---
license: bigscience-openrail-m
---

**NOTE Encoder was not trained this VAE will likely not effect training in anyway**

## Overview

This repository contains evaluation results for an **HDR VAE (High Dynamic Range Variational Autoencoder)** designed for SDXL image reconstruction workflows.

The HDR VAE is evaluated against the **BASE SDXL VAE** using a multi-domain reconstruction analysis covering:

* perceptual similarity
* structural energy preservation
* color distribution retention
* photometric stability
* channel-level color drift

The objective is not only to measure pixel reconstruction accuracy, but to evaluate how well the VAE preserves the complete visual signal of an image through the encode/decode pipeline.

### Photometric Stability

Measured through calibrated marker regions:

* brightness bias
* contrast gain
* RGB channel drift

These evaluate whether reconstruction maintains luminance and color relationships.

---

# Results Summary

## BASE SDXL VAE vs HDR VAE

| Metric                | BASE SDXL VAE |     HDR VAE |
| --------------------- | ------------: | ----------: |
| LPIPS ↓               |    **0.0321** |      0.0593 |
| Gradient Energy Ratio |         0.792 |   **1.167** |
| Color Support Ratio   |         0.951 |   **0.968** |
| Brightness Bias ↓     |       0.00101 | **0.00030** |
| Contrast Error ↓      |       0.00928 | **0.00439** |
| RGB Drift ↓           |        0.1608 |  **0.1035** |

---

# Interpretation

The **BASE SDXL VAE** achieves lower LPIPS error, indicating stronger agreement in learned perceptual feature space. Its reconstruction behavior favors smooth perceptual similarity.

The **HDR VAE** preserves more measurable image information:

* higher structural energy retention
* improved color support preservation
* stronger contrast stability
* reduced channel drift
* improved photometric consistency

The HDR VAE behaves as a more information-preserving reconstruction operator, maintaining image characteristics that are often reduced during standard VAE compression.

---

# Design Goal

The HDR VAE is designed to improve reconstruction fidelity in areas where standard VAEs commonly lose information:

* fine texture
* high-frequency detail
* local contrast
* color variation
* HDR-like tonal relationships

Rather than optimizing only for perceptual closeness, the HDR VAE emphasizes preservation of the original image signal through the latent representation.

---

# Conclusion

The evaluation demonstrates two different reconstruction profiles:

**BASE SDXL VAE**

* lower LPIPS
* smoother reconstruction behavior
* stronger perceptual averaging

**HDR VAE**

* improved structural retention
* improved color preservation
* improved photometric accuracy
* greater signal preservation

The HDR VAE provides a reconstruction profile optimized for users requiring higher information retention, improved detail preservation, and more stable image characteristics through SDXL latent encoding and decoding.