03 · Model Library

The Model Library

Every model family follows the same three steps: download weights, build an engine, use it. This page covers the families, the one-click presets, and building engines for new resolutions.

Weights → engines

Weights are the neural network itself, downloaded once per family. An engine is that network compiled for your hardware and a specific tile size (CoreML on Apple Silicon). The Model Library's left rail shows each family's state: a green dot means ready; Download weights or Build engine mean there is a step left.

A model family page: weights installed (step 1), build (step 2), and the engines already compiled (step 3).

The model families

RefocusFast, consistent CNN upscaler. Used for the standard 4× pass. (Enhance)
Regenerate-QStreaming, quantized generative upscaler. Adds detail with moderate memory use; the default diffusion model. (Enhance)
Frame interpolationGenerates in-between frames to raise the frame rate. (Enhance)
DenoiseTemporal, flow-aligned video denoiser. Used by the Enhance pipeline and by Reprocess in Refine. (Enhance and Refine)
SegmentationSAM2 object segmentation. Powers masks in the Refine workspace. (Refine)
Depth estimationMonocular video depth maps, metric or relative. Powers the Focus tool and depth masks. (Refine)
RegenerateThe full generative upscaler. Maximum detail, heaviest compute. Listed under Beta and only where it can be built.

The rail groups families by the workflow they serve: Enhance, Refine, and Beta. The dashboard's Model Library cell shows one readiness meter per group.

Starter presets

Presets follow the workflows. All installs one engine per model family (depth gets one metric and one relative model). Enhance covers upscaling and cleanup: Refocus, Regenerate-Q, frame interpolation and denoise. Enhance without Diffusion leaves out Regenerate-Q for a smaller download and faster builds. Refine installs the Refine tools: segmentation, metric and relative depth, and denoise. A preset downloads the weights and builds the engines in one step with one overall progress bar; you can keep using the app while it runs. Installs that came with the Essentials bundle already satisfy every preset.

Starter presets. The checklist shows what is already installed and what will be downloaded and built.

Building engines for new dimensions

Engines are compiled per resolution. If you load a video in the Enhance workflow and no engine matches its shape, build one: open the family, click Build engine, and set the dimensions. The drawer validates as you type (dimensions must divide evenly, and it suggests the nearest valid values) and shows a live log while building. Builds take a few minutes per engine.

The build drawer for Refocus: tile size and frame window. Each family exposes its own options (quantization, model size, variant).

Build options

Tile height / widthThe input tile the engine processes at once. For Refocus, output is 4× the tile. Match your footage: 270×480 for vertical phone video, 480×270 for landscape.
Min / Max framesThe temporal window the engine accepts: how many frames it batches per inference.
Quantize (Regenerate)int8 and palettized variants trade a small amount of quality for smaller, faster engines.
Model size (Segmentation)Tiny through Large. Larger models track objects more accurately but load and run slower.
Variant · Size (Depth)Relative or Metric depth, in Small / Base / Large. Metric gives real distances in meters, which the Focus tool can use.
Overwrite toggleRebuild over an existing engine of the same configuration.
Build engines that match the aspect ratios of the footage you work with, for example 270×480 for vertical phone video and 480×270 for landscape. The Enhance workflow picks the best installed engine automatically (Engine · Auto).

Performance: build for your hardware

The Essentials bundle and the presets build one general-purpose shape per model. The largest speedups come from engines shaped for your footage and compiled for your exact hardware, and every build is a one-time step: weights stay installed, and each new engine appears next to the others in step 3.

Custom model shapes on Apple Silicon

CoreML engines are compiled for a fixed tile and frame window. A tile that matches your footage's aspect ratio covers a frame in fewer passes with fewer seams: 480×270 for landscape, 270×480 for vertical video. A larger tile, such as 540×960 for 1080p sources, halves the number of passes again at the cost of memory; a larger frame window (Max frames) keeps the GPU busier per inference. Build in the Model Library, then pick the engine in Enhance or leave Engine · Auto to choose the best fit.

Refocus tileMatch the aspect ratio of your sources. Go larger when memory allows; 16 GB machines are best at 270×480, 32 GB and up can take 540×960.
Regenerate-Q tile512 is the general shape. 768 needs about 2.4 GB more per tile but processes 4K sources in a quarter of the passes; int8 quantize halves the transformer weights for less memory at a small quality cost.
Frame windowRaise Max frames on machines with headroom to batch more frames per inference.
Segmentation sizeSmall is the default. Tiny tracks faster; Base+ and Large track more accurately but load and run slower.
Depth sizeSmall is the default and runs single-pass. Base and Large give finer maps at several times the cost.
Build for the footage you actually process. One engine per aspect ratio you work with is enough; the Enhance workflow and Reprocess pick the best installed engine for each source.

Managing disk space

The rail footer shows the total space used. Engines can be deleted individually from step 3 (trash icon) and rebuilt later. Weights stay installed, so rebuilding an engine does not re-download anything.