All projects
AgentShot
A 5-stage plan-then-generate pipeline of 9 specialized LLM/VLM agents with tool use.
- Type
- Agentic multi-shot video generation & benchmarking
- Context
- Stealth Startup
- Year
- 2025
Highlights
- Built the orchestration framework that decomposes one prompt into a complete multi-shot video, with typed scene contracts, a global main-scene anchor, and chained frame conditioning enforcing cross-shot identity, lighting, and narrative consistency.
- Made every generation backbone hot-swappable (Wan2.2, Veo, HunyuanVideo, Longcat) so the orchestration layer, not a single model, owns output quality.
- Built MSBench, the first multi-shot video benchmark (16 topics, 305 prompts), and a multi-agent LLM-as-judge harness on Qwen3-VL-235B via vLLM; on an open Wan2.2 backbone AgentShot approaches Sora-2 and beats it on cross-shot consistency, with up to +51.7% character consistency over open baselines.