Refocused 1.0.7: square-pixel output for DV and DVD sources, resolution-based output names, and bug fixes

Refocused 1.0.7 is out now for all platforms. This is a small release focused on fixes, most of them for standard-definition tape and disc sources: DV, DVD and HDV footage with non-square pixels, which several of you have been feeding into Enhance since 1.0.6 added interlace support.

Square pixels for anamorphic sources

DV, DVD and HDV footage is stored with non-square pixels: a 720×480 NTSC DV frame is meant to display at 4:3, and the player stretches it on playback. Until now Enhance upscaled the stored frame and passed the stretch flag through, so a 1080p target came out as 1620×1080 with a flag telling the player to squeeze it. Many editors and players honour that flag, but not all of them, and a file that looks right in one place and stretched in another is not a good deliverable.

Enhance now detects non-square source pixels and shows the display size next to the stored size in the Source step. A Square pixels switch, on by default, resamples the output to its display shape with 1:1 pixels: the same 720×480 DV source at 1080p now produces a 1440×1080 file that displays correctly everywhere. Fixed scale fits the display shape inside the 4× canvas, so the model still runs at full resolution and is only ever downscaled. Turn the switch off to keep the source pixel shape, as before. It does not apply in Same-resolution mode, where there is no scaling stage to resample in.

Bug fixes

Engine

• DV footage encoded at the wrong frame rate. Raw .dv files, DV-in-AVI captures, and any chapter segment cut from one, report a placeholder frame rate of 600 or 60000 fps through the DV demuxer. The engine took it at face value, so outputs played about twenty times too fast with a duration of a few seconds. The real frame rate is now used, and the placeholder is logged when it is seen.
• Wrong display width when the output frame shape differs from the source. The output carried the source's pixel-aspect flag verbatim. That was correct for a plain 4× upscale, but any output whose shape differs from the source, such as a square-pixel conversion or a custom size, displayed at the wrong width. The flag is now computed so the output always keeps the source's display aspect.
• Memory Fix for longer clips We fixed a race condition in regenerate-q that could result in memory corruption in longer video clips.

App

• Windows: Reprocess would not run in Refine. The Run button stayed disabled with a warning that VFI and Regenerate require macOS, a leftover from before the Windows engines existed. It now runs wherever the needed engines are installed.
• Output names now say the resolution. Every Enhance output used to be named <source>_4x, including 2× targets and same-resolution jobs. The default name now carries the size the job lands on: _1080p for a ladder target, _2880x1920 for fixed and custom sizes. The name follows your size choice until you edit it or browse for a path, and existing names with the old suffix are still recognised in the library.

What's next in 1.0.8

• AMD support. Bringing the engine to AMD GPUs is in progress. Every model family now runs on our AMD test hardware with output matching the macOS and NVIDIA builds, but we still have performance and packaging work to do.
• Whole-frame Regenerate-Q on macOS. Part of the AMD work was creating a block-sparse attention kernel that lets Regenerate-Q process a whole 1080p canvas as a single tile instead of a grid of smaller ones, which is faster on long jobs and removes the tile grid entirely for 1080p outputs. We are now testing a macOS implementation of that kernel.
• Projects and Videos Revamp. The Projects and Videos sections predate chapters, patches, versions and depth maps. We are working on revamping the UI to make it easier to use, based around the same lineage view Refine now uses. We are focused on making sure the history of a video is visible from the library rather than only inside a session, and improving general usability.

Longer term

Beyond 1.0.8, the focus is on adding models and improving what the diffusion models produce:

1. Deinterlacing. As mentioned in 1.0.6, we currently deinterlace with a conventional filter during decode. A learned deinterlacing model is next, for footage where the filter leaves combing or softness behind.
2. Stabilization. A stabilization model for handheld and tape sources.
3. The next generation of the Regenerate model family. Better detail and consistency from the diffusion models that do the heavy lifting in Enhance.

We will publish a more detailed roadmap in the next week or two, with general release timelines and more information about these projects and other changes that are coming.

To update, download the latest build from the install page, or use the in-app updater in Settings.