Upload moss-video-realtime-sft
Browse files
README.md
CHANGED
|
@@ -31,7 +31,9 @@ This checkpoint is intended for:
|
|
| 31 |
- **Continuous video monitoring** and instant action feedback.
|
| 32 |
|
| 33 |
#### Model Architecture
|
| 34 |
-
|
|
|
|
|
|
|
| 35 |
**MOSS-Video-Preview-Real-Time-SFT** is the flagship model of the series, featuring a **Pioneering Image-Video Unified Cross-Attention Architecture** optimized for streaming:
|
| 36 |
|
| 37 |
- **Native Unified Design**: Unlike traditional models, our architecture supports native frame-by-frame video injection, ensuring the visual context is always up-to-date with the generation process.
|
|
@@ -44,6 +46,10 @@ For architecture diagrams and full system details, see the top-level repository:
|
|
| 44 |
|
| 45 |
The core advantage of this model is its **Asynchronous Streaming Capability**, enabling true "See-and-Say" video intelligence.
|
| 46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
- **Asynchronous Single-Frame Streaming**: Video frames are injected at a stable frequency. The input pipeline is **non-blocking** and decoupled from text generation, ensuring continuous perception.
|
| 48 |
- **Persistent State Maintenance**: Leveraging **Cross-Attention KV Cache** and temporal positional encoding, the model maintains long-range contextual dependencies across continuous frames.
|
| 49 |
- **Instantaneous Streaming Response**: Built on the optimized `MllamaVideoModel`, it performs autoregressive generation alongside the visual stream, achieving ultra-low **Time to First Token (TTFT)**.
|
|
|
|
| 31 |
- **Continuous video monitoring** and instant action feedback.
|
| 32 |
|
| 33 |
#### Model Architecture
|
| 34 |
+
<p align="center">
|
| 35 |
+
<img src="assets/model_structure.png" width="90%" alt="Model Architecture"/>
|
| 36 |
+
</p>
|
| 37 |
**MOSS-Video-Preview-Real-Time-SFT** is the flagship model of the series, featuring a **Pioneering Image-Video Unified Cross-Attention Architecture** optimized for streaming:
|
| 38 |
|
| 39 |
- **Native Unified Design**: Unlike traditional models, our architecture supports native frame-by-frame video injection, ensuring the visual context is always up-to-date with the generation process.
|
|
|
|
| 46 |
|
| 47 |
The core advantage of this model is its **Asynchronous Streaming Capability**, enabling true "See-and-Say" video intelligence.
|
| 48 |
|
| 49 |
+
<p align="center">
|
| 50 |
+
<img src="assets/streaming_structure.jpeg" width="90%" alt="Streaming Inference Mechanism"/>
|
| 51 |
+
</p>
|
| 52 |
+
|
| 53 |
- **Asynchronous Single-Frame Streaming**: Video frames are injected at a stable frequency. The input pipeline is **non-blocking** and decoupled from text generation, ensuring continuous perception.
|
| 54 |
- **Persistent State Maintenance**: Leveraging **Cross-Attention KV Cache** and temporal positional encoding, the model maintains long-range contextual dependencies across continuous frames.
|
| 55 |
- **Instantaneous Streaming Response**: Built on the optimized `MllamaVideoModel`, it performs autoregressive generation alongside the visual stream, achieving ultra-low **Time to First Token (TTFT)**.
|