Research Scientist Intern, Audio-Visual Conversation Scene Generation (PhD)

Meta•Redmond, WA

1d•$7,650 - $12,667

About The Position

The Meta Reality Labs Research Team brings together a world-class team of researchers, developers, and engineers to create the future of virtual and augmented reality, which together will become as universal and essential as smartphones and personal computers are today. Reality Labs Research Audio builds the audio-visual perception that future smart glasses will rely on. That includes understanding who the wearer is talking to, who is speaking, and how a conversation is organized. To train and evaluate these models at scale, we need realistic conversation data that comes with exact ground truth. This internship explores generative models that produce such data. The project: Generate realistic video with audio of multi-person conversations. Inputs can be text alone, text plus one image of a group, or text plus a reference image for each person. Beyond looking real, the scenes should show natural social interaction: turn-taking, gaze, head orientation, listener reactions and group formations. Each scene should also be controllable: the 3D position of each person, how people are grouped, and the position and pose of the egocentric camera. That control is what lets the output serve as accurately labelled training data for audio-visual models, e.g. predicting the glasses wearer's intent or detecting conversation groups.

Requirements

Currently has, or is in the process of obtaining, a PhD in Computer Science, Electrical Engineering, Machine Learning or a related field
Research experience in generative modeling for video or images, with first-author publications at venues such as CVPR, ICCV, ECCV, NeurIPS, ICLR or SIGGRAPH
Hands-on experience training or fine-tuning modern video diffusion or flow-matching models (e.g. DiT-based architectures), including conditioning methods such as adapters, ControlNet or LoRA
Experience with controllable generation from geometric or structural signals: camera pose, 3D layout, depth or human pose
Solid grounding in 3D vision: camera models, multi-view geometry and coordinate transforms
3+ years experience with Python, Matlab, Pytorch, large-scale distributed GPU training or similar
Must obtain work authorization in the country of employment at the time of hire, and maintain ongoing work authorization during employment

Nice To Haves

Experience with joint audio-video generation, speech-driven talking heads or gestures, multi-speaker TTS, or lip sync
Background in spatial audio: binaural rendering, HRTFs, or room acoustics simulation
Research on modeling social interaction: conversational dynamics, gaze, turn-taking or group formations
Familiarity with egocentric perception and datasets (e.g. Ego4D, Ego-Exo4D, Aria), and with tasks such as active speaker detection or conversation grouping
Experience with 3D or 4D human and scene representations (e.g. SMPL-X, Gaussian splatting, NeRF) or multi-view-consistent video generation
Experience with synthetic data generation, game engines, or improving how well models trained on synthetic data perform on real data
Experience using LLMs or VLMs to plan scenes, layouts or dialogue
Track record of open-source contributions or a previous industry research internship
Proven track record of achieving results as demonstrated in accepted papers at top computer vision and machine learning related conferences such as CVPR, ECCV, NIPS, ICASSP, InterSpeech etc
Experience working and communicating cross functionally in a team environment
Intent to return to a degree-program after the completion of the internship/co-op

Responsibilities

Design and train video diffusion or flow-matching models for multi-person conversation scenes, generating audio and video together.
Develop ways to condition generation on 3D layout, group structure and egocentric camera trajectory.
Keep each person's identity consistent across frames and viewpoints, from one or several reference images.
Model natural conversational behavior: speech-driven faces and gestures, gaze, turn-taking and listener reactions.
Build evaluations for realism, audio-visual sync, identity consistency, and how closely the generated 3D positions and poses match the specified ones.
Use the generated data to train downstream egocentric perception models, and measure how much it helps.
Publish results at top venues together with Meta researchers.