Oink
Concepts

AI tools and licences

The AI nodes, how they run on your machine, and the licence tiers of their models.

raw .md

The AIR button on the node toolbar holds nodes that use machine-learning models: rotoscoping, matte refinement, keying, depth and normals, optical flow and retiming, upscaling and denoising, clean plates, generation, and training your own models.

How they run

  • Models run in a separate AI worker process on your own GPU (CPU if there is no GPU); frames are not sent to any online service. The ComfyUI node is the exception by design: it sends work to a ComfyUI server that you run.
  • The worker's Python environment is created the first time an AI node needs it, in ~/.cache/opencomp/ai-venv (several GB).
  • Model files are downloaded the first time a node needs them, into ~/.cache/opencomp/models, and checked against a known size and checksum before use. OPENCOMP_AI_OFFLINE=1 stops all downloads.
  • Heavy jobs take a machine-wide GPU "lease", so two big jobs never fight for GPU memory; a job waiting for the GPU says who holds it and can be cancelled.
  • Results are cached per frame, so scrubbing back does not rerun the model.

AI workflows

These sit on top of the nodes and run on your machine. None of them changes your comp without a click or an approval.

  • Review and fix. A confidence score per frame for AIRoto mattes (no model needed) puts red ticks on the timeline. Alt+] and Alt+[ jump to the next or previous flagged frame. One correction click re-runs only the frames around it, and stops before the next frame you corrected yourself.
  • Auto QC. qc_check (an Assistant tool) looks for NaN or infinite pixels, black or duplicate frames, flicker, holes and jumps in a matte, and missing files. On 12 real frames a clean sequence gave no flags, and five faults added on purpose were all found.
  • Suggestions. The Read, AIRoto, AIDepth and CameraTracker panels show up to three suggestion chips, worked out from rules and the shot's Prep Show report (for example Fix bad frames, Remove markers, Denoise plate, Roto people, Solve camera). A click runs the recipe as one undo step.
  • Regrain. The Grain node has a Match shot button that measures the plate's grain at the current frame and sets size, intensity and black. GenFill's regrain knob adds the plate's grain to the filled area.
  • Whole-show batches (Assistant tools, one approval each, shown in the Jobs panel and undoable): fix bad frames in every shot, roto all the people, and solve cameras for every shot. Hidden cuts stored on the Read stop FrameFixer and PlateDenoise from reaching across a cut.
  • Pre-compute while idle. Off by default. When on (a checkbox in the Jobs panel), after 60 seconds with no edits it warms AIRoto's image data for the current frame and the next two, one job at a time, only if the AI worker is already running and the GPU is free. Any edit cancels it at once.

Some of these were built and tested without running the models; the model results are still being checked on the GPU.

Licence tiers

Every model has a licence tier:

TierMeaning
bundledpermissive licence (Apache, MIT, BSD). Downloads on first use.
optionalcommercial use is allowed but with extra terms (for example the SAM License or NVIDIA's Open Model License). Nothing downloads until you read and accept the licence in the node's panel.
researchthe research pack: non-commercial weights, for research and evaluation only, never for paid or delivered work. Each user reads and accepts the licence in the node's panel before anything downloads; a studio AI policy can forbid the pack.

Oink never redistributes model weights: they come from their publishers to your machine. The full list of models, their licences, sizes and sources is in AI models and licences.

On this page