StreamChar
Model Overview
StreamChar is a undisclosed-parameter model developed by Alibaba. Released on 2026-05-25.
📊 Quick Specs
| Specification | Value |
|---|---|
| Parameters | undisclosed |
| Task | video-generation |
| Modality | text, audio, video |
| License | Other/Custom |
| Type | research-preview |
✨ Key Features
- Decoupled Orchestration Architecture: Separates global semantic planning (LLM orchestrator) from local short-window audio-video DiT
- Progress-Aware Pointer: Maps transcript text directly to continuous audio and video generation, preventing speech-text drift
- Sink-Chunk Memory Mechanism: Employs persistent visual anchors ('sink chunks') to eliminate identity and quality drift
- Two-Stage Decoupled Distillation: Combines DMD step-compression with online chunk rollout fine-tuning for low latency
- Real-Time Single-GPU Streaming: Continuous streaming inference on a single NVIDIA H100 GPU (1.34s latency per 33-frame chunk)
🔗 Resources
- Hugging Face: StreamChar
- GitHub: Repository
- Paper: arXiv
- Website: Project Page
📜 License & Access
Other/Custom — See repository for specific license details.
Key Features
Decoupled Orchestration Architecture: Separates global semantic planning (LLM orchestrator) from local short-window audio-video DiT
Progress-Aware Pointer: Maps transcript text directly to continuous audio and video generation, preventing speech-text drift
Sink-Chunk Memory Mechanism: Employs persistent visual anchors ('sink chunks') to eliminate identity and quality drift
Two-Stage Decoupled Distillation: Combines DMD step-compression with online chunk rollout fine-tuning for low latency
Real-Time Single-GPU Streaming: Continuous streaming inference on a single NVIDIA H100 GPU (1.34s latency per 33-frame chunk)
You might also want to compare
Verified Sources
Tags
Model Specs
Parameters
Undisclosed
Context Window
undisclosed
License
Other/Custom
Deployment
Resources & Links
Curator Notes
Partially enriched via migration on 2026-07-25. Manual review recommended.
Compare Specs
Compare parameters, context windows, modalities, and benchmark scores of this model side-by-side with others.
Compare Model