Obviously, the SDR-to-HDR option applies machine learning to increase the bit-depth of the input video if necessary, but does it do the same to the chroma sampling as well? For example, if I run an 8-bit 4:2:0 video file through one of the Hyperion models, and my output codec is ProRes 422 HQ, am I getting fully-calculated 10-bit 4:2:2 video in the final output? Or am I getting 10-bit 4:2:0 video in a 4:2:2 container?
Interesting question, following it for Topaz’ answer 
The model is actually doing 16-bit 4:4:4. For ProRes 422 HQ, the results are downsampled to 10-bit 4:2:2.
All of the GAN-based models actually also process at 16-bit 4:4:4 and get converted to 4:2:0 or 4:4:4, depending on the encoding settings chosen.
Excellent!
Shouldn’t this be better documented?
We’re talking about HDR, which relies heavily on high bit depth and, in some workflows, full 4:4:4 chroma to preserve its full visual potential.
ALL the GAN-based models? That’s basically every local model that isn’t Stabilization, Motion Deblur, or Frame Interpolation, isn’t it?
Sorry, meant to say 4:4:4 for GAN models output and then converted to 4:2:0 or 4:2:2 as set by encoders. This would be the enhancement models aside from the new Starlight family.
GAN (Generative Adversarial Network)
An architectural framework based on a duel between a Generator and a Discriminator to invent sharp textures (sometimes adding artifacts or “hallucinated” details, like Iris or some specific face/text refiners).
CNNs (Convolutional Neural Networks), or Temporal Transformers. These models are strictly “predictive” because they analyze existing neighboring pixels or surrounding frames to mathematically deduce and reconstruct the missing data without “generating” brand-new artificial noise or textures out of thin air.
→ Starlight
So, let me make sure I understand this: The non-Starlight enhancement models DO upsample the chroma, but Hyperion does NOT? So I would need to do an enhancement to expand the pixel format before running Hyperion if I want 4:4:4 HDR?
The output codec will determine what the downsample will be. If you wish to maintain 4:4:4, you would need to export in a codec that supports that, either ProRes or DNxHR, depending on whether you are on Mac or Windows.
Is FFV1 excluded from these combinations?
The user might have a (very) large drive at home.
Fair point, I forgot about FFV1 as that can handle 12-bit 4:4:4 for export as well.
Okay, sorry for the repetitive question, but I want to make sure I’m not misinterpreting you: Hyperion, by itself, WILL give me (say) true neural-upsampled 4:2:2 or 4:4:4 pixel output from a 4:2:0 input, so long as I specify such a format in the export settings with a proper codec, correct?
Had to ask for more clarity on this situation to be sure I did not give wrong information as I am not a true color scientist and have spent minimal time playing with the HDR workflows at this point. Here is what I got back from the team amd I am quoting it.
“If there is an input of a 4:2:0, and a user outputs a codec with 4:4:4, this will be more of an expanded output of information with more shades of the existing colors for up to 4:4:4. That being said it’s best to assume that in most cases it would be somewhere between the two. Moving up to 4:4:4 gives the clip the highest probability to be a true 4:4:4 but not necessarily fully equate to it. This still gives the video output more room for compositing and color workflows.”
Thank you. So the GAN models do full neural upsampling of chroma while Hyperion all by itself does a kind of fill-in function that isn’t just padding but isn’t what the GAN models do either. I guess that’s why the original Hyperion support page suggests combining it with enhancement models.* Thank you again, and sorry that you had to answer the same question multiple different ways.
*On an unrelated note, the way the filters are laid out in the most recent release of Topaz Video makes it impossible to run an Enhancement and SDR-to-HDR model at the same time. Because Starlight Precise 2.5 and Hyperion 2 are the models at the top of each list, both automatically close all other functions when highlighted due to being standalone models, and the program doesn’t remember the last selected model, so a Catch-22 happens: selecting SDR-to-HDR will always disable Enhancement and vice-versa, making it impossible to pair the original Hyperion with an enhancement model.
Welcome, just wanted to make sure that we got the information clearly laid out for everyone.
You can enable the Hyperion model, swap to Hyperion 1, then enable Enhancements and change to a different model and then go back and re-enable the Hyperion model and that should allow both to be turned on. I will make a note to the devs that the app defaulting to those two options is causing a headache to see if we can get that addressed in a future update.
