DiffHDR Converts 8-Bit Video to HDR and Generates Detail in Clipped Highlights and Shadows

DiffHDR, a research framework from Texas A&M University, Eyeline Labs, and Netflix, converts ordinary 8-bit low dynamic range video into HDR by generating plausible radiance in regions the original footage blew out or crushed to black. The method also supports controllable re-exposure, giving post teams latitude to re-expose footage that previously held no recoverable detail.

• It generates detail rather than recovering it. DiffHDR synthesizes plausible highlight and shadow radiance in over- and underexposed regions; that detail is inferred by the model, not restored camera-original sensor data.
• It treats LDR-to-HDR as inpainting. The method formulates the conversion as generative radiance inpainting in the latent space of a video diffusion model, operating in Log-Gamma color space.
• Re-exposure can be steered. Conversion can be guided by text prompts or reference images, so the same clip can be pushed toward different exposures or looks.
• Training data is synthetic. To get around the scarcity of paired HDR video, the team built a pipeline that synthesizes HDR training footage from static HDRI maps.
• Code and models are public. DiffHDR ships with a paper, code, and model weights, and is listed for presentation at ECCV 2026.

How DiffHDR fills in radiance the footage never captured

Most video lives in 8-bit LDR, where saturation and quantization discard the scene's original high dynamic range. Once a sky clips to white or a shadow falls to black, that information is gone, which caps how far a colorist can push exposure and blocks accurate luminance mapping to an HDR display.

DiffHDR works around that by leaning on the spatio-temporal priors already baked into a pretrained video diffusion model. Instead of stretching the existing signal, it paints new radiance into the blown and crushed regions while recovering continuous scene radiance across the frame. This is generative SDR-to-HDR conversion and re-exposure, not spatial upscaling. The resolution stays the same; what changes is the dynamic range and the invented detail inside it.

DiffHDR teaser comparing low-dynamic-range input with controllable HDR re-exposure results. Credit: DiffHDR project team / Texas A&M University, Eyeline Labs, and Netflix.

Because the highlight and shadow detail is synthesized, it should be treated as plausible reconstruction rather than a faithful record of what the sensor saw. For workflows where fidelity to the original capture matters, that distinction matters.

Text prompts and reference images control the look

The controllable side is where DiffHDR diverges from a one-button conversion. The researchers show re-exposure guided by text prompts or reference images, letting an operator direct how the recovered highlights and shadows resolve rather than accepting a single deterministic output.

The project page also demonstrates the method on in-the-wild video, movie footage, and synthetic defocus, alongside still and video comparisons against prior LDR-to-HDR approaches. The team reports that DiffHDR improves radiance fidelity and temporal stability over state-of-the-art methods; those are author-reported experimental results and have not been independently verified.

Where it sits among the HDR conversion tools

DiffHDR lands in a crowded lane of 8-bit-to-HDR work already moving into production pipelines. It is closest to systems that reconstruct HDR from existing LDR footage, like Beeble's SwitchHDR, which pulls 16-bit HDR from 8-bit video.

A parallel reconstruction tool, Topaz Labs' Hyperion 2.5, converts 8-bit video into 10-bit ProRes and 16-bit EXR.

That is a different problem from generating HDR at the source. Tools like LTX-2.5 produce native HDR as part of the generation step, so there is no LDR footage to reconstruct from. DiffHDR belongs with the conversion group. It takes finished 8-bit clips, real or generated, and expands them after the fact. With the code and weights posted publicly, teams can test how the inferred detail holds up against the reconstruction tools already in the pipeline.