Research pack
Optional non-commercial AI models (MatAnyone 2, MiniMax-Remover): licences, the acceptance step, and the studio switch.
The research pack is a set of optional AI models whose licences do not allow commercial use. They are useful for tests, personal projects, research and look development, but you must never use them on paid, client or delivered work.
Oink never bundles or redistributes these models. Nothing is downloaded until you, the person logged in, read the licence and accept it in the node's panel. The model code and weights then come from their authors to your machine.
Models
| Model | Used by | What it does | Licence | Commercial use |
|---|---|---|---|---|
| MatAnyone 2 | AIMatteRefine, Model = MatAnyone 2 | Video matting: starts from the rough matte (for example AIRoto) and follows it through the shot, keeping hair and motion blur soft and steady | NTU S-Lab License 1.0 (code and weights) | No: ask the authors (NTU S-Lab) for a commercial licence |
| MiniMax-Remover | AICleanPlate, stage 2 model = MiniMax-Remover | Video object removal: paints the background behind the masked object | Weights: CC BY-NC 4.0. Code: no licence file (the authors keep all rights) | No |
The Auto model choice never picks a research-pack model, even after you accept its licence: you choose it yourself. Pickers that list these models (for example GenFill and GenVideo, which also show FLUX.2 dev, another non-commercial model) mark them with a non-commercial badge.
VideoMaMa (video matting on Stable Video Diffusion; code CC BY-NC 4.0, weights under the Stability AI Community License) was looked at and is not included.
Using a research-pack model
- Pick the model in the node (AIMatteRefine model, or AICleanPlate generate unseen + model).
- The panel shows a red research pack: non-commercial use only box with the licence summary. Click read full licence to read the whole text.
- Click accept (non-commercial use). Your user name, machine, time and a fingerprint of the licence text are recorded. Acceptance is per user: someone else on a shared model cache must accept for themselves.
- The model downloads the first time the node runs. revoke withdraws your acceptance; the node stops using the model even if its files are on disk.
From a terminal: python -m opencomp_ai.registry accept matanyone2, ... revoke matanyone2.
Turning the pack off
| Who | How |
|---|---|
| Studio (authoritative) | Admin AI policy file (/etc/opencomp/ai_policy.json or $OPENCOMP_AI_POLICY) with "allow_research_pack": false. The models disappear from the panels and cannot be accepted, downloaded or run. With "locked": true (the default for an admin file) nobody can turn it back on. |
| One user | "allow_research_pack": false in your own ~/.config/opencomp/ai_policy.json (only when no locked admin file exists). |
| One process | OPENCOMP_AI_RESEARCH=0. |
Licence summaries
NTU S-Lab License 1.0 (MatAnyone 2): redistribution and use, with or without changes, for non-commercial purposes only, keeping the copyright notice and disclaimer; the names of the authors may not be used to promote derived products. Commercial use needs the contributors' permission (Peiqing Yang, Shangchen Zhou, Chen Change Loy).
Creative Commons Attribution-NonCommercial 4.0 (MiniMax-Remover weights): share and adapt for non-commercial purposes, with credit to the authors (Bojia Zi, Weixuan Peng et al.). The inference code has no licence at all, so it is fetched only to your machine for your own non-commercial use.
The full texts ship with Oink in opencomp_ai/models/licences/ and are shown in the panel before you accept.
How they compare
Measured on an RTX 4090 against the commercial-safe models (details and method in docs/notes/ai-research-pack.md):
| Result | |
|---|---|
| MatAnyone 2, real plate (2K, 8 frames, AIRoto matte) | the steadiest refiner: edge flicker 0.032 vs 0.046 (ViTMatte), 0.070 (BiRefNet HR), 0.050 (RVM); 0.17 s per frame, 3 GB |
| MatAnyone 2, synthetic hair with exact ground truth | per-frame error only slightly below the rough matte (SAD 28.9 vs 29.9); BiRefNet HR (21.0) and ViTMatte (25.4) are more accurate per frame; flicker lowest of all (0.0028 vs 0.009 to 0.016) |
| MiniMax-Remover, clip with a known background | fills pixels no frame shows at 23.6 dB PSNR (stage 1 alone 18.1 dB, Wan 2.1 VACE 9.3 dB on this clip); 3.1 s per frame, 15 GB GPU memory |
BiRefNet HR, ViTMatte and Wan 2.1 VACE stay the defaults because they are safe for commercial work.
Prep Show & key shots
Analyse every shot of a conform in one step: hidden cuts, bad frames, grain, camera move, people, faces, tracking markers, lens, exposure, blur and focus, with a shot report per shot, suggested prep nodes for each comp, and key shots per look.
Review and editorial
Oink Review: playback, review notes, the editorial timeline, conform and comp round trips.