pyultrahdr

Turn any ordinary SDR image into a real Ultra HDR JPEG — in pure Python, no libultrahdr required.

Before — plain SDR	After — Ultra HDR JPEG

Open this README in Chrome on an HDR display (Windows 11 24H2 / macOS Sonoma / Android 14+ / iOS 17.4+) to see the right image light up. On an SDR display the two look identical — that's the point: Ultra HDR degrades gracefully everywhere and pops wherever HDR is available.

The output is a standard JPEG on the surface, but on an HDR display the highlights punch up to ~1000 nits. On SDR displays it looks like a perfectly normal photo. One file, two worlds.

python sdr_to_hdr.py your_photo.jpg
# → your_photo_ultrahdr.jpg

Open that file in Chrome on an HDR monitor. The sun will actually be bright.

Why this exists

Google's Ultra HDR format (also standardised as ISO 21496-1) is the easiest way to ship HDR photos on the web today — it's just a regular JPEG plus an embedded gain map, so it degrades gracefully everywhere. But the only production-quality tooling is Google's C++ libultrahdr, which is painful to build, wrap, and deploy — especially on Windows.

pyultrahdr is a ~230-line Python script that produces byte-compatible Ultra HDR JPEGs from any SDR input. Zero C dependencies, zero build system, just numpy and Pillow. Drop it into a pipeline, a Jupyter notebook, or a batch job.

Features

• Pure Python. numpy + Pillow. Works on Windows/macOS/Linux out of the box.
• Byte-compatible MPF/XMP. The Multi-Picture Format APP2 segment and hdrgm: XMP metadata are structured to match Google's reference samples — verified by hex-diffing against real Ultra HDR files from the wild.
• Monotonic inverse tone mapping. Every pixel gets a visible boost; no knee-curve artefacts that leave mid-tones looking flat.
• BT.709 → BT.2020 gamut expansion in linear light.
• Configurable peak brightness (default 1000 nits) and SDR reference white (default 203 nits, per BT.2408).
• Single-file script — copy, paste, ship.

Install

pip install numpy pillow

Then grab the script:

curl -O https://raw.githubusercontent.com/hanfeisun/pyultrahdr/main/sdr_to_hdr.py

Usage

python sdr_to_hdr.py input.png

Flags:

Flag	Default	What it does
--peak-nits	1000	Target HDR peak brightness
--sdr-white	203	SDR reference white in nits (BT.2408)
--quality	92	Base JPEG quality (1-100)
--gainmap-quality	85	Gain map JPEG quality (1-100)

Batch convert a folder:

from pathlib import Path
import subprocess, sys

for p in Path('photos').glob('*.jpg'):
    subprocess.run([sys.executable, 'sdr_to_hdr.py', str(p)])

How to view the result

• Chrome 116+ on an HDR-capable display (Windows 11 24H2, macOS Sonoma+, ChromeOS). Make sure HDR is toggled on in your OS display settings.
• Android 14+ Gallery / Google Photos.
• iOS 17.4+ Photos.
• Windows 11 Photos app (24H2 and later).

You can test your own HDR chain with samples from gregbenzphotography.com/hdr-demo or the MishaalRahmanGH/Ultra_HDR_Samples repo.

How it works

 SDR sRGB 8-bit
      │
      ▼  sRGB EOTF
 linear [0,1]
      │
      ▼  inverse tone mapping:  boost(L) = s + (p - s) · L^γ
 HDR in nits
      │
      ▼  BT.709 → BT.2020 matrix
 HDR BT.2020 (nits)
      │
      ▼  per-pixel log2(HDR / SDR)  → gain map
 gain map (grayscale, 8-bit)
      │
      ▼  assemble:  SOI + XMP(hdrgm:) + MPF(APP2) + baseJPEG + gainmapJPEG
 Ultra HDR JPEG

The resulting file has:

1. A normal baseline JPEG (your SDR image, re-encoded).
2. An XMP APP1 block with Adobe's hdrgm:1.0 namespace describing the gain map parameters (GainMapMin, GainMapMax, HDRCapacityMin, HDRCapacityMax, OffsetSDR, OffsetHDR).
3. An MPF APP2 block (CIPA DC-007 Multi-Picture Format) pointing from the primary image to the secondary (gain map) image.
4. The gain map itself as a second JPEG concatenated to the end.

An HDR-aware decoder reads the gain map, applies it per-pixel (in log2-linear space) to reconstruct HDR radiance, and sends PQ/HDR10 signal to the display. An SDR decoder just sees the first JPEG and ignores the rest.

Why not just a PQ PNG / AVIF / JXL?

• PNG/AVIF/JXL with PQ works, but only the handful of apps that understand HDR containers will render correctly. Everywhere else the image looks washed out.
• Ultra HDR degrades gracefully to a normal SDR JPEG everywhere, and lights up on HDR-capable platforms. That's why the web is converging on it.

Limitations

• The inverse tone mapping is heuristic. It looks great on natural photos and screenshots, but it's not a learned model — don't expect it to recover clipped highlights that were never captured. If you need a learned ITM, look at ExpandNet, HDRCNN, or SingleHDR and feed their output into this script's assembly stage.
• The gain map is single-channel (luminance). RGB gain maps are allowed by the spec but rare in the wild.

License

MIT. See LICENSE.

References

• Google, Ultra HDR Image Format v1.0
• ISO/IEC 21496-1 — Gain map metadata for image conversion
• Adobe, Gain Map Specification (hdrgm:)
• CIPA DC-007 — Multi-Picture Format
• ITU-R BT.2408 — SDR reference white at 203 nits
• Banterle et al., Advanced High Dynamic Range Imaging (inverse tone mapping background)

---

If this saved you a day of wrestling with C++ toolchains, star the repo. PRs welcome.