Upload README.md
Browse files
README.md
CHANGED
|
@@ -24,7 +24,7 @@ license: apache-2.0
|
|
| 24 |
|
| 25 |
|
| 26 |
## Updates
|
| 27 |
-
|
| 28 |
- [2026/06] Release the inference code and pretrained weights 🎉
|
| 29 |
|
| 30 |
|
|
@@ -32,7 +32,6 @@ license: apache-2.0
|
|
| 32 |
|
| 33 |
|
| 34 |
## ✨ Highlights
|
| 35 |
-
---
|
| 36 |
|
| 37 |
- **Mask-free shifted-window self-attention (MFSWA).** Each spatial window is **pre-gathered into a dense tensor**, so every attention call reduces to a single standard scaled-dot-product (SDPA) call — *no attention mask, cyclic shift, or padding ever enters the graph*. This gives a **1.62× throughput gain over its full-attention teacher** at essentially identical quality, with **no dedicated sparse kernel**.
|
| 38 |
- **Restoration-aware Autoencoder (ReAE).** A lightweight encoder–decoder jointly fine-tuned with the DiT in pixel space removes the heavy-3D-VAE / tiled-decoding bottleneck.
|
|
@@ -40,7 +39,6 @@ license: apache-2.0
|
|
| 40 |
|
| 41 |
|
| 42 |
## 📊 Results
|
| 43 |
-
---
|
| 44 |
|
| 45 |
### Efficiency at 2560×1440 (single H100, causal streaming, 24 frames)
|
| 46 |
|
|
@@ -59,7 +57,6 @@ license: apache-2.0
|
|
| 59 |
|
| 60 |
|
| 61 |
## 🛠 Installation
|
| 62 |
-
---
|
| 63 |
|
| 64 |
```bash
|
| 65 |
git clone https://github.com/Holiday/SwiftVR.git
|
|
@@ -86,7 +83,7 @@ pip install -e .
|
|
| 86 |
|
| 87 |
|
| 88 |
## 🗂 Model Zoo
|
| 89 |
-
|
| 90 |
| Model Name | Date | Backbone | Link |
|
| 91 |
|---|---|---|---|
|
| 92 |
| SwiftVR | 2026.06 | Wan2.2-TI2V-5B | [🤗 HuggingFace](https://huggingface.co/H-oliday/SwiftVR) |
|
|
@@ -108,7 +105,7 @@ checkpoints/
|
|
| 108 |
|
| 109 |
|
| 110 |
## 🚀 Quick Start
|
| 111 |
-
|
| 112 |
### Python API
|
| 113 |
|
| 114 |
```python
|
|
@@ -161,7 +158,7 @@ Use `--png` to write a PNG sequence.
|
|
| 161 |
|
| 162 |
|
| 163 |
## 🎬 More Visual Results
|
| 164 |
-
|
| 165 |
> Full-length restored clips (low-quality input → SwiftVR, played back to back).
|
| 166 |
|
| 167 |
|
|
@@ -175,7 +172,7 @@ Use `--png` to write a PNG sequence.
|
|
| 175 |
|
| 176 |
|
| 177 |
## 🙏 Acknowledgements
|
| 178 |
-
|
| 179 |
SwiftVR builds on [Wan2.2-TI2V-5B](https://github.com/Wan-Video), the lightweight autoencoder [TAEHV](https://github.com/madebyollin/taehv), and the [RealBasicVSR](https://github.com/ckkelvinchan/RealBasicVSR) degradation pipeline. We thank the authors of [DOVE](https://github.com/zhengchen1999/DOVE), [SeedVR2](https://github.com/ByteDance-Seed/SeedVR), and [FlashVSR](https://github.com/OpenImagingLab/FlashVSR) for releasing strong baselines, and the [UltraVideo](https://github.com/Tele-AI/UltraVideo) team for the training corpus.
|
| 180 |
|
| 181 |
|
|
|
|
| 24 |
|
| 25 |
|
| 26 |
## Updates
|
| 27 |
+
|
| 28 |
- [2026/06] Release the inference code and pretrained weights 🎉
|
| 29 |
|
| 30 |
|
|
|
|
| 32 |
|
| 33 |
|
| 34 |
## ✨ Highlights
|
|
|
|
| 35 |
|
| 36 |
- **Mask-free shifted-window self-attention (MFSWA).** Each spatial window is **pre-gathered into a dense tensor**, so every attention call reduces to a single standard scaled-dot-product (SDPA) call — *no attention mask, cyclic shift, or padding ever enters the graph*. This gives a **1.62× throughput gain over its full-attention teacher** at essentially identical quality, with **no dedicated sparse kernel**.
|
| 37 |
- **Restoration-aware Autoencoder (ReAE).** A lightweight encoder–decoder jointly fine-tuned with the DiT in pixel space removes the heavy-3D-VAE / tiled-decoding bottleneck.
|
|
|
|
| 39 |
|
| 40 |
|
| 41 |
## 📊 Results
|
|
|
|
| 42 |
|
| 43 |
### Efficiency at 2560×1440 (single H100, causal streaming, 24 frames)
|
| 44 |
|
|
|
|
| 57 |
|
| 58 |
|
| 59 |
## 🛠 Installation
|
|
|
|
| 60 |
|
| 61 |
```bash
|
| 62 |
git clone https://github.com/Holiday/SwiftVR.git
|
|
|
|
| 83 |
|
| 84 |
|
| 85 |
## 🗂 Model Zoo
|
| 86 |
+
|
| 87 |
| Model Name | Date | Backbone | Link |
|
| 88 |
|---|---|---|---|
|
| 89 |
| SwiftVR | 2026.06 | Wan2.2-TI2V-5B | [🤗 HuggingFace](https://huggingface.co/H-oliday/SwiftVR) |
|
|
|
|
| 105 |
|
| 106 |
|
| 107 |
## 🚀 Quick Start
|
| 108 |
+
|
| 109 |
### Python API
|
| 110 |
|
| 111 |
```python
|
|
|
|
| 158 |
|
| 159 |
|
| 160 |
## 🎬 More Visual Results
|
| 161 |
+
|
| 162 |
> Full-length restored clips (low-quality input → SwiftVR, played back to back).
|
| 163 |
|
| 164 |
|
|
|
|
| 172 |
|
| 173 |
|
| 174 |
## 🙏 Acknowledgements
|
| 175 |
+
|
| 176 |
SwiftVR builds on [Wan2.2-TI2V-5B](https://github.com/Wan-Video), the lightweight autoencoder [TAEHV](https://github.com/madebyollin/taehv), and the [RealBasicVSR](https://github.com/ckkelvinchan/RealBasicVSR) degradation pipeline. We thank the authors of [DOVE](https://github.com/zhengchen1999/DOVE), [SeedVR2](https://github.com/ByteDance-Seed/SeedVR), and [FlashVSR](https://github.com/OpenImagingLab/FlashVSR) for releasing strong baselines, and the [UltraVideo](https://github.com/Tele-AI/UltraVideo) team for the training corpus.
|
| 177 |
|
| 178 |
|