Instructions to use Felldude/Wan2.1-Diffusers-HDR-VAE with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use Felldude/Wan2.1-Diffusers-HDR-VAE with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("Felldude/Wan2.1-Diffusers-HDR-VAE", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
| license: apache-2.0 | |
| base_model: | |
| - Wan-AI/Wan2.1-T2V-14B | |
| # VAE Reconstruction & Color Expansion Evaluation | |
| ### Wan2.1 Base VAE vs HDR VAE | |
| --- | |
| ## ๐ At-a-Glance Summary | |
| **Color expansion (HDR VAE vs GT):** | |
| - LAB volume: **~+50%** | |
| - Saturation: **~+27%** | |
| - Unique colors: **~+30%** | |
| - Gradient strength: **~+7%** | |
| **Structural cost:** | |
| - SSIM: **~-2%** | |
| - MSE: **~+50%** | |
| ### Key takeaway | |
| HDR VAE significantly increases chromatic richness and color diversity, | |
| but reduces structural fidelity and increases reconstruction error relative to Wan2.1 Base VAE. | |
| --- | |
| ## ๐งช Evaluation Setup | |
| - Task: Image reconstruction (VAE decoding comparison) | |
| - Models: | |
| - Wan2.1 Base VAE | |
| - HDR VAE | |
| - Ground Truth (GT) | |
| - Mode: Deterministic image-to-image reconstruction | |
| - Metrics: | |
| - Color statistics (HSV / LAB / RGB) | |
| - Structural metrics (SSIM, MSE, PSNR) | |
| - Frequency features (edges, gradients) | |
| - Entropy + color diversity | |
| --- | |
| ## ๐ Key Findings | |
| ### ๐จ Color behavior | |
| - **Wan2.1 Base VAE:** slight compression of GT color space | |
| - **HDR VAE:** strong expansion of color space (+30% to +50%) | |
| ### ๐งฑ Structure | |
| - **Wan2.1 Base VAE:** closer to GT, smoother output | |
| - **HDR VAE:** sharper, more high-frequency artifacts | |
| ### ๐๏ธ Perceptual behavior | |
| - **Wan2.1 Base VAE:** higher fidelity, lower distortion | |
| - **HDR VAE:** more vivid but less faithful reconstruction | |
| --- | |
| ## ๐ Quantitative Results vs GT | |
| ### Wan2.1 Base VAE | |
| - Brightness: -0.71% | |
| - Contrast: -0.23% | |
| - Saturation: -0.11% | |
| - Entropy: -0.08% | |
| - Dynamic range: -0.46% | |
| - Edge density: -15.35% | |
| - Gradient strength: -5.90% | |
| - LAB color volume: -5.00% | |
| - Quantized colors: -6.60% | |
| - Unique colors: +7.90% | |
| - Sharpness: -45.65% | |
| --- | |
| ### HDR VAE | |
| - Brightness: +0.90% | |
| - Contrast: +0.93% | |
| - Saturation: +26.90% | |
| - Entropy: +0.17% | |
| - Dynamic range: -0.07% | |
| - Edge density: +14.98% | |
| - Gradient strength: +7.10% | |
| - LAB color volume: +48.60% | |
| - Quantized colors: +45.60% | |
| - Unique colors: +29.50% | |
| - Sharpness: +32.15% | |
| --- | |
| ## ๐งฎ Reconstruction Quality | |
| ### Wan2.1 Base VAE | |
| - SSIM: **0.822** | |
| - PSNR: **25.86** | |
| - MSE: **180.88** | |
| โก๏ธ Best structural fidelity and lowest reconstruction error | |
| --- | |
| ### HDR VAE | |
| - SSIM: **0.806** | |
| - PSNR: **24.06** | |
| - MSE: **278.08** | |
| โก๏ธ Higher distortion but stronger chromatic expansion | |
| --- | |
| ## ๐ง Interpretation | |
| **Color axis โ Winner: HDR VAE** | |
| - Largest increase in LAB volume, saturation, and color diversity | |
| **Structure axis โ Winner: Wan2.1 Base VAE** | |
| - Closer SSIM and lower MSE | |
| **Fidelity axis โ Winner: Wan2.1 Base VAE** | |
| - More faithful reconstruction of GT distribution | |
| --- | |
| ## ๐ Final Conclusion | |
| HDR VAE acts as a **chromatic expansion decoder**, increasing color space occupancy and high-frequency detail. | |
| Wan2.1 Base VAE remains closer to the **ground-truth manifold**, prioritizing structural and perceptual fidelity over color amplification. |