Refocused 1.0.9: block-sparse Regenerate-Q on NVIDIA, ProRes profiles, stalled-job detection, and bug fixes

Refocused 1.0.9 is out now for all platforms. The block-sparse attention kernel that 1.0.8 brought to Apple Silicon and AMD now runs on NVIDIA GPUs as well, so the Model Library on Windows and Linux can build whole-frame Regenerate-Q engines. This release also adds a ProRes profile choice to the Output settings, has the engine fail a job that stops making progress instead of leaving it running, and lists your recent jobs on the Activity screen. The bug fixes include an app crash on macOS 27, Regenerate-Q on Radeon RX 9000 cards, and Refocus stalls on 12 GB NVIDIA cards.

Block-sparse Regenerate-Q on NVIDIA

In 1.0.8, block-sparse attention was available on Apple Silicon and AMD only, and the NVIDIA builds used dense square tiles. 1.0.9 adds it to the Windows and Linux NVIDIA builds. Build on the Regenerate-Q page of the Model Library now has the Block-sparse attention switch on NVIDIA too, off by default. With it on you can build rectangular tiles, with height and width in multiples of 128, including 1152×1920, which covers a 1080p output in a single tile with no tile seams, and 768×1280 for 720p. trt_regenerate_q_builder --sparse-attn now works in the release builds as well.

Whole-frame tiles need a lot of memory: 1152×1920 needs a 40 GB-class GPU, while a block-sparse 768×768 tile runs on 24 GB cards. Auto on NVIDIA now takes your card's memory into account. It estimates how much memory each installed engine needs for the job, considers only the ones that fit, and picks the one with the fewest tiles. A 40 GB-class card gets the whole-frame block-sparse tile. A 24 GB card stays on the dense 768 tile even when a whole-frame engine is installed, instead of running out of memory on the first chunk. The Tile size row still lets you pick any installed engine yourself.

The engines in the installers are unchanged: a dense 512 tile, plus a dense 768 tile that setup builds on cards with 24 GB or more.

ProRes profiles

ProRes has no rate control; the profile sets the quality and the file size. Until now the Output card showed the CRF quality slider for the ProRes encoders even though it had no effect, and every ProRes file was written as ProRes 422 HQ, or 4444 when the pixel format was 4:4:4. For ProRes the slider is now replaced by six profiles: 422 Proxy, 422 LT, 422, 422 HQ (the default), 4444 and 4444 XQ. Picking a profile also sets the matching pixel format, since the 4444 profiles write 4:4:4 chroma. To use a profile for every new video, set it in Workflow settings → Encoding defaults.

Stalled-job detection

When a job stopped making progress, for example because the GPU driver had started paging video memory to system RAM, it used to stay "running" with no end. The app showed a Stalled? hint after five minutes, but the engine kept waiting. The engine now tracks progress at every stage of a job. If nothing moves for 10 minutes (30 minutes before the first frame), it marks the job failed and names the stage that stopped, for example encode step made no progress for 10 min. The reason shows on the job in Videos. You can change both limits in app_config.json (pipeline.stall_timeout_seconds and pipeline.stall_startup_timeout_seconds), and 0 turns a limit off.

Recent jobs on the Activity screen

When no job is running, the Activity screen now shows your last four finished jobs as cards. Each card shows the total time, the frame count and the speed, plus the time and frame rate of each stage. Click a card to open the job in Videos. A failed job that has no timings shows its error instead.

Bug fixes

Engine

• Radeon RX 9000 (RDNA4): Regenerate-Q failed. On 1.0.8.1, Check this computer and every Regenerate-Q job failed on the first chunk with transformer chunk_first failed: CUDA error: invalid argument. The cause was a fault in AMD's ROCm 10.0 runtime on RDNA4 cards. The AMD build now ships ROCm 10.1.0, which fixes it.
• Refocus stalled on 12 GB NVIDIA cards. By default, Refocus always processed the largest batch of frames its engine allowed, regardless of the card's memory. With a 1920×832 source that batch needs about 14.5 GB. On a 12 GB card the NVIDIA driver paged the overflow through system RAM, and the job crawled or stopped. Refocus now sizes the batch to your card's memory on NVIDIA and AMD (14 frames instead of 24 for that clip on a 12 GB card). If you set Frames Per Processing Unit yourself, your value is kept, and the engine log warns when it does not fit.
• Regenerate-Q on NVIDIA failed partway through when the scratch drive filled up. On NVIDIA, Regenerate-Q keeps the working state of tiles that do not fit in GPU memory on the scratch drive, about 1.2 GB per 512×512 tile. A 2× Reprocess of a 1080p clip can use around 40 tiles, or about 46 GB. When the drive ran out, the job failed 5 to 10 minutes in with tile state writeback failed. The job now stops before it writes any tile state. The error gives the space needed, the space free, and the setting to change (Settings → Engine → Scratch Location). Refine's Reprocess card also shows the scratch estimate next to the tile count.

App

• macOS 27: the app crashed during jobs. On macOS 27 the app could quit mid-job when an audio device changed, for example AirPods connecting or disconnecting, a display with speakers going to sleep, or USB or HDMI audio being plugged in. The job then showed as failed. The cause was a bug in the CoreAudio code of the bundled mpv player library, which mpv has since fixed. 1.0.9 includes that fix. Preview playback on macOS now mixes audio down to stereo. Outputs are not affected.
• Model Library: building an engine that was already installed. A build whose configuration matched an installed engine ran to the end (35 minutes for a Refocus GPU preset) and then threw the result away. The Build drawer now stops right away and names the installed engine. To rebuild it, turn on Overwrite existing engine of the same configuration.
• Empty Help menu. The Help menu had nothing in it on Windows and macOS. It now has User Guide (the in-app guide), User Guide on refocused.ai, Release Notes and Settings.
• Settings: changes needed a second step to apply. Save Changes in a Settings section only took effect after a separate click on Apply Changes at the top of the page. Each section now has a Save & Apply button that saves and applies in one step. Save Changes still lets you edit several sections and apply them together.
• Refine: the queue label was hidden. The progress strip at the bottom of the workspace covered the bottom of the docked Reprocess card, including Waiting in engine queue. The card now sits above the strip.
• Model Library on NVIDIA now shows the optimization level each TensorRT engine was built with (fast, standard or max) next to its size.

What's next

• Continued AMD testing. We are testing more Radeon cards and will add them to the supported list as we verify them.
• Deinterlacing. A learned deinterlacing model, for footage where the decode-time filter leaves combing or softness behind.
• Stabilization. A stabilization model for handheld and tape sources.
• The next generation of the Regenerate model family. Better detail and consistency from the diffusion models that do the heavy lifting in Enhance.

Thank you

A special thank-you to runnermx and Anton Kulagin for their detailed reports and logs, and patience while testing this cycle. Some of the fixes above exist because they took the time to write in. Thank you to everyone else who sent bug reports and feedback too, and keep them coming!

To update, download the latest build from the install page, or use the in-app updater in Settings → Application. On AMD, run Benchmark again under Settings → Application → Check this computer after updating. The stored measurements are tied to the GPU runtime version, so until you re-run it, Auto uses its built-in estimates.