The AI Machine
for all VFX Pipelines

Fixed

• Fixed OFX rendering in Nuke when the Driver input is disconnected

Performance

• Declare RGBA cutout output so the updated AIM Client can keep it in shared memory instead of serializing it.

Performance

• Avoid a redundant float32 copy after clipping, preserving output pixels and precision.
• Keep RGBA results in shared memory with the updated AIM Client instead of falling back to serialization.

Added

• DaVinci Resolve (OFX) version. Resolve keeps the output the size of the timeline, so there is no Expand Output there. Instead:
  • Source Size tells the effect how small the clip really is (1/2, 1/3 or 1/4 of the timeline). Resolve has already enlarged the clip by then, so the effect shrinks the frame back to about the clip's own size and upscales that with the model, which gives real detail instead of Resolve's interpolation.
  • Zoom In works as in Premiere, with Zoom Center X/Y as percent of the frame (Y counts up from the bottom, like Resolve's own positions).

Fixed

• The 4xNomos8kDAT model failed to load in every host ("Expected tuple, got torch.Size") because the compiled build rejected the model's tensor shape.
• Frames that take longer than 30 seconds (1080p and up on slower GPUs, or a cold first render) failed with "AIM error: 30 seconds". The effect now keeps its progress moving every few seconds while it renders, which keeps the host waiting. In Resolve this also needs the AI Machine OFX template fix (build tooling 1.0.9).
• Resolve could ask for the same frame several times in a row and the effect upscaled it again each time. Finished frames are now reused while the frame and settings are unchanged.

Fixed

• Premiere: clips smaller than the sequence frame rendered off-center, pushed toward the bottom-right (1.3.1 and 1.3.3 didn't fix it). The cause was the expand-buffer flag, which makes Premiere give the effect a frame-sized buffer; Super Resolution now drops it in Premiere.

Changed

• Rename the Resolution control to Custom Resolution in all hosts, since it only applies when Quality is Custom. Existing projects keep their saved value.

Fixed

• Lower peak memory per rendered frame on Windows by freeing the input frame before reading the result, to reduce bad allocation render failures (AIM Client 5.27).
• Group plugin error reports by error and host command so each error is tracked as one issue.

Changed

• Premiere now shows Resolution and grays it out unless Quality is Custom, instead of never showing it. Premiere ignores dynamic hiding on an effect's first instance, but supports graying out.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Fixed

• Keep the render-destination reminder hidden during local rendering. In After Effects, show it below Help / Contact only for an active remote destination; keep it suppressed in Premiere.
• Preserve saved parameter IDs and parameter-change callbacks, and prevent automatic parameter reveal.

Added

• After Effects-only "256 at Non-Full Res" option below Resolution. Uses 256 for non-1/1 renders and preserves the editable Resolution choice for full-resolution renders. Hidden in Premiere and omitted from Resolve.

Changed

• Use the dashless license code facedirector (the license server ignores dashes, so existing licenses keep working).

• Use the dashless license code aidepthoffield (the license server ignores dashes, so existing licenses keep working).

Fixed

• Release the YOLO source-resolution cap to reduce unnecessary upscaling and GPU memory use.
• Include the missing YOLO weights check so Ultralytics cannot attempt a download into the plugin install directory.

• Fix unintended Cutout color and alpha shifts in Adobe 16-bpc projects by rebuilding with the corrected AIM Client RGBA converters.
• Retain optimized, NaN-safe Adobe16 conversion; 8/32-bpc conversion and model inference are unchanged.

Changed

• Pass explicit bit_depth=8 to image_to_cv2, preserving byte-oriented processing with the strict AIM API.
• Refresh shared development and publishing documentation.

Changed

• Pass explicit bit_depth=8 to image_to_cv2, preserving byte-oriented processing with the strict AIM API.
• Refresh shared development and publishing documentation.

Changed

• Document the explicit input bit-depth contract for the upcoming AIM server; native-range conversion remains explicit and preserves the existing depth fix.

Fixed

• Fixed incorrect or nearly static depth maps with 16/32-bit input in DaVinci Resolve and Adobe hosts. QuickDepth now preserves the input pixel range before inference instead of clipping most pixels to white.
• Applied the same correction to Quality Boost tiles and depth-slice source colors and alpha.

Changed

• Updated internal documentation for publishing and development workflows.

Fixed

• Include the source and driver frame memory reduction from 1.11.1 in the production release.

Changed

• Synchronize plugin version metadata for the Adobe, OFX, and Photoshop release.

Fixed

• Build and publish the OFX plugin (DaVinci Resolve, Nuke) again. v1.2.2 was released for Adobe only because the plugin config had no build_for entry, so aim publish defaulted to Adobe; build_for now lists adobe and ofx.

• Fix per-frame Metal buffer and texture leaks by declaring native return ownership to PyObjC.
• Add bounded memory regressions for source and compiled plugins, including HDR, resizing, cancellation, and renderer recreation.
• Gate release publication on the packaged Mac plugin's Metal memory regression, with pinned test dependencies and archived results.
• Record native ownership pitfalls in shared agent context and document the incident and proposed CI memory safeguards.

• Release Alpha Hint dimension validation without the temporary reproduction script and CI workflow.

• Fix render failure when Tile Overlap is larger than half of Tile Size

• Defer AIM Server startup in Premiere Pro until the effect needs it, instead of launching during plugin discovery. After Effects warmup is unchanged.

Fixed

• Align SDXL VAE encoding and decoding inputs with the current convolution precision, preserving upstream overflow protection and latent precision.
• Log SDXL precision and runtime metadata on inference failures.

Changed

• Installed custom Civitai models now appear in the main model picker; choosing a model applies immediately, and model downloads are labeled "Download Models".
• Remove the SDXL diagnostic test suite and its CI workflow.

• Fix tiled rendering failures on Apple GPUs while preserving model weights and float32 output.
• Limit high-quality MPS tile batches to reduce memory use and improve repeated render times on the tested 24 GiB M2.
• Validate pixel-identical repeated renders, document randomized performance comparisons, and correct the CI release instructions.

• Keep BiRefNet deformable convolutions on Apple GPUs instead of falling back to CPU, preserving the model weights, resolution, and float32 precision.
• Add guarded, randomized performance comparisons and numerical render checks.
• Use a version representable in Adobe's four-bit patch-version field.
• Update shared development instructions and document the benchmark workflow.

Fixed

• Accept GGRFADR aescripts licenses alongside existing BSKFADR licenses.

• Add automatic subject focus using the bundled U2NETP saliency model and robust subject depth sampling.
• Enable subject autofocus by default and place its checkbox above the Focus Mask group; disable manual focus controls while it is enabled.
• Set default near/far radius to 30 and Highlight Glow to 10, retaining Focus Region Size at 5.
• Make Show Focus Mask display the normalized blur radius independently of Focus Clarity, including split and disabled radii.
• Cache subject detection separately from depth inference and preserve frame-independent rendering.
• Add verified subject-model provisioning, release packaging, and autofocus/focus-preview regression coverage.
• Reorganize agent documentation around current behavior, validation, and complete model-inclusive releases.

• Migrate AI Depth of Field to AIM with the original Depth Anything V2 Small model.
• Enable PSS licenses for product ID ai-depth-of-field, retaining original aescripts license compatibility.
• Port depth-aware bokeh, point focus, split radii, highlight glow and focus clarity.
• Preserve alpha, legacy AE match name and parameter disk IDs.
• Cache inference independently from focus and blur settings.
• Use float16 for MPS depth inference, retaining float32 depth normalization and bokeh.
• Restore native Metal texture sampling with the original bokeh shader equations.
• Add a Release standalone legacy shader harness and repeatable renderer comparisons.
• Reduce Metal texture bandwidth with float16 storage for normal-range images; retain float32 for HDR/negative inputs and exact disabled-radius passthrough.
• Skip bokeh samples beyond conservative local depth bounds and simplify accumulation without changing sampling quality.
• Accelerate the original transformer weights with a bundled Apple Silicon MLX backend, retaining the Torch decoder and fallback.
• Add paired timing, adversarial bokeh parity, and hybrid inference validation; improve measured 1080p and 4K render times by approximately 21%.
• Build a separately named Legacy comparison effect and fix its bundle resource lookup, failed initialization cleanup, and teardown guards.
• Prepare verified weights in the renamed Legacy effect's model directory to prevent model-not-found errors.

Changed

• Publish Photoshop packages for macOS and Windows alongside Adobe and OFX packages.
• Document the release host selection and Photoshop manifest fields.

Changed

• Reduce macOS inference contention by disabling idle ONNX Runtime thread spinning, improving measured 4K rendering time by approximately 5% on Apple M2 without changing output.

Improved

• Rebuild with AIM's Adobe cancellation-log cleanup: normal user cancellation no longer appears as an error in render diagnostics or shared error logs. Cancellation propagation and parameter cleanup are unchanged; genuine failures retain their diagnostics.
• Retain the SmartFX bounds fix and failure diagnostics from v1.6.17.