Skip to content

Guides / Using Lumit / Effects

Updated

Addons and models

An addon is an optional download: the model runtime, or a model pack. Each pack is one trained model that reads your footage and hands back one thing: an in-between frame, a depth map, a matte, or the outline of what you clicked. They run on your own graphics card. Lumit works without any of them.

Settings ▸ Addons with the runtime and all five packs installed.
Settings ▸ Addons with the runtime and all five packs installed.
AddonSizeWhat it makesUsed by
Model runtimeNothing on its own. Every pack needs it.All of them
RIFE 4.920 MBThe frame between two real ones.Flow
Depth Anything V2 Small94 MBA depth map from a single frame.Depth
Robust Video Matting14 MBA matte for a person, carried frame to frame.Remove background
BiRefNet Lite214 MBA matte for the main subject of a frame.Remove background
SAM 2.1 Hiera Tiny111 MBThe outline of the thing you click on.Roto brush

Every addon is an analysis model. Nothing generative is listed.

1. Open Edit ▸ Settings… ▸ Addons. Nothing is installed yet.

2. Press Check for addons. Lumit fetches the catalogue. This button is the only time it looks. Each pack lists its licence and size, and reads Needs the runtime until the runtime is in.

The catalogue, before the runtime is installed.
The catalogue, before the runtime is installed.

3. Install the Model runtime. Press Install on its row.

4. Install the packs you want. Each download is checked against its published hash before Lumit uses it.

5. Press Load now, or leave the runtime to load the first time an effect asks for it. The line under it names the graphics provider it is running on.

Install from file installs a pack from an addon.json with its files beside it, for a machine with no network. Where they are kept ▸ Show opens the folder. Remove on a row deletes that addon. Licence opens its licence in the browser.

Without the pack it needs, an effect wears a badge naming the pack and leaves the picture alone.

Flow makes the frames between real ones when a layer is slowed down. Turn it on with the Flow switch in the layer’s Modes column, or Layer ▸ Flow, then set Engine in its Flow group to RIFE.

Engine set to RIFE on a shot slowed to 25%, then played through.

The Flow group is in the layer’s rows in the Timeline, and in Effect controls while Settings ▸ Timeline ▸ Transform in Effect controls is on.

A machine without RIFE previews the layer with the built-in engine and says so on the row. An export of that layer is refused.

Effect ▸ Utility ▸ Depth, then Analyse. The shot is read once, in the background, and nearer is brighter.

Depth on the whole shot, read by DirectML.
Depth on the whole shot, read by DirectML.

Point a Depth of field’s Matte row, a Set matte or a track matte at the layer to use it.

Effect ▸ Utility ▸ Remove background, then Analyse. Model picks the pack:

  • Robust Video Matting for a person in video. Detail sets Portrait or Full body.
  • BiRefNet for the main subject, whatever it is, a frame at a time.
Remove background with BiRefNet, and the Model list open.
Remove background with BiRefNet, and the Model list open.

View shows the cut-out or the matte. Another layer reads the matte through its Matte row, a Set matte or a track matte.

Set Seed from on the Roto brush to Segment. Pick the Roto brush tool (Alt+W) and click the subject on the base frame. Each click refines the outline, and Alt-click marks what is not the subject. Then press Propagate.

Three clicks on the runner, propagated through the shot. View is Boundary.
Three clicks on the runner, propagated through the shot. View is Boundary.

Without SAM 2.1, a propagation seeded by Segment is refused.

Each pack keeps its own licence, named on its row. BiRefNet’s weights are trained on DIS5K, whose terms are non-commercial.