import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Felldude/FLUX.2-HDR-VAE", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]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.
- Downloads last month
- 23