How Windows Auto HDR Uses AI and DirectX

Windows Auto HDR upgrades legacy Standard Dynamic Range (SDR) games to High Dynamic Range (HDR) automatically, without requiring developer updates. By combining lightweight machine learning algorithms with DirectX runtime hooks, Windows analyzes 8-bit SDR output frames in real time and intelligently remaps color values, luminance, and dynamic range into a 10-bit or 16-bit HDR color space. This process brightens specular highlights and deepens color saturation while preserving the original artistic intent and user interface clarity.

Intercepting the Pipeline with DirectX Hooks

Auto HDR operates at the operating system and graphics driver level rather than inside the game's internal game engine. It targets games built on DirectX 11 and DirectX 12.

When an eligible game runs, Windows uses presentation-level hooks via the DirectX Graphics Infrastructure (DXGI). As the game completes its render loop and prepares to present the final frame, DXGI intercepts the standard 8-bit swapchain buffer (typically formatted as DXGI_FORMAT_B8G8R8A8_UNORM or R8G8B8A8_UNORM).

Instead of sending this limited 8-bit frame directly to the display, the system upgrades the swapchain to a high-precision format—typically a 10-bit integer format (DXGI_FORMAT_R10G10B10A2_UNORM) or a 16-bit floating-point format (DXGI_FORMAT_R16G16B16A16_FLOAT, often used with scRGB). This provides the necessary headroom to hold expanded color and luminance data without introducing banding artifacts.

The Role of Machine Learning

Traditional tone-mapping tools use simple math to stretch luminance evenly across a frame, which often causes blown-out UI text or washed-out midtones. Auto HDR avoids this by utilizing a deep neural network trained on thousands of side-by-side SDR and native HDR game frames.

During training, the model learned how game engines construct scenes and how lighting physics behave in native HDR:

  • Semantic Classification: The neural network recognizes the difference between diffuse surfaces (such as concrete, skin, or cloth), intense emissive light sources (such as neon signs, fire, sunlight, and laser blasts), and static 2D HUD/UI elements.
  • Selective Intensity Calculation: Because the model understands scene context, it knows to leave UI elements and paper-white diffuse values at typical SDR brightness (around 100 to 200 nits) while boosting targeted light sources to modern HDR peak thresholds (often reaching 1,000 nits or more depending on the calibrated display).

For live deployment, Microsoft distilled this complex training into an optimized, hardware-accelerated post-processing shader. This inferencing model executes in fractions of a millisecond directly on the GPU, avoiding noticeable input latency or frame-rate drops.

Expanding the 8-Bit Color Space

Standard 8-bit SDR buffers are restricted to the narrow Rec. 709/sRGB color gamut, offering only 256 steps per RGB channel and a low dynamic range.

To convert this limited data into wide color gamut (WCG) HDR:

  1. Bit-Depth Expansion: The 8-bit integer values are uncompressed into floating-point color representations. This eliminates rounding errors during mathematical transformations.
  2. Color Gamut Remapping: The sRGB color coordinates are mapped into wide-gamut spaces, specifically BT.2020 or linear scRGB. The system calculates where saturated colors should shift along the wider chromaticity curve without distorting natural skin tones or neutral colors.
  3. Luminance De-Quantization: The inferred brightness values are plotted along the Perceptual Quantizer (PQ / SMPTE ST 2084) curve. By distributing luminance logarithmically rather than linearly, the system allocates more precision to high-contrast areas.

Once the frame is transformed and tone-mapped to match the specific peak brightness and black-level capabilities of the connected HDR display, DXGI presents the final frame to the monitor, producing an authentic HDR image from an SDR source.