Instructions to use Felldude/Wan2.2-Diffusers-HDR-VAE with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use Felldude/Wan2.2-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/Wan2.2-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
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,115 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
+
base_model:
|
| 4 |
+
- Wan-AI/Wan2.2-Animate-14B
|
| 5 |
+
- nvidia/Cosmos3-Super
|
| 6 |
+
- nvidia/Cosmos3-Nano
|
| 7 |
---
|
| 8 |
+
|
| 9 |
+
# VAE Reconstruction & Color Expansion Evaluation
|
| 10 |
+
### Wan 2.2 Base HDR VAE vs Ground Truth (GT)
|
| 11 |
+
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
## ๐ At-a-Glance Summary
|
| 15 |
+
|
| 16 |
+
**Color expansion (Wan 2.2 Base HDR VAE vs GT):**
|
| 17 |
+
- LAB volume: **~+34.09%**
|
| 18 |
+
- Saturation: **~+22.27%**
|
| 19 |
+
- Unique colors: **~+23.68%**
|
| 20 |
+
- Quantized colors: **~+30.32%**
|
| 21 |
+
- Sharpness: **~+4.25%**
|
| 22 |
+
|
| 23 |
+
**Structural cost:**
|
| 24 |
+
- SSIM: **~-1.42%**
|
| 25 |
+
- PSNR: **~-2.31%**
|
| 26 |
+
- MSE: **~+13.09%**
|
| 27 |
+
|
| 28 |
+
### Key takeaway
|
| 29 |
+
Wan 2.2 Base HDR VAE significantly increases perceptual color richness and detail density,
|
| 30 |
+
while introducing a mild reduction in structural fidelity and a moderate increase in reconstruction error relative to ground truth.
|
| 31 |
+
|
| 32 |
+
---
|
| 33 |
+
|
| 34 |
+
## ๐งช Evaluation Setup
|
| 35 |
+
|
| 36 |
+
- Task: Image reconstruction (VAE decode comparison)
|
| 37 |
+
- Model:
|
| 38 |
+
- Wan 2.2 Base HDR VAE
|
| 39 |
+
- Reference:
|
| 40 |
+
- Ground Truth (GT)
|
| 41 |
+
- Mode: Deterministic image reconstruction evaluation
|
| 42 |
+
- Metrics:
|
| 43 |
+
- Color statistics (HSV / LAB / RGB)
|
| 44 |
+
- Structural similarity (SSIM)
|
| 45 |
+
- Pixel fidelity (PSNR, MSE)
|
| 46 |
+
- Texture features (edges, gradients, sharpness)
|
| 47 |
+
- Entropy and color diversity measures
|
| 48 |
+
|
| 49 |
+
---
|
| 50 |
+
|
| 51 |
+
## ๐ Key Findings
|
| 52 |
+
|
| 53 |
+
### ๐จ Color behavior
|
| 54 |
+
- Wan 2.2 Base HDR VAE significantly expands GT color space
|
| 55 |
+
- Strong increases in saturation and LAB distribution volume
|
| 56 |
+
- Noticeably richer and more expressive chromatic output
|
| 57 |
+
|
| 58 |
+
### ๐งฑ Structure
|
| 59 |
+
- Slight reduction in SSIM and PSNR indicates mild structural deviation from GT
|
| 60 |
+
- Increased sharpness suggests enhanced edge emphasis and micro-detail amplification
|
| 61 |
+
|
| 62 |
+
### ๐๏ธ Perceptual behavior
|
| 63 |
+
- Outputs appear more vivid and detailed than GT
|
| 64 |
+
- Trade-off between realism fidelity and perceptual enhancement
|
| 65 |
+
|
| 66 |
+
---
|
| 67 |
+
|
| 68 |
+
## ๐ Quantitative Results vs GT
|
| 69 |
+
|
| 70 |
+
### Wan 2.2 Base HDR VAE
|
| 71 |
+
|
| 72 |
+
- Brightness: **+0.79%**
|
| 73 |
+
- Contrast: **+0.73%**
|
| 74 |
+
- Saturation: **+22.27%**
|
| 75 |
+
- Entropy: **+0.13%**
|
| 76 |
+
- Dynamic range: **-0.13%**
|
| 77 |
+
- Edge density: **-1.41%**
|
| 78 |
+
- Gradient strength: **-0.68%**
|
| 79 |
+
- LAB color volume: **+34.09%**
|
| 80 |
+
- Quantized colors: **+30.32%**
|
| 81 |
+
- Unique colors: **+23.68%**
|
| 82 |
+
- Sharpness: **+4.25%**
|
| 83 |
+
|
| 84 |
+
---
|
| 85 |
+
|
| 86 |
+
## ๐งฎ Reconstruction Quality
|
| 87 |
+
|
| 88 |
+
### Wan 2.2 Base HDR VAE vs GT
|
| 89 |
+
|
| 90 |
+
- SSIM: **-1.42%**
|
| 91 |
+
- PSNR: **-2.31%**
|
| 92 |
+
- MSE: **+13.09%**
|
| 93 |
+
|
| 94 |
+
โก๏ธ Slight reduction in structural fidelity with moderate increase in reconstruction error
|
| 95 |
+
|
| 96 |
+
---
|
| 97 |
+
|
| 98 |
+
## ๐ง Interpretation
|
| 99 |
+
|
| 100 |
+
**Color axis โ Strong winner: Wan 2.2 Base HDR VAE**
|
| 101 |
+
- Large gains in LAB volume, saturation, and color diversity
|
| 102 |
+
|
| 103 |
+
**Structure axis โ Slight trade-off**
|
| 104 |
+
- Small degradation in SSIM/PSNR balanced by increased sharpness
|
| 105 |
+
|
| 106 |
+
**Perceptual axis โ Winner: Wan 2.2 Base HDR VAE**
|
| 107 |
+
- More visually rich and detailed outputs compared to GT
|
| 108 |
+
|
| 109 |
+
---
|
| 110 |
+
|
| 111 |
+
## ๐ Final Conclusion
|
| 112 |
+
|
| 113 |
+
Wan 2.2 Base HDR VAE acts as a **perceptual and chromatic enhancement VAE**, expanding color richness and visible detail density beyond ground truth.
|
| 114 |
+
|
| 115 |
+
It trades a small amount of structural fidelity for significantly improved perceptual richness, making it more suitable for downstream generative pipelines where visual expressiveness is prioritized over strict reconstruction accuracy.
|