Tian (Owen) Ye

I co-led the research workflow behind SoL-Pi. Starting from 152 proposed directions, agents autonomously implemented, tested, and refined candidate mechanisms; four survived validation. My work covered research priors, workflow design and implementation, and refinement of the surviving code for release.

On EdgeBench, SoL-Pi used 45–49% fewer tokens and about one-third lower API-equivalent cost than Pi, retaining approximately 94% of Pi’s average score across GPT-5.6 Sol and Claude Opus 5.

With one coordinator and 20 workers, the SoL-Pi swarm achieved 17.5% fewer simulated cycles and 26.8% lower API-equivalent cost than the stock-Pi swarm in a two-hour kernel-optimization comparison. Results reflect one run per configuration.

1,300+ GitHub stars in the first three days of public release.

I contributed to GenEvolve, which combines evidence gathering, visual reference selection, and generation skills through multi-turn tool use. Visual experience distillation turns differences between higher- and lower-quality trajectories into reusable experience.

Co-developed an open-source world foundation model as a core contributor, supporting minute-scale 720p generation and precise 6-DoF camera control on one GPU.

Co-developed open-source 5B and 14B hybrid video diffusion transformers. The 5B stack generates a five-second 720p video in 13.06 seconds on one H100.

Led data-model co-design across the MultiAspect-4K-1M dataset, aspect-ratio-aware positional encoding, VAE post-training, and an aesthetic curriculum for native 4K synthesis.

Led LucidFlux, which directly maps degraded images to photorealistic 2K/4K restorations without requiring captions.

Architected an open-source masked generative transformer for efficient high-resolution text-to-image synthesis.

Co-developed a system for arbitrary-length video generation from text, voice, and reference images. The system combines temporal coherence and speaker-aware lip sync with inference distillation for deployment in Hedra’s product.