All work
Hedra
Contributor to Character-3 (MagicInfinite), Hedra’s flagship omnimodal foundation model for digital-human video — launched to 3M+ users and anchoring Hedra’s a16z-led $32M Series A.
- Role
- Research Engineering Intern
- Period
- Mar. 2025 – Aug. 2025
- Website
- hedra.com
Highlights
- Contributed to the from-scratch design and training of Character-3, a 13B Diffusion Transformer that turns one portrait, free-form text, and a voice clip into arbitrarily long, identity-stable talking videos across realistic humans, anime, sculptures, and animals, trained at scale on 200×H100.
- Worked on the architecture (image-as-text-token identity injection, face-masked per-frame audio cross-attention with an adaptive face-area loss), a DMD2 + shared-base dual-LoRA distillation compressing 50 sampling steps to 4 (20× speedup; a 1-minute 540p video in 60s on 8×H100), and a sliding-window denoising scheme for infinite-length generation at no extra training cost.
- Owned the data side: designed and built the Slurm-based data orchestration pipeline from scratch (job scheduling, quality filtering, distributed processing), curating 10M+ usable clips and the 1.85M-clip training set; promoted to lead 3+ interns in the second month.
- Character-3 won 91.33% user preference over prior state of the art and launched as the first omnimodal foundation model in production (3M+ users, 10M+ videos), anchoring Hedra’s a16z-led $32M Series A.