README.md

---
license: other
language:
  - en
library_name: diffusers
pipeline_tag: text-to-image
tags:
  - flux
  - flux2
  - vae
  - hdr
  - ultrahd
  - cinematic
  - high-detail
  - image-enhancement
  - texture-recovery
  - sharpness-enhancement
  - color-enhancement
  - microcontrast
  - raw
  - generative-ai
  - stable-diffusion
  - diffusion
  - latent-space
  - photorealistic
  - neural-compression
  - reconstruction
  - fp16
  - inference
  - image-generation
  - visual-quality
  - machine-learning
  - ai-art
  - flux-compatible
model_type: Variational Autoencoder (VAE)
datasets:
  - internal-evaluation-suite
metrics:
  - lpips
  - gradient-energy
  - unique-colors
  - brightness-bias
  - contrast-gain
  - rgb-channel-shift
inference: true
---

# FLUX.2 UltraHD RAW VAE

## Overview

FLUX.2 UltraHD RAW VAE is a high-fidelity enhancement-focused Variational Autoencoder engineered for the FLUX.2 ecosystem. Unlike reconstruction-pure VAEs optimized exclusively for perceptual similarity, UltraHD RAW VAE is tuned for cinematic rendering, HDR enhancement, sharper edge response, texture recovery, and richer gradient reproduction.

The model is fully compatible with all FLUX.2 variants and other models including:

- FLUX.2
- FLUX.2 Klein 9B & 4B
- Ernie

---

# Scientific Assessment

## Design Objective

The primary design goal of UltraHD RAW VAE is not strict latent reconstruction accuracy, but perceptual enhancement quality.

Key optimization targets include:

- Increased texture fidelity
- Enhanced microcontrast
- Expanded color richness
- Sharper edge gradients
- Improved HDR-style rendering
- Stable luminance preservation
- Dynamic range consistency

---

# Quick Assessment

## Observed Visual Characteristics

### Color Enhancement

UltraHD RAW VAE increases effective color richness through:

- Higher unique color utilization
- Expanded HDR gradient smoothness
- Slight saturation enhancement
- Preserved luminance integrity
- Preserved dynamic range behavior

### Detail Recovery

The VAE demonstrates:

- Stronger edge energy
- Sharper fine detail reconstruction
- Increased perceived texture depth
- Enhanced microcontrast behavior

---

# Training Notes

Although the visual enhancement effect may appear subtler compared to previous enhancement VAEs, the underlying optimization process was substantially more difficult.

More than double-digit failed training attempts were required to achieve:

- Stable HDR enhancement
- Low LPIPS divergence
- Contrast preservation
- Brightness neutrality
- Balanced channel behavior

Achieving an LPIPS value near the already extremely low FLUX.2 baseline proved exceptionally challenging.

---

# Direct Metric Comparison

## Evaluation Setup

- 150-image benchmark comparison
- Tested directly against Base FLUX.2 VAE
- Metrics computed across identical latent reconstruction conditions

---

# LPIPS (Perceptual Similarity)

| Model | LPIPS |
|---|---|
| FLUX Base | 0.0073 |
| UltraHD RAW VAE | 0.0303 |

### Difference

- 4.1× higher than FLUX Base

### Interpretation

The model intentionally trades strict reconstruction fidelity for perceptual enhancement quality.

Despite the increase, both values remain significantly below typical human perceptual thresholds.

---

# Gradient Energy

| Model | Gradient Energy |
|---|---|
| FLUX Base | 405.5 |
| UltraHD RAW VAE | 633.5 |

### Difference

- +56% increase

### Interpretation

Substantially sharper detail rendering and stronger edge definition.

---

# Unique Colors

| Model | Unique Colors |
|---|---|
| FLUX Base | 37,369 |
| UltraHD RAW VAE | 39,184 |

### Difference

- +4.9% increase

### Interpretation

Improved color richness and smoother HDR tonal transitions.

---

# Brightness Bias

| Model | Brightness Bias |
|---|---|
| FLUX Base | 0.0047 |
| UltraHD RAW VAE | 0.0030 |

### Interpretation

Excellent luminance preservation with slightly improved brightness neutrality.

---

# Contrast Gain

| Model | Contrast Gain |
|---|---|
| FLUX Base | 0.984 |
| UltraHD RAW VAE | 0.983 |

### Interpretation

Near-identical dynamic range and HDR tone retention.

---

# RGB Channel Shift Analysis

## Red Shift

| Model | Red Shift |
|---|---|
| FLUX Base | +0.0048 |
| UltraHD RAW VAE | +0.0188 |

Interpretation:

- Slightly warmer red response
- Improved cinematic warmth

---

## Green Shift

| Model | Green Shift |
|---|---|
| FLUX Base | -0.0037 |
| UltraHD RAW VAE | -0.0104 |

Interpretation:

- Minor reduction in green balance
- Helps reduce sterile tonal appearance

---

## Blue Shift

| Model | Blue Shift |
|---|---|
| FLUX Base | +0.0181 |
| UltraHD RAW VAE | +0.0434 |

Interpretation:

- Stronger blue/cyan HDR emphasis
- Enhanced atmospheric rendering

---

# Summary

FLUX.2 UltraHD RAW VAE is an enhancement-oriented VAE optimized for:

- HDR rendering
- Texture recovery
- Cinematic sharpness
- Richer gradients
- Enhanced microcontrast
- Improved perceptual detail

Compared to the base FLUX VAE, UltraHD RAW VAE produces:

- Significantly stronger edge detail
- Improved color richness
- Better texture clarity
- Enhanced HDR-style rendering
- Stable brightness preservation
- Preserved contrast behavior

The model is best suited for users prioritizing perceptual image quality and cinematic rendering over strict pixel-perfect reconstruction fidelity.

---

# Recommended Usage

Recommended for:

- Photorealistic generations
- Cinematic compositions
- HDR workflows
- Texture-heavy scenes
- High-detail portrait rendering
- Atmospheric lighting
- Stylized realism

Less suitable for:

- Pixel-faithful reconstruction tasks
- Scientific image preservation
- Reconstruction benchmarking applications

---

# Compatibility

Compatible with:

- FLUX.2
- FLUX.2 Klein
- Ernie

Precision support:

- FP16
- BF16
- FP32

---

# Disclaimer

This VAE intentionally modifies reconstruction characteristics to improve perceptual aesthetics. Metric increases in LPIPS are expected and are part of the enhancement-oriented design philosophy.
--