Multimodal Research Engineer: AI Video & Image Gen Job San Francisco area,California USA,IT/Tech

Requirements

Strong passion for pushing the boundaries of visual AI, with a self-driven, hands-on approach to solving complex technical problems
Proficient in PyTorch with end-to-end experience across data processing, model training, and deployment
Solid understanding of video and image generation architectures, including diffusion models, DiT, Control Net, and SOTA video generation models
Experience with multimodal model training, including working with audio, vision, and language modalities together
Experience with distributed training tools (FSDP, Deep Speed, etc.)
Experience with large-scale data processing, dataset construction, and automated data cleaning
(Desirable) Experience with joint audio-visual or speech-conditioned generation models
(Desirable) Experience with AIGC, video effects, character animation, or asset generation products
(Desirable) Familiarity with ML deployment and orchestration (Kubernetes, Slurm, Docker, cloud platforms)
(Desirable) Publications in relevant venues (NeurIPS, ICLR, CVPR, ECCV, ICCV, or similar)

What the job involves

Joining us as a Research Engineer on the Multimodal team, you'll be at the forefront of building and advancing video and image generation models that bring AI characters to life in entirely new ways
Your work will directly shape how millions of users experience rich, expressive, and visually compelling AI interactions every day
The Multimodal team is responsible for training, fine-tuning, and deploying cutting-edge image, audio and video generation models that power Character.

AI's visual experiences
We work across the full model lifecycle — from data pipelines and training to deployment and product integration
As a Multimodal Research Engineer, you will own and advance our video model training efforts, including joint audio-visual generation and image-to-video. You will collaborate across research, product, and infrastructure to push the boundaries of what AI-generated visuals can look and feel like at scale
Lead fine-tuning and continued training of video generation models, including image-to-video and joint audio-visual generation
Design and experiment with novel model architectures for multimodal generation, including multimodal conditioning (voice, structured text, reference images)
Leverage techniques such as LoRA, RLHF, and full-parameter fine-tuning to improve model quality across diverse visual scenarios
Design and build large-scale data pipelines and automated annotation workflows to support continuous model improvement
Explore model compression, inference acceleration, and serving optimizations to enable efficient real-time video processing at scale

#J-18808-Ljbffr