Update README.md
Browse files
README.md
CHANGED
|
@@ -8,6 +8,8 @@ LocalDiT builds upon the architecture of [PixArt-α](https://huggingface.co/PixA
|
|
| 8 |
- **Parameters**: 0.52B
|
| 9 |
- **Resolution**: Supports generation up to 1024×1024 pixels
|
| 10 |
- **Language Support**: English text prompts
|
|
|
|
|
|
|
| 11 |
|
| 12 |
# Usage
|
| 13 |
Details on code execution will be released at a later date.
|
|
@@ -33,6 +35,10 @@ image.save("generated_image.png")
|
|
| 33 |
- Implemented window-based local attention in alternating transformer blocks
|
| 34 |
- Reduced parameter count through efficient attention design
|
| 35 |
- Optimized for both quality and computational efficiency
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
# Results
|
| 38 |
LocalDiT achieves comparable image quality to PixArt-α while offering:
|
|
|
|
| 8 |
- **Parameters**: 0.52B
|
| 9 |
- **Resolution**: Supports generation up to 1024×1024 pixels
|
| 10 |
- **Language Support**: English text prompts
|
| 11 |
+
- **Text Encoder**: FLAN-T5-XXL (4.3B parameters)
|
| 12 |
+
- **VAE**: SDXL VAE for high-quality latent encoding/decoding
|
| 13 |
|
| 14 |
# Usage
|
| 15 |
Details on code execution will be released at a later date.
|
|
|
|
| 35 |
- Implemented window-based local attention in alternating transformer blocks
|
| 36 |
- Reduced parameter count through efficient attention design
|
| 37 |
- Optimized for both quality and computational efficiency
|
| 38 |
+
- **Components**:
|
| 39 |
+
- Diffusion Backbone: Custom LocalDiT architecture (0.52B parameters)
|
| 40 |
+
- Text Encoder: FLAN-T5-XXL (4.3B parameters) for rich text embedding
|
| 41 |
+
- VAE: SDXL's Variational Autoencoder for high-fidelity latent space encoding/decoding
|
| 42 |
|
| 43 |
# Results
|
| 44 |
LocalDiT achieves comparable image quality to PixArt-α while offering:
|