Diffusers
Safetensors
Felldude's picture
Update README.md
d102dfa verified
|
Raw
History Blame Contribute Delete
3.11 kB
---
license: apache-2.0
base_model:
- Wan-AI/Wan2.2-Animate-14B
- nvidia/Cosmos3-Super
- nvidia/Cosmos3-Nano
---
# VAE Reconstruction & Color Expansion Evaluation
### Wan 2.2 Base HDR VAE vs Ground Truth (GT)
---
## ๐Ÿ“Š At-a-Glance Summary
**Color expansion (Wan 2.2 Base HDR VAE vs GT):**
- LAB volume: **~+34.09%**
- Saturation: **~+22.27%**
- Unique colors: **~+23.68%**
- Quantized colors: **~+30.32%**
- Sharpness: **~+4.25%**
**Structural cost:**
- SSIM: **~-1.42%**
- PSNR: **~-2.31%**
- MSE: **~+13.09%**
### Key takeaway
Wan 2.2 Base HDR VAE significantly increases perceptual color richness and detail density,
while introducing a mild reduction in structural fidelity and a moderate increase in reconstruction error relative to ground truth.
---
## ๐Ÿงช Evaluation Setup
- Task: Image reconstruction (VAE decode comparison)
- Model:
- Wan 2.2 Base HDR VAE
- Reference:
- Ground Truth (GT)
- Mode: Deterministic image reconstruction evaluation
- Metrics:
- Color statistics (HSV / LAB / RGB)
- Structural similarity (SSIM)
- Pixel fidelity (PSNR, MSE)
- Texture features (edges, gradients, sharpness)
- Entropy and color diversity measures
---
## ๐Ÿ“ˆ Key Findings
### ๐ŸŽจ Color behavior
- Wan 2.2 Base HDR VAE significantly expands GT color space
- Strong increases in saturation and LAB distribution volume
- Noticeably richer and more expressive chromatic output
### ๐Ÿงฑ Structure
- Slight reduction in SSIM and PSNR indicates mild structural deviation from GT
- Increased sharpness suggests enhanced edge emphasis and micro-detail amplification
### ๐Ÿ‘๏ธ Perceptual behavior
- Outputs appear more vivid and detailed than GT
- Trade-off between realism fidelity and perceptual enhancement
---
## ๐Ÿ“Š Quantitative Results vs GT
### Wan 2.2 Base HDR VAE
- Brightness: **+0.79%**
- Contrast: **+0.73%**
- Saturation: **+22.27%**
- Entropy: **+0.13%**
- Dynamic range: **-0.13%**
- Edge density: **-1.41%**
- Gradient strength: **-0.68%**
- LAB color volume: **+34.09%**
- Quantized colors: **+30.32%**
- Unique colors: **+23.68%**
- Sharpness: **+4.25%**
---
## ๐Ÿงฎ Reconstruction Quality
### Wan 2.2 Base HDR VAE vs GT
- SSIM: **-1.42%**
- PSNR: **-2.31%**
- MSE: **+13.09%**
โžก๏ธ Slight reduction in structural fidelity with moderate increase in reconstruction error
---
## ๐Ÿง  Interpretation
**Color axis โ†’ Strong winner: Wan 2.2 Base HDR VAE**
- Large gains in LAB volume, saturation, and color diversity
**Structure axis โ†’ Slight trade-off**
- Small degradation in SSIM/PSNR balanced by increased sharpness
**Perceptual axis โ†’ Winner: Wan 2.2 Base HDR VAE**
- More visually rich and detailed outputs compared to GT
---
## ๐Ÿ Final Conclusion
Wan 2.2 Base HDR VAE acts as a **perceptual and chromatic enhancement VAE**, expanding color richness and visible detail density beyond ground truth.
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.