Hyperion 2.5

Hyperion 2.5 is Topaz Labs' SDR to HDR video model, released on 18 August 2026. Topaz describes it as automatically redistributing luminance and colour across highlights, midtones and shadows while preserving fine detail, then expanding the available colour and luminance range. In practice that means inverse tone mapping, bit-depth expansion, and a conversion from Rec.709 gamma to BT.2100 PQ on wider primaries. It generates nothing and adds no resolution.

Topaz publishes two output formats for this model: 10-bit ProRes in BT.2020 PQ and 16-bit half-float EXR in ACES 2065-1. ProRes is the editorial and grading format; EXR is what a compositor expects to receive. The Topaz Video desktop app offers a wider codec list than the hosted model does, so check which route you are using before promising a delivery format.

No conversion controls. The original Hyperion filter exposed Adjust Exposure, Boost Saturation and a Highlight Threshold that defaulted to 0.70, and those were removed in Hyperion 2 in favour of automatic highlight classification. What you still choose is the output format, codec and bit depth.

Not the clipped part. Those values are absent from the file and no model can read data that was never written. What the conversion does is redistribute what survived across a much wider range, so highlights that still hold detail gain separation. Topaz states the limit directly: on video that is too bright and blown out there can be too little range to build upon, and its own guidance calls the result faux HDR.

Interlaced footage converts poorly and should be deinterlaced first. Very dark or overexposed source leaves the conversion almost no range to work with. Moving shots with inconsistent lighting can show frame-to-frame variation, and compression artifacts get expanded along with real detail, which is why any cleanup or upscaling belongs before the conversion rather than after it.

Ask Copilot to convert a clip to HDR, or pick Hyperion 2.5 from the model library and attach the video. There are no conversion settings to fill in, so the only choice is the output format. Because it sits alongside the generation and upscale models, a clip can be generated, upscaled and converted without leaving the conversation.