Spoiler: MediaInfo Code: General Format : Matroska Format version : Version 4 File size : 507 MiB Duration : 3 min 6 s Overall bit rate : 22.8 Mb/s Frame rate : 29.970 FPS Encoded date : 2026-10-28 13:28:01 UTC Writing application : mkvmerge 102.0 ('Little Houses') 64-bit Writing library : libebml v1.4.7 + libmatroska v1.7.2 Video ID : 1 Format : HEVC Format/Info : High Efficiency Video Coding Format profile : Main 10@L4@Main Codec ID : V_MPEGH/ISO/HEVC Duration : 3 min 6 s Bit rate : 21.3 Mb/s Width : 1 920 pixels Height : 1 080 pixels Display aspect ratio : 16:9 Frame rate mode : Constant Frame rate : 29.970 (30000/1001) FPS Color space : YUV Chroma subsampling : 4:2:0 (Type 1) Bit depth : 10 bits Bits/(Pixel*Frame) : 0.342 Stream size : 473 MiB (93%) Language : English Default : Yes Forced : Yes Color range : Limited Color primaries : BT.709 Transfer characteristics : BT.709 Matrix coefficients : BT.709 Audio ID : 2 Format : PCM Format settings : Little / Signed Codec ID : A_PCM/INT/LIT Duration : 3 min 6 s Bit rate mode : Constant Bit rate : 1 536 kb/s Channel(s) : 2 channels Sampling rate : 48.0 kHz Frame rate : 25.003 FPS (1920 SPF) Bit depth : 16 bits Stream size : 34.1 MiB (7%) Language : English Default : Yes Forced : No Spoiler: Download Hidden Content: You must either reply or click 'Like' to see the hidden information contained here.
@JL92 are you doing this upscales as a first step for the future AI upscale ? that's a very smart move, i see you do the color transfer too... Very pro of you, great job and thank you for sharing!
For a perfect 4K Starlight Precise results, the best first step is a non-ai upscale from a master source. I am in an organic era, I'll stick with the non-ai version lol
o Yeah , I agree. At the current stage of AI it hallucinates each frame, I never used AI upscale, I don't have the pc for it but from what in can see the AI doesn't have temporal awareness, it only hallucinates from each frame it sees so if it sees a small face far away it creates a different face, if it had temporal awareness it draw the same face from past or future frames. It will get there but is not there yet, so mathematical upscale is still best. And i asked an AI just that and look what it said :
I like both, mathematical (for an organic result) and a generative AI (only if it's Starlight Precise for 4K) For example, I could never reach this 8K version without a good non-ai 1080p source https://sharemania.us/threads/beyoncé-lady-gaga-video-phone-starlight-precise-4k-8k.307045/
ok but if you have the hardware can't you just use topaz with no hallucinations, just keep blurred the part it doesn't have enough details to upscale to a clear image ? Those freaky face morphs are nightmare fuel !
Topa Topaz has a lot of models, but since Starlight was created all other models looks trash for me. Starlight Precise was created to use 1080p source videos, for the "perfect" 4K without hallucinations I still love it, but I think it needs some updates, like a professional tool to avoid some parts to keep it more natural, or even scan faces first to be sure that the results will be more real My biggest inspiration for non-ai organic upscales was the channel MHD (2006-2008) and Palladia (2008-2016). I really love their results Hidden Content: You must either reply or click 'Like' to see the hidden information contained here.
Tv stations such as MHD definitely have better quality sources anyway, but for fans like us doing the mathematical upscales and keeping them is very smart because they can be used when AI upscales advance to a better level where you can input a still photo of a person and the AI uses it as reference, i think I've seen a video on YT saying Hollywood is working on this, or somethin, still you're doing non-Ai upscales is a great move!
The most impressive fact is that MHD did a real-time upscaling... in 2006! With Teranex computers and original master tapes
Hoh!, yeah, they have the hardware too, years ago i was looking at madVR Envy Core (the hardware box) made by the guys that made made the madVR software, so yeah they always had the hardware and cash to do good upscales but all the slop posted on YT with shitty official upscales made by official labels is just lazy or from noobs making them. Sharemania has much better upscales than them thanks to guys such as you!
Those labels used the same preset, and add their entire catalog to process. Poor results Too much contrast, halo, overshapened. Everything is too much basically
Yeah, it makes no sense to me how they can post such garbage to the public, are they just stupid, or what, why they don't do a better job, i never understood.
with a master prores source, in less than 1 hour I can produce a good non-ai.. those labels are so boring
this is even worse than we though... bad source files with softwares that is not specific for upscaling
