Hi everyone, today we’re publishing the second alpha of our SDR-to-HDR conversion model, Hyperion.
We considered the helpful feedback from the previous alpha and have made the following changes.
The exposure control no longer linearly brightens every part of the scene.
In general, midtones will get somewhat brighter, but highlights will get significantly brighter. With this change, a dark scene enhanced with a high exposure setting will no longer look artificially brightened across the entire frame.
We’ve added a new slider to control highlights.
At low values, the slider tells the model to reconstruct washed out highlights, and at high values the model will increase the contrast in the existing highlights. This change makes colorspace expansion more consistent – reducing the color fluctuation across different shots/angles within a scene.
The model architecture is entirely different, with speeds up to 30 fps at 1080p on an NVIDIA RTX 4090.
Hyperion will no longer denoise or otherwise perform general enhancement, unlike in the first alpha.
We believe that with the new speed improvements, stacking an enhancement model (like Nyx) with Hyperion provides a better alternative with more precise controls. This change also resolves any warping artifacts seen in the previous alpha.
This model is designed to perform inverse tonemapping by increasing the bit depth, expanding the color gamut, and brightening the highlights of input videos.
This model improves contrast and recovers detail in highlight and shadow areas of standard dynamic range inputs, and produces HDR10 (BT.2020/PQ) output.
HDR playback is not currently supported in the app, and the output preview section has been updated for this model with an overlay informing users about this limitation. We recommend mpv player on Windows (mpv.io), or IINA on macOS (iina.io).
For optimal export settings, we recommend the H.265 codec’s Main10 profile for this model. Additionally, on macOS, exporting as ProRes 4444 XQ can result in higher fidelity (12-bit).
We’ve uploaded some updated samples of Hyperion’s output at different brightness settings to demonstrate the highlight expansion capabilities of the model:
Is it possible to use this model so you’d take rec709/sRGB SDR footage and then use the model to increase the dynamic range and export that into ProRes/EXR with a log/linear gamma rather than going to an HDR PQ (ST2084) BT.2020 format
@tony.topazlabs@kyle.topazlabs Try to add HDR metadata in output file too. Two Step Enhancement can be setup First Deniose Than SDR to HDR.Youtube Link of Squirel Done in Alpha 2 Prothere
Same Footage But Enhanced With RHEA and SDR To HDR here
Would be great to have that baked in, but fyi (in case it’s helpful to anyone reading) you can use DoVi_Scripts in the meantime to do it once you have the HDR version.
In case it’s useful to anyone, I threw together a little Python script, topaz_hdr.py, that uses the cli to encode the hdr rawvideo from topaz using an ffmpeg that has the x265 encoder. If you’re using windows you can do winget install ffmpeg to get a suitable ffmpeg binary.
it’s in the first post.
banding shouldn’t be a thing in 10 bit (HDR) content. That’s what this model is supposed to do. increase bitdepth and enhance color space.
HDR 10 metadata can be added to the output file.
Consider HDR range below a limit 500 (0-499)400 (0-399) My S2721DGF monitor is Display Visa HDR 400 but peak 480.So it would be even ready to be uploaded to Youtube.You Can Do what Youtube does .If monitor SDR it will play video in SDR otherwise HDR.Also you will need to detect the screen Apple SDR,XDR,HDR TV.TVs have Tone Mapping in them .On iMac Tony Alpha will only write High10 Into outfile in MOV (H265).
Main10 Profile is being Replaced With Main on Imy iMac when writing MP4 File Doing Main10 .
This Link has Resolve Studio Projects for the Your file at 200 Nits And 600 here
This is working for me, Alpha 1 did throw me errors when trying to run the SDR->HDR, Alpha 2 does not.
Although I think the actual output probably could be tuned.
To me the output kinda looks a bit oversaturated for some reason.
But I have faith that the model can and will be tweaked to produce a more natural looking image still with the HDR properties.
May you patch pause/resume functionnality ASAP please, it’s actually unusable
If you pause and resume you’ll lose >= 20% of done job
If you pause and resume + close the app. , you’ll lose >=30% of done job
Adjust exposure text area : behaviour completely crazy, turn into 800 or 1000 whatever you type, we have to backspace + delete tons of time to clear the area
Also but not blocking
Red shifting : doing more and more test to confirm that
At least @ 400 nits, 0 exposure, over exposed red colours
Trying @ 300 / 30 aso
I suspect a RAM memory leak. It is not completely restored when TVAI is closed. Also during its use, the used value increases, but not in TVAI. Any ffmpeg use may crash after TVAI use, system has to be rebooted.
At the moment I did not find the fine tuning to turn SDR to fine HDR. It mostly turned into more contrasted color space. Lights are fine, white also, but blacks and darks tone look clogged.
Depenging of the player, the output is either recognized in correct BT.2020 (case of Nvidia SHIELD) with Kodi 21, or incorrectly BT.709 in x86/x64 players (case of HTPC, RDNA2 GPU)
Both play usually flawlessly HDR “regular content”.
It looks like the output file does not meet Main10 HDR specifications.
(will give it a try on Vero 4K+ asap)
