Generating machine-learned inverse rig models

A computing system comprising one or more computing devices configured to:

• receive one or more training examples, each training example comprising target mesh data for a video game object, wherein the target mesh data specifies a configuration of a shape of the video game object;
• for each training example:
• generate predicted rig parameters using an inverse rig model, comprising processing the target mesh data of the training example;
• generate predicted mesh data, comprising processing the predicted rig parameters using a forward rig model, wherein each of the target mesh data and the predicted mesh data comprise vertex data, the vertex data comprising co-ordinate information for each of a plurality of corresponding vertices;
• and
• determine a loss, comprising performing a comparison between the predicted mesh data and the target mesh data, wherein determining a loss for each training example comprises determining a mesh loss, comprising determining a distance between each pair of vertices that correspond to each other in the target mesh data and the predicted target mesh data;
• and
• train the inverse rig model, comprising updating parameters of the inverse rig model based on the losses of one or more training examples.

A computer-implemented method for generating a machine-learned inverse rig model, wherein the machine-learned inverse rig model outputs rig parameters for a facial animation rig used to generate animations of facial expressions in video games, the method comprising:

• receiving one or more training examples, each training example comprising target mesh data for a face portraying a particular facial expression;
• for each training example:
• generating predicted rig parameters using an inverse rig model, comprising processing the target mesh data of the training example;
• generating predicted mesh data, comprising processing the predicted rig parameters using a forward rig model, wherein the forward rig model comprises a differentiable function obtained by:
• extracting, from the facial animation rig, operations used to transform the rig parameters into mesh data;
• and
• converting the operations into a sequence of differentiable subfunctions;
• and
• determining a loss, comprising performing a comparison between the predicted mesh data and the target mesh data;
• and
• training the inverse rig model, comprising updating parameters of the inverse rig model based on the losses of one or more training examples.