Option to NOT execute Nyx upscale after Starlight

Topaz Video Ideas
1

I am upscaling 480p originals. When using Starlight Sharp, selecting any resolution for the output, it appears to first do a 1080p upscale and then uses Nyx to upscale to the selected output resolution.

What I would like to do is NOT execute the Nyx process and stop at the 1080 output. I do my own post processing to get to 2160 in Proteus. As it is I have to watch the end of a multi-day process and catch it before it finishes the Nyx run to save the 1080p intermediate file.

A simple check box to ā€˜Save intermediate fileā€™ would be great.

2

So, is there really that much difference between the primary SL Sharp output without NYX and SL mini?

3

IMHO, yes. Starlight Sharp is sharper than Starlight ā€˜Miniā€™ without any post processing. I dont know what parameters they are passing to the Nyx model, but I suspect it was chosen as the model that does the least modification of the content.

Also, doing the ā€˜oddā€™ upscale (225% to go from 480 to 1080) is better done by the diffusion model and have it hand a simple 200% upscale job to Proteus. I do a lot of sharpening in my Proteus run, and a LOT of noise reduction. The NR cleans up any edges that may get over-sharpened.

1 Like
4

AFAIK Topaz stated that Starlight Sharp is a Starlight with Nyx on top of that.

Maybe you should try Starlight and Proteus/Rhea afterwards?

5

Would someone be kind enough to copy and paste an example CLI command for a Starlight Sharp export please? (I canā€™t do SLS directly on my machines)

I / we will then have a better idea of how this works.

Thanks.

Andy

1 Like
6

All I can say is that the diffusion process runs faster when running Starlight vs. Starlight Sharp, which means SOMETHING is different between the two.

Edited to add: Furthermore, Starlight DOES NOT produce an intermediate file, instead directly scaling to the resolution selected from the menu, and displays the upscaling process on the Export Queue tab.

7

Good point, Iā€™m interested too, I tried but something must be wrong, it says ā€œcannot verify filesā€

"C:\\ProgramData\\Topaz Labs LLC\\Topaz Video\\models\\/runner.exe"  ,agrs:  QList("C:/TEMP2/testvideo.mp4", "C:/TEMP2/testvideo.mkv", "--start-frame-idx", "0", "--end-frame-idx", "512", "--ffmpeg-encoding", "-level 3 -c:v ffv1 -pix_fmt yuv420p -slices 4 -slicecrc 1 -g 1 -movflags frag_keyframe+empty_moov+delay_moov+use_metadata_tags+write_colr -bf 0", "--ffmpeg-preproc-filters", "", "--upscale-factor", "2", "--max-gpu-mem", "32", "--model_name", "astrasharp")  working dir:  "C:\\ProgramData\\Topaz Labs LLC\\Topaz Video\\models\\"
8

The option to disable nyx into SLS is a good idea. I would even go further: They could give us ā€œcustom resolutionā€ today for Starlight + Starlight Sharp, but instead we get predefined fixed scaling factors with it often you do not reach desired target resolution or overshoot it. Here it becomes absurd: SLS is an FHD model and does always FHD (final scale Nyx), but you canā€™t select FHD as output :zany_face:

They could even let the user freely choose which model to use for the final scale: Lanczos, Nyx, Proteus, Iris, Gaia, or Theia. Topaz is not taking advantage of all the possibilities they have.

1 Like
9

Having queued up several Starlight Sharp jobs over the last few days, the 1440p output after the SLS/Nyx process is of inferior quality to what I routinely get when I can use the 1080p intermediate file. As I said, I donā€™t know what parameters Nyx is getting, but faces look horribly plastic.

I donā€™t know who said Starlight Sharp is just Starlight mini with Nyx added, but if that were so, why are there two different model files (astra-1.0 and astrasharp-1.0), why do they produce diiferent resolutions, and display differently in the export Queue (no preview with SLS)?

10

Yes SLS is a different model then they run NYX and additional something else over the SLS encode.

Unfortunately, they donā€™t give us the parameters to turn off nyx and to vary the strength. Itā€™s also not clear why you donā€™t have more target resolution options in the Topaz program, the SLM diffusion model can do different scalings via the CLI, so why do they only specify us with fixed factors 2x, 3x 4x

It would be better if SLS is also trained to 2k and 4k, only FHD is simply not enough. if I use nyx, then I would like to do this 1x