Ultimate 48bit Tool
When training with 48bit images you must use FP32 with no 8bit optimizer as this rule:
float16: collapses everything toward a fixed noise floor (~1e-7)
float32: preserves the original image quantization structure almost perfectly
48-bit advantage only survives in float32 (and higher), not in float16
This custom model was trained on a dataset exceeding 22 GB, achieving strong evaluation scores that indicate meaningful HDR reconstruction rather than simple gradient smoothing or noise-based approximation.
Images enhanced with this tool require only minimal adjustments to integrate with common training frameworks such as Onetrainer or Kohya SS. Standard tooling such as OpenCV and widely used VAE-based workflows already support higher bit-depth processing, making the model compatible with existing image-to-image and diffusion training setups with little modification.
Standard 8-bit pipelines that normalize inputs by dividing by 255, Using 48bit images you can use normalization by 65,535 during conversion to float32. This preserves significantly more tonal precision, allowing finer gradations between pixel values and improved reconstruction fidelity.
16-bit = 256× more levels per pixel when cast to FP32
That is a 25,500% increase in precision (approx.)
NOTE: KEEP PIL out of your python OPEN-CV ONLY NO PILLOW
Perceptual Metrics (60-image test set)
LPIPS (Alex) 0.044 Low perceptual distance / high similarity
Gradient Energy 0.088 → 0.108 Preserved fine detail, slight sharpening
FFT Structure Score 1.07 → 1.23 Improved high-frequency retention
Histogram Continuity 11.2 → 11.3 Stable tonal distribution
MAE Distribution (8-bit → 16-bit reconstruction)
MAE Range Accuracy Comment Percent (%)
≥1000 Occasionally visible in uniform areas 1.06
600–1000 Almost never visible 10.03
400–600 Fully imperceptible 27.39
200–400 Near perfect 59.95
≤200 Near exact scientific 1.57
I can not repeat this enough DO NOT EVER use pillow with 48bit images, do not even import it.
For Onetrainer change the MGDS in your VENV Source (LoadImage.py)
For Kohya change the latent creation process force use of Open CV (I forget the .py)
ALLWAYS change normalization Logic. You have to divide by 65k
# If 8-bit, normalize to [0, 1]; if 16-bit, divide by 65535
if img_tensor.dtype == torch.uint8:
img_tensor = img_tensor.float() / 255.0
elif img_tensor.dtype == torch.uint16:
img_tensor = img_tensor.float() / 65535.0
else:
img_tensor = img_tensor.float()