融光: Agent-Native Video Production for the Short-Drama Era

Agent-Orchestrated Production Pipeline

融光 adopts a director-agent architecture that decomposes video production into discrete cognitive tasks, eschewing monolithic generation for modular agency:

Core Abstractions

• SceneContext: Persistent memory object maintaining character appearance, lighting conditions, and narrative state across agent handoffs
• ShotPlan: Agent-generated storyboard metadata that decouples narrative intent from visual execution
• AssetLedger: Immutable record of generated clips enabling non-destructive agent collaboration

The TypeScript/Java split reveals architectural maturity: TypeScript handles the event-driven agent choreography (where async/await patterns excel), while Java manages the resource-intensive video encoding and model inference orchestration.

The breakthrough isn't generating videos—it's generating consistent videos. 融光 treats temporal coherence as a multi-agent consensus problem rather than a model inference issue, using agent critique loops to enforce character identity and lighting continuity across scenes.

Specific Technical Innovations

1. Character Lock Protocol: A specialized agent extracts visual embeddings from reference images and injects consistency constraints into each generation prompt, maintaining facial structure and costume details across disconnected inference calls.
2. Narrative-Aware Shot Sequencing: Unlike prompt-chaining approaches, the CutterAgent analyzes emotional beats in source scripts to determine optimal shot duration and transition timing, effectively automating cinematic grammar.
3. Short-Drama Optimization: Hardcoded workflow templates for 1-3 minute vertical video formats (9:16 aspect ratio, hook-first structure, cliffhanger endings) tailored for Douyin/Kuaishou content formats.
4. Multi-Modal Asset Coordination: Synchronizes B-roll generation with dialogue timing through a shared timeline abstraction, ensuring visual cuts align with audio beats without manual keyframing.
5. Failsafe Rollback Mechanism: Agents maintain checkpoints at each production stage; if visual coherence checks fail, the system regenerates specific shots rather than entire sequences, reducing compute waste by ~60% compared to end-to-end regeneration.

Throughput Characteristics

As an agentic orchestration layer atop heavy video models, 融光's performance is bounded by inference costs rather than code efficiency:

Scalability Constraints

The architecture faces inherent bottlenecks in temporal consistency validation—as video length scales beyond 5 minutes, the combinatorial complexity of cross-scene coherence checks grows quadratically. Current implementation caps automated sequences at 20 scenes before requiring human-in-the-loop validation. Additionally, the Java backend's thread pool architecture limits concurrent project processing to ~10 active workflows per instance without horizontal scaling.

Competitive Positioning

Integration Landscape

• Model Backends: Pluggable architecture supports WAN 2.1 (Alibaba), CogVideoX (Zhipu), and commercial APIs via adapter pattern
• Distribution: Native export presets for Douyin, Kuaishou, and Xiaohongshu (Little Red Book) metadata formats
• Content Supply: Direct ingestion from novel/script platforms (likely targets Chinese web-novel IP conversion)

融光 occupies a unique niche: it's not competing with video models, but with the manual labor of prompt engineering and clip selection that current tools require. In the exploding Chinese short-drama market (projected $50B+ by 2026), this automation layer has immediate commercial utility.