Skip to content

Temporal Denoise (FastDVDnet)

FastDVDnet is an Apple-Silicon-only temporal denoise option for image-sequence work. The accepted reference strength is 2.5. It reads two neighboring frames on each side of the current frame and reflects the available sequence at short endpoints.

Review and export

  1. Add a Temporal Denoise node and select FastDVDnet.
  2. Start at strength 2.5. Choose flat boundaries for ordinary images or 360° wrap for equirectangular plates. For 360° work, place FastDVDnet before Transform 360. This denoises the original equirectangular pixels with the seam wrapped correctly, then applies pan, tilt, and roll downstream. If a transformed view must be denoised, render that view to a new image sequence first.
  3. Use Cached A/B Preview to build a bounded 9, 25, or 49 frame comparison. Closing the preview releases it; this is not a full-shot disk cache.
  4. Export an image sequence through Export or Process. Process uses Resume verified sequence. Completed frames and the hidden verification journal remain when processing is cancelled.
  5. To continue an interrupted export, use the same source, graph, settings, frame range, format, and destination. Verified frames are skipped and the first missing or invalid frame is rendered. Choose a new directory when you intentionally want a clean output.

Rolling temporary storage was about 1.3 GiB in the qualified 12K workflow and does not grow with shot duration. Completed output files accumulate separately in the chosen destination. The earlier approximately 14-second M4 Max result used the development Python/MPS runtime. The portable app uses bundled Core ML; its 12K rate must be measured separately and varies with image content, graph complexity, storage, and Apple Silicon generation.

Qualified boundary

The production graph admission currently accepts one unambiguous image-sequence source and ordinary spatial processing around the FastDVDnet node. It rejects ambiguous masks, multiple image sources, downstream temporal consumers, retime, and other routes whose five effective input occurrences cannot be frozen exactly. The error is reported before neural processing begins.

FastDVDnet requires its verified bundled Core ML model on Apple Silicon. When that runtime is unavailable or damaged, the node remains in the project but preview and export stop with a repair message. Existing non-FastDVDnet denoisers remain available.