| --- |
| license: bigscience-openrail-m |
| base_model: |
| - MiniMaxAI/MiniMax-H3-Image |
| --- |
| |
| **NOTE Encoder was Frozen Using this VAE for Training should not effect outcome** |
|
|
| ## Overview |
|
|
| This repository contains evaluation results for a **MiniMax Image HDR VAE (High Dynamic Range Variational Autoencoder)** designed for MiniMax-H3 image reconstruction and enhancement workflows. |
|
|
| The HDR VAE is evaluated against the **BASE MiniMax Image VAE** using a multi-domain reconstruction analysis covering: |
|
|
| * perceptual similarity |
| * structural energy enhancement |
| * color distribution changes |
| * saturation enhancement |
| * sharpness and high-frequency detail |
| * channel-level color differences |
|
|
| The objective is not only to measure pixel reconstruction accuracy, but to evaluate how the VAE enhances image characteristics through the encode/decode pipeline. |
|
|
| ### Image Enhancement |
|
|
| Measured through reconstruction comparisons: |
|
|
| * RGB pixel deviation |
| * saturation changes |
| * gradient energy |
| * sharpness retention |
| * contrast gain |
| * high-frequency energy |
| * perceptual color difference (DeltaE) |
|
|
| These evaluate how reconstruction changes image detail, color intensity, sharpness, and structural information. |
|
|
| --- |
|
|
| # Results Summary |
|
|
| ## BASE MiniMax Image VAE vs HDR VAE |
|
|
| | Metric | BASE MiniMax VAE | HDR VAE | |
| | ----------------------------- | ---------------: | ------: | |
| | PSNR ↑ | **32.4389** | 28.4966 | |
| | SSIM ↑ | 0.87133 | **0.87810** | |
| | RGB MAE ↓ | **0.015623** | 0.028214 | |
| | Saturation MAE ↓ | **0.034885** | 0.091786 | |
| | Gradient Energy Retention | 0.98014 | **1.05029** | |
| | Sharpness Retention | **0.96952** | 0.91151 | |
| | Contrast Retention | 1.01329 | 1.00179 | |
| | High-Frequency Retention | 0.96159 | **1.04110** | |
| | DeltaE ↓ | **2.7502** | 7.0100 | |
|
|
| --- |
|
|
| # Interpretation |
|
|
| The **BASE MiniMax Image VAE** achieves lower pixel reconstruction error and DeltaE, indicating closer numerical agreement with the original input. |
|
|
| The **HDR VAE** demonstrates a different reconstruction profile: |
|
|
| * increased gradient energy |
| * increased high-frequency energy |
| * higher structural similarity |
| * stronger saturation changes |
| * enhancement-oriented reconstruction behavior |
|
|
| The HDR VAE behaves as an enhancement-oriented reconstruction operator rather than a pixel-preservation model, producing measurable changes to image characteristics during reconstruction. |
|
|
| --- |
|
|
| # Design Goal |
|
|
| The HDR VAE is designed to improve image characteristics that standard VAEs commonly soften or reduce: |
|
|
| * fine texture |
| * high-frequency detail |
| * local contrast |
| * color variation |
| * edge definition |
| * perceived sharpness |
|
|
| Rather than optimizing only for pixel-level similarity, the HDR VAE emphasizes image enhancement through decoder fine-tuning. |
|
|
| --- |
|
|
| # Conclusion |
|
|
| The evaluation demonstrates two different reconstruction profiles: |
|
|
| **BASE MiniMax Image VAE** |
|
|
| * lower PSNR error |
| * lower RGB reconstruction error |
| * lower DeltaE |
| * stronger numerical color preservation |
| * more faithful pixel reconstruction |
|
|
| **HDR MiniMax Image VAE** |
|
|
| * higher SSIM |
| * increased gradient energy |
| * increased high-frequency energy |
| * enhancement-oriented reconstruction |
| * greater image characteristic variation |
|
|
| The HDR VAE provides an enhancement-oriented reconstruction profile for users seeking increased structural information, richer image characteristics, and enhanced visual detail rather than strict pixel-perfect reconstruction. |
