Topaz Video AI Beta 6.2.0.0.b

Hi everyone,

Today we’re releasing an updated beta for Video AI 6.2

6.2.0.0.b

This release includes a new sign-in experience with Google and Apple OAuth support, along with some backend preparation for the first local-rendering model in the Starlight series. We expect to release this model update to all users later this Spring!

More news from us very soon,

– Topaz Video AI team

15 Likes

The previews in 6.x continue to make no sense. If I select the blue render Xs button, there’s to be no way to turn looping on and off and no way to create multiple previews. And “real time preview” appears to be on and operating simlutaneous with the fixed time preview, with no way to turn it off.

Also, comparing my test source to previous builds, what upsamples at 20fps in 5.3.6 is upsampling at 14fps in this build with the same settings. That’s a 30% performance loss.

Basically, the only reason to have this on my system is to serve as a Starlight gateway.

2 Likes

You don’t even need it for that, they have a web IU for Starlight as well :grin:

I’ve got something weird happening here with the web front end. Unless I open it in my browser’s “incognito mode,” it just keeps reloading itself and won’t accept any input.

Nice to hear, really that is we are waiting for!! But there are unfortunately mayour bugs with TVAI and I don’t understand why these have not been fixed, there was enough time. I overlook now the new UI design, which nobody likes. Not a solid base running a “local Starlight” on it, and does TVAI 6 really has become slower than V5, why?

2 Likes

RTX5000 opt work

Speeds are back to normal on my 3080 ti, except Nyx, it’s about half as slow as it used to be in the benchmark.
No errors running the benchmark this time.

1 Like

Any idea on how much vram would be required for local Starlight upscaling ?

A great question – we are working on optimization day-by-day and we hope to detail the system requirements soon. For now, it’s safe to assume this model will require a more recent and more high-end GPU than any of our previous models.

6 Likes

Any idea if the local Starlight model will run on Apple Max or Ultra? I don’t mind having to wait two weeks for a two minute video render since that’s the max amount of time I usually do for my slowmo montages.

1 Like

Does “more high end” mean higher than an RTX 4090/5090? Will we need to start shopping AI accelerator cards?

I think even with a 5090, processing more than a few minutes of video will be too time consuming with their local Starlight version, judging from the performance from the cloud version with much more powerful hardware. Even a x10 improvement in performance would be too slow for a 1 hour video is my guess. But ofc, I might be wrong :wink:

2 Likes

For now, I see Starlight as a tool for relatively short clips. I just did a project in which two or three scenes had very low quality compared to the rest but had actions that were too important to just delete. That’s the sort of thing that Starlight is currently useful for.

2 Likes

Since I do two - three minute videos, I don’t mind waiting two weeks or more for it to process.

1 Like

Just popped in my 5090FE and gave it a shot. APFast still definitely needs some work, but a nice improvement over my 4090 otherwise.

Topaz Video AI Beta  v6.2.0.0.b
System Information
OS: Windows v11.24
CPU: Intel(R) Core(TM) Ultra 9 285K  47.469 GB
GPU: NVIDIA GeForce RTX 5090  31.349 GB
GPU: Intel(R) Graphics  0.125 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	53.18 fps 	2X: 	17.46 fps 	4X: 	04.42 fps 	
Iris		1X: 	47.78 fps 	2X: 	18.92 fps 	4X: 	04.78 fps 	
Proteus		1X: 	60.03 fps 	2X: 	18.77 fps 	4X: 	04.81 fps 	
Gaia		1X: 	19.10 fps 	2X: 	13.76 fps 	4X: 	04.60 fps 	
Nyx		1X: 	17.51 fps 	2X: 	12.65 fps 	
Nyx Fast		1X: 	42.21 fps 	
Rhea		4X: 	04.63 fps 	
RXL		4X: 	04.57 fps 	
Hyperion HDR		1X: 	24.30 fps 	
4X Slowmo		Apollo: 	40.70 fps 	APFast: 	67.38 fps 	Chronos: 	46.82 fps 	CHFast: 	45.85 fps 	
16X Slowmo		Aion: 	31.16 fps 	

AsRock Z890I Nova WiFi Mainboard (BIOS v2.23.AS05, 02/18/2025, Intel uCode 0x116, Intel MEI v19.0.5.1948)
Intel Core Ultra 9 (Series 2) 285K
24GB x2 (48GB) DDR5-9300 42/56/56/134/968/2T
nVidia 5090 Founders Edition
nVidia Game Ready Driver v572.83 (03/18/2025)

Very nice. 50-series optimization in this release perform even better than in the ALPHA-release.

In general, seems to be about a 10% improvement compared to ALPHA, with a few exceptions that probably are more dependent on other parts than tensor cores.

AION stands out, as it’s seeing aa almost 20% improvement.

Topaz Video AI Beta  v6.2.0.0.b
System Information
OS: Windows v11.24
CPU: AMD Ryzen 9 9900X 12-Core Processor              31.555 GB
GPU: NVIDIA GeForce RTX 5080  15.517 GB
Processing Settings
device: -2 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	39.46 fps 	2X: 	20.03 fps 	4X: 	05.57 fps 	
Iris		1X: 	38.12 fps 	2X: 	21.56 fps 	4X: 	06.43 fps 	
Proteus		1X: 	39.25 fps 	2X: 	24.25 fps 	4X: 	06.28 fps 	
Gaia		1X: 	11.82 fps 	2X: 	08.47 fps 	4X: 	06.11 fps 	
Nyx		1X: 	10.81 fps 	2X: 	08.51 fps 	
Nyx Fast		1X: 	26.33 fps 	
Rhea		4X: 	04.73 fps 	
RXL		4X: 	04.57 fps 	
Hyperion HDR		1X: 	33.65 fps 	
4X Slowmo		Apollo: 	53.76 fps 	APFast: 	76.39 fps 	Chronos: 	30.76 fps 	CHFast: 	42.95 fps 	
16X Slowmo		Aion: 	55.91 fps 	

MSI MAG X870E Tomahawk
CPU +200mhz boost clock, CO value per core, x10 scalar, curve shaper manually tuned, mb ppt limits (~250W).
Memory:
ZenTimings_Screenshot7

The bug with only ‘generic’ models being downloaded and used for RTX 3090 on Win10 is solved in this beta. :slight_smile:
obraz

BUT. Temp wipe bug is STILL present. It’s dangerous. :stuck_out_tongue:

Good News, some models for the 5080 improves, the benchmark overall shows improvement, but real performance do not for Artemis HQ, please get it optimized, it is 99.9% of my upscaling.

Dione Robust with 2nd pass Scale 1x Iris LQ 720 to 4k - 22% Improvement, GPU ~95 - 100%

Artemis HQ 1080 to 4k - No Improvement, GPU underutilized 50-60%

Topaz Video AI Beta  v6.2.0.0.b
System Information
OS: Windows v11.24
CPU: AMD Ryzen 9 9950X 16-Core Processor              63.555 GB
GPU: NVIDIA GeForce RTX 5080  15.517 GB
Processing Settings
device: 0 vram: 1 instances: 1
Input Resolution: 1920x1080
Benchmark Results
Artemis		1X: 	35.80 fps 	2X: 	19.42 fps 	4X: 	04.53 fps 	
Iris		1X: 	33.41 fps 	2X: 	17.97 fps 	4X: 	05.45 fps 	
Proteus		1X: 	35.55 fps 	2X: 	21.35 fps 	4X: 	05.62 fps 	
Gaia		1X: 	10.70 fps 	2X: 	07.78 fps 	4X: 	05.16 fps 	
Nyx		1X: 	09.73 fps 	2X: 	07.96 fps 	
Nyx Fast		1X: 	23.55 fps 	
Rhea		4X: 	04.21 fps 	
RXL		4X: 	04.11 fps 	
Hyperion HDR		1X: 	31.86 fps 	
4X Slowmo		Apollo: 	46.47 fps 	APFast: 	71.05 fps 	Chronos: 	27.94 fps 	CHFast: 	40.13 fps 	
16X Slowmo		Aion: 	52.64 fps 	

It’s primarily bottlenecked by RAM speed due to the CPU/RAM used to scale the image before sending it to the GPU.
When optimizing RAM, we always see uplift in 2X and 4X results.

1 Like

Ran ran a batch from the same TV series, 6 errors out of 30 episodes.

logsForSupport.zip (4.6 MB)
DxDiag.txt (111.5 KB)