Refocused 1.0.5: much faster targeted-resolution upscales, FFV1 lossless output, and fixes

Refocused 1.0.5 is out now for all platforms. This is a smaller, focused release: while working on the next round of features we found that targeted-resolution upscales were doing far more processing than they needed to, and that was worth fixing right away.

Targeted-resolution upscales are much faster

When you pick a target resolution in Enhance (say, 1080p from a 720p source), previous versions reached it indirectly: the diffusion model generated the full 4x canvas, and the result was then downscaled to your target. For a 720p source, that meant tiling and generating a 5120x2880 image — more than 7x the pixels of the 1920x1080 you actually asked for — and then throwing most of that work away.

The engine now generates directly at the target size. The model fits your footage to exactly the canvas you chose and runs there, so a sub-4x target processes only the pixels you keep. Depending on your source and target, targeted upscales can be several times faster than in 1.0.4, with no change in output quality.

Two related improvements shipped alongside this:

• No more tile seams. Regenerate and Regenerate-Q process video in tiles, and at some output sizes the boundaries between tiles could be visible in the result. Tiles now overlap by a sensible default margin and are blended together, so seams are gone without any configuration.
• Accurate size readouts. Everywhere Enhance shows an output size — the settings cards, the pipeline sidebar, the advanced scaling dialog — it now shows exactly the dimensions the engine will run. The advanced scaling dialog also edits your Output-size choice directly, so there is one source of truth for what you get.

FFV1 lossless encoding

FFV1, the lossless codec the archival world standardizes on, is now available in the encoder list on every platform. If you want output with every pixel preserved exactly — for archiving masters, or as an intermediate before your own encode — pick FFV1 in the Output card. It supports 8- and 10-bit output in 4:2:0, 4:2:2, and 4:4:4, writes to MKV (recommended) or AVI, and includes per-slice checksums so file corruption is detectable. Expect very large files; lossless means no compression tradeoffs in either direction.

Workflow improvements

• Saved pipelines now capture everything. Saving a pipeline snapshots your full configuration — model choices, interpolation, denoise, grain, blend, encoder and output settings, and any advanced per-stage tweaks. Applying it to a new video restores all of it (the clip range stays with the video, as you'd expect).
• Default encoder in workflow settings. You can now set your preferred encoder, container, and format once in workflow settings, and every new video starts from it instead of the built-in default.

Additional Bug fixes

Windows

• Fixed Regenerate-Q jobs stalling near the start of a job. On systems where Windows limits the amount of pinned memory available to the GPU, which can happen dynamically, mid-job, Regenerate-Q's disk-tier staging could deadlock, leaving the job pinned at around 1% forever with no error. The engine now handles the constrained case correctly and logs clearly if the disk tier is slowing a job down, instead of hanging silently.

What's next

The 1.0.4 roadmap still stands: the next feature release (1.0.6) focuses on an overhaul of the Refine workspace, along with improvements to sign-up and the model building experience. We're also continuing to investigate tiling in some of the refine models, like depth mapping.

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