HDR VAE v3

HDR VAE v3 is the most capable release in the HDR VAE series to date, delivering substantial improvements in perceived image quality while maintaining strong reconstruction fidelity.

Compared to the standard SDXL VAE, HDR VAE v3 focuses on preserving detail, expanding color reproduction, and increasing visual impact. The result is sharper images, richer colors, improved edge retention, and a noticeably larger effective color gamut.

These gains are visible in standard workflows and become even more pronounced when working in high bit-depth pipelines.


Highlights

✨ Sharper details and textures

🎨 Expanded color volume and color diversity

🌈 Richer saturation without excessive clipping

πŸ“ˆ Improved edge preservation

πŸ” Better retention of fine image structure

⚑ Slightly increased dynamic range and contrast


At a Glance

Category Winner
Reconstruction Accuracy SDXL Base VAE
Sharpness HDR VAE v3
Edge Preservation HDR VAE v3
Color Gamut HDR VAE v3
Color Diversity HDR VAE v3
Saturation HDR VAE v3
Visual Impact HDR VAE v3
Scientific Fidelity SDXL Base VAE

The base SDXL VAE remains slightly stronger for pure reconstruction metrics, while HDR VAE v3 produces significantly more visually compelling outputs.


Evaluation Results

Reconstruction Fidelity

HDR VAE v3 remains very close to the original SDXL VAE in structural accuracy.

Model SSIM
SDXL Base VAE 0.8555
HDR VAE v3 0.8481

A small SSIM reduction is expected for enhancement-oriented VAEs. In exchange, HDR VAE v3 delivers substantial gains in perceived image quality.


Detail Preservation

One of the largest improvements in HDR VAE v3 is the retention of fine detail.

Sharpness (Laplacian Variance)

Model Value
Ground Truth 172.68
SDXL Base VAE 102.12
HDR VAE v3 165.55

HDR VAE v3 remains remarkably close to the source image while avoiding the softening commonly seen with the default SDXL VAE.

Edge Density

Model Difference from Ground Truth
SDXL Base VAE -24.8%
HDR VAE v3 -7.3%

The improvement in edge retention helps preserve texture, architectural detail, hair, fabric, and other fine structures.


Color Performance

Color reproduction is where HDR VAE v3 shows its largest gains.

LAB Color Volume

Model Value
Ground Truth 1.387M
SDXL Base VAE 1.297M
HDR VAE v3 1.822M

This represents roughly a 38% increase in color volume over the source image.

Saturation

Model Value
Ground Truth 118.96
SDXL Base VAE 118.27
HDR VAE v3 139.81

Colors appear more vibrant, richer, and more expressive while maintaining overall image coherence.

Color Diversity

Metric SDXL Base VAE HDR VAE v3
Quantized Colors -10.1% +16.6%
Unique Colors +1.4K +17.0K

HDR VAE v3 consistently produces a wider variety of distinguishable colors throughout the image.


Dynamic Range & Contrast

HDR VAE v3 introduces modest but measurable improvements in image depth and contrast.

Metric Ground Truth SDXL Base VAE HDR VAE v3
Brightness 68.645 68.928 69.740
Contrast 56.36 55.56 56.62
Dynamic Range 244.3 243.0 246.3

These changes contribute to a stronger HDR-like appearance without dramatically altering image balance.


Recommended Usage

Choose SDXL Base VAE if you need

  • Maximum reconstruction fidelity
  • Scientific or medical imaging
  • Archival preservation workflows
  • Exact reference matching
  • Conservative image processing

Choose HDR VAE v3 if you need

  • A better default VAE for image generation
  • Stronger photorealism
  • Sharper textures and facial details
  • Richer color reproduction
  • Improved environmental and architectural renders
  • More visually striking outputs
  • Enhanced presentation quality

Conclusion

HDR VAE v3 intentionally trades a small amount of reconstruction accuracy for substantial gains in perceived image quality.

While the SDXL Base VAE remains the better choice for strict numerical fidelity, HDR VAE v3 delivers superior sharpness, edge retention, color reproduction, and overall visual impact, making it the recommended choice for most generative image workflows.

Downloads last month
22
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Felldude/SDXL-HDR-VAE

Finetuned
(10)
this model