SDR to HDR — ACEScg Converter

Turn any 8-bit image into a production-ready ACEScg EXR. AI-reconstructed highlights, physically plausible dynamic range, ready to light a scene.

pip install .
sdr-hdr photo.png --mode model -o photo_acescg.exr

Weights download automatically. Tile size adapts to your GPU. Output is half-float RGBA EXR with premultiplied alpha — drops straight into Nuke, RV, 3ds Max, Houdini, or any ACES-aware app.

Install

git clone https://github.com/Geddart/SDR_HDR.git
cd SDR_HDR
pip install .

Python 3.10+, PyTorch 2.0+. CUDA GPU with 8+ GB VRAM for model mode.

Usage

# AI highlight reconstruction (recommended — needs GPU)
sdr-hdr photo.png --mode model -o photo_acescg.exr

# Fast math-only conversion (no GPU needed)
sdr-hdr photo.png -o photo_acescg.exr

# Batch convert a folder
sdr-hdr ./textures/ -o ./textures_hdr/

Scene Presets

Presets are tuned for common scene types. Pick the one that matches your source and go.

sdr-hdr photo.png --preset day -o photo_acescg.exr
sdr-hdr photo.png --preset night -o photo_acescg.exr
sdr-hdr photo.png --preset hdri -o photo_acescg.exr

# Round-trip: restore ACEScg data that was saved as sRGB PNG
sdr-hdr texture.png --preset roundtrip -o texture_acescg.exr

Preset	Peak	Gain	Power	What it's for
night	150	3.0	2	Night exteriors — point lights very hot, deep shadows
day	30	4.0	2	Daylight — lower contrast, ambient-filled shadows
interior	60	3.5	2	Indoor scenes — windows and practicals moderately hot
overcast	15	5.0	2	Flat lighting — gentle expansion, lifted shadows
hdri	250	3.0	2	Environment maps for CG lighting — very hot light sources
roundtrip	1	1.0	50	Restore ACEScg data saved as sRGB PNG — no expansion, no dither

--peak, --gain, and --power override preset values if you need to fine-tune:

# Start with night preset but cap highlights lower
sdr-hdr photo.png --preset night --peak 100

Without a preset, defaults are peak=150, gain=3.0, power=2.

Two Modes

Model Mode (--mode model)

A Refusion-HDR diffusion model (ConditionalNAFNet + IRSDE) runs 100 reverse diffusion steps to reconstruct what was behind clipped highlights. PU21 decodes to absolute luminance, then the same luminance-preserving inverse tonemap as linear mode expands the dynamic range using --peak and --gain. Both modes produce matching output ranges — the model contributes smarter highlight reconstruction.

Weights (~292 MB) auto-download on first run. Tile size auto-adapts to your VRAM.

Linear Mode (default)

Pure math, no AI. Fast, runs on CPU.

sRGB → EOTF decode → TPDF dither → luminance-based inverse tonemap → Rec.709 → ACEScg

Applies hdr = Y * (gain + (peak - gain) * Y^power) to luminance only, scales RGB proportionally. Preserves chromaticity, prevents solarization. --power controls how steeply highlights ramp up: low power (2) spreads expansion across the range, high power (50) concentrates it near white. TPDF dithering kills 8-bit banding before the nonlinear expansion; disable with --no-dither for clean round-trips.

Parameters

sdr-hdr input -o output [options]

Flag	Default	Description
--mode	linear	linear or model
--preset	—	Scene preset: night, day, interior, overcast, hdri, roundtrip
--exposure	0.0	Exposure adjustment in stops
--peak	150	Max HDR value — or set by preset
--gain	3.0	Mid-tone multiplier — or set by preset
--tile-size	auto	Tile dimension, auto from VRAM (model mode)
--overlap	64	Tile overlap for blending (model mode)
--power	2	Highlight rolloff steepness — or set by preset
--no-dither	off	Disable TPDF dithering (cleaner for round-trips)
--weights	auto	Path to model checkpoint (model mode)

How It Works

All color math is pure PyTorch on GPU — no OCIO, no oiiotool.

• Input: sRGB (Rec.709 primaries, sRGB EOTF)
• Output: ACEScg (AP1 primaries, D60 white point, scene-linear)
• Color transform: 3x3 matrix with Bradford chromatic adaptation D65 → D60
• Dithering: TPDF at ±1 LSB before nonlinear expansion
• Alpha: Premultiplied, preserved from source

Project Structure

SDR_HDR/
├── convert.py              # CLI entry point + presets
├── pyproject.toml          # pip install .
├── pipeline/
│   ├── inference.py        # Conversion pipeline (linear + model)
│   ├── colorspace.py       # sRGB EOTF, Rec.709→ACEScg, inverse tonemap
│   ├── exr_writer.py       # Half-float RGBA EXR output
│   └── download.py         # Auto-download model weights
├── models/
│   ├── nafnet.py           # ConditionalNAFNet (Refusion-HDR)
│   ├── sde.py              # IRSDE diffusion engine
│   └── pu21.py             # PU21 perceptual encoder/decoder
├── tests/                  # 53 tests
└── weights/                # Auto-downloaded, gitignored

License

MIT

Acknowledgments

• Refusion-HDR — ConditionalNAFNet + IRSDE architecture (AIM 2025 ITM Challenge)
• PU21 — Perceptually Uniform encoding (Mantiuk & Azimi 2021)