HDR Effect VAE V3.0

Note: ALL METRICS use FP32 version - FP8 version is supplied for research only.

Bottom Line

HDR VAE is the recommended choice for approximately 80–90% of real-world use cases.

Although Base SDXL VAE maintains a small advantage in objective reconstruction metrics, HDR VAE consistently produces images that appear sharper, richer, more vibrant, and more professionally polished. The modest reduction in PSNR and SSIM is a worthwhile tradeoff for the substantial gains in perceived visual quality.

Full Evaluation

Note HDR VAE had a SSIM score of 0.8481 (Higher is better) vs SDXL BASE 0.8555 - A slight drop is expected for an enhancement VAE and this result is far better then prior versions.

A = SDXL Base VAE

B = HDR VAE version 3.0

Brightness & Contrast

• Brightness:

  • GT: 68.645

  • A: 68.928 (+0.28) → very close

  • B: 69.74 (+1.09) → noticeably brighter

  • A is closer to Ground Truth

• Contrast:

  • GT: 56.36

  • A: 55.56 (-0.80) → slightly flatter

  • B: 56.62 (+0.26) → slightly higher contrast

  • B better preserves contrast.

Sharpness & Detail

• Sharpness (Laplacian variance):

  • GT: 172.68

  • A: 102.12 (-70.56) → much softer

  • B: 165.55 (-7.13) → much closer to GT

  • Clear win for B. VAE-A loses a lot of sharpness.

• Gradient Strength:

  • A: 23.36 (-2.21) → weaker edges/gradients

  • B: 24.68 (-0.89) → better

  • B wins but had slight loss in GS from Ground Truth

• Edge Density:

  • A: 0.0190 (-24.8% fewer edges)

  • B: 0.0232 (only -7.3% fewer)

  • Strong win for B, SDXL was prone to flattening and this VAE seems to counter that.

Pattern: VAE-B preserves sharpness, edges, and fine details much better.

Color & Information Content

• Saturation:

  • GT: 118.96

  • A: 118.27 (almost identical)

  • B: 139.81 (+20.84) → significantly more saturated

  • B produces more vibrant colors.

• LAB Color Volume:

  • GT: 1.387M

  • A: 1.297M (-6.5%)

  • B: 1.822M (+37.9%) → much wider color gamut

  • Strong win for B.

• Quantized Colors (32 bins):

  • A: 2457 (-10.1%)

  • B: 3183 (+16.6%)

  • B wins.

• Unique Colors:

  • A: +1.4k

  • B: +17.0k

  • B wins (more color variety).

Information & Distribution

• Entropy:

  • A: 6.555 (+0.047) → slightly higher

  • B: 6.517 (+0.008) → very close to GT

  • Minor win for A (but not very meaningful here).

Dynamic Range

• A: 243.0 (slightly lower)

• B: 246.3 (slightly higher)

• Both very close to GT (244.3).

Key Takeaway

Base SDXL VAE achieves higher reconstruction accuracy and slightly stronger objective metrics such as PSNR and SSIM.

HDR VAE delivers significantly better perceived image quality, with major improvements in sharpness, detail retention, color vibrancy, and dynamic range.

While Base SDXL VAE wins on technical reconstruction accuracy, HDR VAE wins decisively on visual quality. For most generative AI workflows, perceived image quality is more important than small improvements in reconstruction metrics.

Recommended Use Cases

Use Base SDXL VAE when you need:

• Maximum numerical fidelity

• Scientific, medical, or archival imaging

• Close color matching to reference images

• Natural, subtle, or muted aesthetics

• High consistency across generations

• Extensive post-processing workflows

Use HDR VAE when you need:

• General-purpose image generation (recommended default)

• Photorealistic outputs with greater depth and impact

• Artistic or stylized generations

• Better detail retention in faces, textures, and architecture

• HDR-like contrast and richer color reproduction

• Strong visual appeal and presentation quality