This page is best viewed on HDR-capable displays with browsers like Safari or Chrome. Some browsers/systems cannot display the HDR MP4-stills images.
On the left: SDR Source // on the right: SDR-to-HDR with beeble.ai
Recently I saw some posts on LinkedIn about a tool that up-converts SDR footage to HDR using AI. I got in touch with the founder of beeble.ai, and he pointed me to an HDR-encoded demo video on YouTube.
(https://youtu.be/oUmxX8_MOxA?si=XRYAv4c3q-O3pnnp).
For my taste, the SDR footage in the videos often looks too flat, and the converted HDR footage looks bright, but overall too dark. So I wanted to test the tool with some of my own footage.
I rediscovered these two HDR test clips I created in 2021 using bracketed photos and graded in FCPX. Back then, my workflow was as follows:
- Merge the bracketed photos with Photomatix Pro to linear ProPhoto EXR files (not a recommended workflow anymore).
- Export an ARRI LogC ProRes 444 file from Nuke.
- Grade the footage in FCPX in HDR mode (linear Rec.2020 PQ)
- At that time, on a 500 nits iMac 27″ and check on an iPad Pro XDR at 1000 nits.
Here is the idea for a test that came to my mind:
- Converting the HDR grade to SDR in FCPX with HDRTools.
- Export a high-quality H.264 file from each clip and use that as the source for beeble.ai.
- Set a highlight threshold in beeble.ai and process the two scenes to generate the EXR AP-O output.
- In Resolve, create a side-by-side comparison of the original HDR, converted SDR, and upconverted HDR footage.
Overall, I am not really interested in AI generating new content in clipped areas of an image. I find it much more intriguing to observe how the AI determines which pixels receive the highest values in the newly generated HDR content derived from the SDR source.
I’ve seen every demo footage processed by beeble.ai so far, and it’s always ended up quite dark, even darker than the original. To fix this, the EXR files I received from beeble.ai have an exposure compensation of plus one stop.
I created two comparison sheets in Resolve that show two points of view. In both contact sheets, the original HDR footage is on the left, the SDR content is in the middle, and the beeble.ai SDR to HDR conversion is on the right.
For the initial contact sheet, I created a Resolve project with a PQ output in a non-RCM environment. This setup highlights the significant contrast between the SDR and HDR content.
- The HDR images on the left looks like intended
- In the middle, the SDR images are inside a PQ container that clips at 100 nits. Naturally, it appears less vibrant than the HDR on the left.
- The HDR result from beeble.ai on the right reveals more detail in the headlights but suffers from a lack of overall light intensity. This is particularly noticeable in the light reflection on the sidewalk from another car behind the Prius.
- In the overall appearance, I find the middle and right results closer to each other; only the left HDR source sticks out (as it hopefully should)
Another approach is to examine the results in a Resolve ACES project using an HDR ODT. The SDR content appears remarkably similar to both the HDR stills on the left and right.
- The HDR image on the left looks like intended.
- In the middle, the SDR image version employs the inverse Display Rec.709 IDT. This results in values reaching around 16, which allows for a surprisingly good match to the HDR look on the left. However, the SDR material appears less bright and slightly less colourful compared to the HDR master.
- On the right is again the beeble.ai result. The main differences are the detail in the headlights (which was generated) and the lack of brightness in the light reflection behind the Prius.
Finally, here are both contact sheets as HDR-encoded clips on YouTube.