pWe are looking for a Research Scientist to develop the generative modeling technology behind Athlence's rendering pipeline. /ppThe focus of this role is on developing generative models to improve the visual quality of reconstructed scenes while preserving geometry, multi-view consistency, and temporal coherence. You will work with diffusion, flow-based, video, and other generative architectures, developing methods that incorporate geometry, camera information, multi-view observations, and temporal context. /ppThe role sits at the intersection of generative modeling and 3D vision. You will work closely with the 3D reconstruction team to connect geometric representations with generative models and develop methods that can operate reliably on real-world multi-camera captures. /ph3What you will be doing /h3ulliResearch and develop generative models for high-quality, temporally consistent rendering from 3D and 4D scene representations /liliExplore diffusion, flow matching, rectified flow, video generation, and related generative modeling approaches /liliAdapt and fine-tune large-scale 2D, video, and 3D generative models to Athlence's multi-view and temporal data /liliDevelop conditioning mechanisms that incorporate geometry, camera information, temporal context, and multi-view observations /liliExplore feed-forward architectures and 3D/4D representations that connect sparse multi-view inputs directly to renderable scene representations /liliDesign and run experiments to understand the trade-offs between perceptual quality, geometric consistency, temporal stability, latency, and compute /liliDevelop evaluation metrics and benchmarks that capture perceptual quality and temporal/multi-view consistency beyond conventional pixel-based metrics /liliBuild training, evaluation, and inference pipelines for large-scale generative models /liliDefine and improve the data distribution used to train generative models, including dataset curation, filtering, and preprocessing /liliCollaborate closely with the 3D reconstruction and systems teams to integrate generative models into the broader rendering pipeline /li /ulh3What you must have /h3ulliStrong practical experience with modern generative models, particularly diffusion or flow-based models /liliA rigorous understanding of the principles behind diffusion models, noise schedules, flow matching, or related generative modeling methods /liliHands-on experience training, fine-tuning, adapting, or distilling 2D, video, or 3D generative models /liliStrong Python and PyTorch skills, with experience building training and evaluation pipelines /liliGood foundations in computer vision and 3D, including familiarity with camera models and basic multi-view geometry /liliExperience designing experiments, analyzing model behavior, and investigating failure modes /liliAbility to take research ideas from an initial hypothesis through implementation, experimentation, and evaluation /liliAbility to work independently and communicate technical results and trade-offs clearly /li /ulh3Nice to have /h3ulliExperience with large-scale distributed training using technologies such as FSDP, DeepSpeed, or similar /liliExperience with video generation, temporal modeling, or multi-view generative models /liliExperience with novel-view synthesis, neural rendering, Gaussian Splatting, NeRFs, or other 3D representations /liliExperience conditioning generative models on geometry, depth, camera parameters, or other spatial information /liliExperience optimizing generative models for inference latency, memory usage, or throughput /liliPublications at leading computer vision or machine learning venues such as CVPR, ICCV, ECCV, SIGGRAPH, NeurIPS, or ICML /liliOpen-source work or impactful research projects demonstrating practical generative modeling experience /li /ulh3What we offer /h3ulliAccess to unique, proprietary multi-camera sports data captured in real-world environments /liliThe opportunity to work on challenging generative modeling and 3D vision problems where research ideas are evaluated against real-world constraints /liliHigh ownership across the full research cycle, from research exploration and experimentation to integration and deployment /liliA small, highly technical team where you can directly influence the research direction and systems we build /li /ul #J-18808-Ljbffr
Research Scientist — Generative AI
