start menu and can be removed from Settings > Apps. Or take the portable
ZIP: unzip anywhere and run RUDRA.exe. A GPU with DirectX 12 is used when
there is one; the CPU otherwise.

Movies, both systems. RUDRA runs ffmpeg and ffprobe from the PATH and
needs a build with libx265, prores_ks and zscale:
macOS brew install ffmpeg@6, with $(brew --prefix ffmpeg@6)/bin on PATH; Windows the "full" build from gyan.dev with
its bin folder on the PATH. rudra-native ffmpeg-check says whether yours
has everything. Stills need nothing extra.

Known in this beta

• macOS: HDR output on Metal EDR, and inference on Core ML, are being
  measured on hardware; until then, check the first-run card (Help > Check the
  display and the model), which reports the display's real peak.
• The model does not yet beat the analytic baseline on every condition of
  the bench (see STATUS.md in the repository); the reconstruction is shown
  where it applied, so you can judge it frame by frame.
• Windows: not signed with a code-signing certificate yet, so SmartScreen
  may ask once.
• Linux builds are made from source (scripts/native_app.sh); there is no
  Linux package in this beta.

Reporting

Issues and results: https://github.com/fxtdstudios/RUDRA/issues. Include the
version (RUDRA.app > Get Info, or RUDRA.exe > Properties > Details), what
Help > About shows, and for a crash the report macOS or Windows offers.

RUDRA v0.9.0-beta.4

RUDRA 0.9.0 beta 4

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the fourth public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG,
  ProRes 422 HQ or ProRes 4444 (the finishing master), with the audio,
  checked before it is published, through a queue you can stop and resume
  (ProRes 4444 with alpha from the command line:
  rudra-native video --format prores4444). The ProRes files come graded in
  PQ, or scene-referred in ACEScct (AP1) or ARRI LogC4 (AWG4): the
  master before peak and knee, to grade like camera footage. A log file is
  tagged unknown primaries and transfer (there are no codes for these
  curves); set the clip's input to the curve in Resolve. Its sidecar names
  the curve and gamut.
• Masters as ACES 2065-1 EXR sequences.
• The model (optional; masters default to the analytic inverse):
  sdr2hdr_image_v8, level with the inverse on real SDR, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 3

• Source panel. Name the curve that made the SDR (unknown, ACES,
  filmic/Hable, AgX, camera log, plain Rec.709 clip) and the baseline becomes
  its inverse, drawn against the ACES inverse with readouts in nits.
• Three-click calibration. Click black, 18% grey or a highlight and type
  the nits it should be; one to three anchors are fitted over the picked curve.
• Reference match. Load the graded HDR of the frame (EXR, or 16-bit PQ
  PNG/TIFF) and the curve is fitted to it at every code, with the residual in
  stops.
• Painted masks for the first four Region EV bands (left drag paints, Alt
  erases, the wheel sizes the brush); they travel with the master as
  <master>.masks.png.
• Anchor live on the viewer, with an Anchor knee slider: the master's
  anchor stage is what you see while Anchor is on.
• ProRes 4444 beside 422 HQ, and on both an Encoding row: graded PQ,
  or scene-referred ACEScct (AP1) or ARRI LogC4 (AWG4).
• The model is sdr2hdr_image_v8 (optional; masters default to the
  analytic inverse). shadow_v1 lost to the inverse on every real Meridian
  frame (−4.11 dB) and is retired. A bare start now follows the default when
  it changes, so an install upgraded from an earlier beta opens v8 instead of
  the package it used last; a model you picked yourself stays picked.
• Model parity on real 1080p frames: the ONNX graphs reduce GroupNorm in
  stages, so ONNX Runtime matches PyTorch to 9e-6 on all 287 Meridian test
  frames (it was 5.6e-3).
• Half precision: packages carry an fp16 tile graph, held to a 1% tolerance;
  --precision fp16 on rudra-native diff and bench, and a TensorRT device
  (--device tensorrt) where ONNX Runtime has it.
• Every live control (peak, region EV, strength, mode, preserve, source curve,
  calibration) is one composite and display pass: 1.1 to 1.3 ms at 4K on an
  RTX 4080, under a tenth of a 60 Hz frame. Inference time is separate:
  rudra-native bench reports and budgets it per machine.

Install

macOS (Apple silicon, macOS 13 or later). Open the DMG and drag RUDRA to
Applications. The DMG is signed with RUDRA's Developer ID and notarized, so it
opens with no warning. A DMG built without the certificate (the build log says
"ad hoc") makes macOS warn on first open: System Settings > Privacy &
Security > Open Anyway, or xattr -dr com.apple.quarantine /Applications/RUDRA.app.

Windows (x64, Windows 10 or 11). Run RUDRA-0.9.0-beta.4-windows-x64-setup.exe:
it installs for your user by default (no administrator prompt), adds RUDRA to
the Start menu and can be removed from Settings > Apps. Or take the portable
ZIP: unzip anywhere and run RUDRA.exe. A GPU with DirectX 12 is used when
there is one; the CPU otherwise.

Movies, both systems. RUDRA runs ffmpeg and ffprobe from the PATH and
needs a build with libx265, prores_ks and zscale:
macOS brew install ffmpeg@6, with $(brew --prefix ffmpeg@6)/bin on PATH; Windows the "full" build from gyan.dev with
its bin folder on the PATH. rudra-native ffmpeg-check says whether yours
has everything. Stills need nothing extra.

Known in this beta

• macOS: HDR output on Metal EDR, and inference on Core ML, are being
  measured on hardware; until then, check the first-run card (Help > Check the
  display and the model), which reports the display's real peak.
• The model does not yet beat the analytic baseline on every condition of
  the bench (see STATUS.md in the repository); the reconstruction is shown
  where it applied, so you can judge it frame by frame.
• Windows: not signed with a code-signing certificate yet, so SmartScreen
  may ask once.
• Linux builds are made from source (scripts/native_app.sh); there is no
  Linux package in this beta.

Reporting

Issues and results: https://github.com/fxtdstudios/RUDRA/issues. Include the
version (RUDRA.app > Get Info, or RUDRA.exe > Properties > Details), what
Help > About shows, and for a crash the report macOS or Windows offers.

RUDRA v0.9.0-beta.3

RUDRA 0.9.0 beta 3

The desktop RUDRA: SDR footage in, scene-linear HDR out, with the places the
model reconstructed shown to you. This is the third public beta of the native
app (C++20, Qt 6.8, no Python at runtime). It is for evaluation and
non-commercial use (PolyForm Noncommercial 1.0.0, see LICENSE; the model
weights: LICENSE-weights).

What is in it

• The Studio in a native window: Reconstruct, Grade and Deliver; residual
  strength, recovery modes, preserve, Region EV, undo; the scopes (waveform,
  histogram, vectorscope), the probe, false colour, wipe, guides.
• An HDR viewer: on an HDR display the image is shown above SDR white
  (Windows: Direct3D 12 scRGB and HDR10; macOS: Metal EDR on the XDR panel).
• Compare that tells you what changed: under the Compare bar, the share of
  the frame RUDRA changed against the analytic baseline, by how many stops,
  and how much of it the view peak shows, with a button that raises the peak
  far enough to see it. Changes tints what RUDRA changed (amber brighter,
  blue darker) on either side of a wipe.
• The invented-pixel map: the Invented layer marks in magenta the values
  RUDRA made up where the SDR clipped or crushed, and in cyan where it read SDR
  detail differently from the baseline. Every master's sidecar carries the
  same shares (invented_pixels).
• Stills and sequences: PNG, JPEG, TIFF, WebP, BMP, a frame or a folder.
• Movies: open, scrub, and export HDR10 (HEVC 10-bit with its metadata), HLG
  or ProRes 422 HQ, with the audio, checked before it is published, through a
  queue you can stop and resume (ProRes 4444 with alpha from the command
  line: rudra-native video --format prores4444).
• Masters as ACES 2065-1 EXR sequences.
• The model: sdr2hdr_shadow_v1, the shipped RUDRA model, run by ONNX
  Runtime (DirectML on Windows GPUs, Core ML on Apple silicon, CPU
  everywhere). Every package is checked against its reference frames the
  first time it runs on a device, and refused if it drifts.
• rudra-native, the command line: diff, video, deliver, batch,
  ffmpeg-check.

Changes since beta 2

• Compare looked like it did nothing: at the default 203-nit view peak the
  reconstruction and the baseline clip to the same white, because RUDRA's
  changes sit above diffuse white or in deep shadow. The compare line now says
  so, and Show at N nits raises the view peak (the master is unchanged).
• New: the Changes tint and the Invented layer (above), held to the
  CPU reference on every GPU path.
• macOS: the DMG is signed with a Developer ID and notarized by Apple when the
  release is built with the signing certificate, so it opens with no warning.
• rudra-native bench times 1080p and 4K and checks per-machine latency
  budgets (native/bench/latency_budgets.json).
• Packages identify themselves as 0.9.0-beta.3. The shipped model remains
  sdr2hdr_shadow_v1; this is an application beta, not newly trained weights.

Install

macOS (Apple silicon, macOS 13 or later). Open the DMG and drag RUDRA to
Applications. The DMG is signed with RUDRA's Developer ID and notarized, so it
opens with no warning. A DMG built without the certificate (the build log says
"ad hoc") makes macOS warn on first open: System Settings > Privacy &
Security > Open Anyway, or xattr -dr com.apple.quarantine /Applications/RUDRA.app.

Windows (x64, Windows 10 or 11). Run RUDRA-0.9.0-beta.3-windows-x64-setup.exe:
it installs for your user by default (no administrator prompt), adds RUDRA to
the Start menu and can be removed from Settings > Apps. Or take the portable
ZIP: unzip anywhere and run RUDRA.exe. A GPU with DirectX 12 is used when
there is one; the CPU otherwise.

Movies, both systems. RUDRA runs ffmpeg and ffprobe from the PATH and
needs a build with libx265, prores_ks and zscale:
macOS brew install ffmpeg@6, with $(brew --prefix ffmpeg@6)/bin on PATH; Windows the "full" build from gyan.dev with
its bin folder on the PATH. rudra-native ffmpeg-check says whether yours
has everything. Stills need nothing extra.

Known in this beta

• macOS: HDR output on Metal EDR, and inference on Core ML, are being
  measured on hardware; until then, check the first-run card (Help > Check the
  display and the model), which reports the display's real peak.
• The model does not yet beat the analytic baseline on every condition of
  the bench (see STATUS.md in the repository); the reconstruction is shown
  where it applied, so you can judge it frame by frame.
• Windows: not signed with a code-signing certificate yet, so SmartScreen
  may ask once.
• Linux builds are made from source (scripts/native_app.sh); there is no
  Linux package in this beta.

Reporting

Issues and results: https://github.com/fxtdstudios/RUDRA/issues. Include the
version (RUDRA.app > Get Info, or RUDRA.exe > Properties > Details), what
Help > About shows, and for a crash the report macOS or Windows offers.