v8 models have been restored.
ThanksâŠitâs so bizarre to me that itâs one of us that keeps having to tell them that their models are missing. How hard can it be to monitor that?
I have the box left to download faster models. But it hasnât done anything. I also have no idea where to browse for more Models.
I sat down and did systematic testing. VEAI is broken in that most updates hurt performance by large margins.
Ryzen System:
Ryzen 3700x, 32GB DDR 3600, EVGA 2060KO
Intel System:
i7-8700k, 48GB DDR 3000, EVGA 1080ti
DVD Clip is from 720x480 â 225% 1440x1080
HD Clip is from 1920x1080 â 200% 3840x2160
Both Windows 10 v 20H@
Nvidia Studio v 457.30
Topaz VEAI v1.7.1
Clip---------- Model---------- 2060KO---------- 1080ti
DVD---------- ALQ-8---------- 1.07spf---------- 0.16spf
DVD---------- AHQ-8---------- 1.09spf---------- 0.16spf
DVD---------- GHQ-5---------- 0.18spf ----------0.33spf
DVD---------- GCG-5 ---------- 0.18spf---------- 0.33spf
HD------------ ALQ-8---------- 7.63spf---------- 9.91spf
HD------------ AHQ-8---------- 7.53spf----------10.04spf
HD------------ GHQ-5---------- 1.25spf---------- 2.18spf
HD------------ GCG-5---------- 1.26spf---------- 2.20spf
Topaz VEAI v1.3.8
Clip---------- Model-------------- 2060KO---------- 1080ti
DVD------ Artemis-LQ---------- 0.35spf---------- 0.21spf
DVD----- Artemis-HQ---------- 0.35spf---------- 0.21spf
DVD---------- Gaia-HQ---------- 0.58spf---------- 0.26spf
DVD---------- Gaia-CG---------- 0.55spf---------- 0.25spf
HD-------- Artemis-LQ---------- 2.16spf---------- 1.28spf
HD-------- Artemis-HQ----------2.16spf---------- 1.29spf
HD------------ Gaia-HQ---------- 3.21spf---------- 1.87spf
HD------------ Gaia-CG---------- 3.02spf---------- 1.74spf
Yes, Iâve experienced this as well but specifically only if itâs a video that has already been pre-processed by VEAI. I do a denoise/deblock pass first, and using that resulting video to upscale further gives me the frame duplication error. The frame issues donât happen if I straight up upscale from source. Source is already reencoded in both cases.
You donât browse them. It downloads whatâs appropriate for what youâre doing.
abbadon posted:
With a Nvidia RTX 3090 FE:
720p to 1080p upscale, 0.19sec per frame
I get about 1sec per frame with a Nvidia GTX 1080 Ti doing the same.
@tpirowski:
I presume you are using an AMD RX 6800 XT and comparing it to a Nvidia RTX 3080?
If youâre getting 10x faster, that would be about 0.023sec per frame, or ~43 frames per second. Please confirm. My next box I plan to use two RX 6800 XT GPUs.
When upscaling DSLR 1080p footage to UHD it seems that the highlights are more blown out (the scopes also change compared to original). Tried to balance it with adjustment layer and bring highlights down. Also tried saturation but it doesnât fix the issue. Is there a way to avoid any exposure and color shifts when upscaling? Or a way to fix it in post quickly with some settings on an adjustment layer?
My workflow
-
work on 1080p timeline when project finished export in prores
-
import in Video Enhance
-
UHD upscale / 200% for 1080p TIFF 16 bit output (which setting should I use for this?)
-
Render and bring back on 4k timeline in PP and export again with original audio from 1080p timeline.
Any tips for best settings + quality and no exposure and color shifts?
For those that use Artemis HQ v6 models, the 100% denoise/deblock models for fp32/ox are still offline:
http://veai-models.topazlabs.com/ahq-v6-fnet-fp32-256x352-1x-ox.tz
http://veai-models.topazlabs.com/ahq-v6-fnet-fp32-288x288-1x-ox.tz
http://veai-models.topazlabs.com/ahq-v6-fnet-fp32-384x480-1x-ox.tz
http://veai-models.topazlabs.com/ahq-v6-fnet-fp32-384x672-1x-ox.tz
http://veai-models.topazlabs.com/ahq-v6-fnet-fp32-448x576-1x-ox.tz
http://veai-models.topazlabs.com/ahq-v6-fnet-fp32-480x384-1x-ox.tz
http://veai-models.topazlabs.com/ahq-v6-fnet-fp32-576x384-1x-ox.tz
http://veai-models.topazlabs.com/ahq-v6-fnet-fp32-576x416-1x-ox.tz
http://veai-models.topazlabs.com/ahq-v6-fnet-fp32-576x672-1x-ox.tz
And also these 2 for Apple/ml:
http://veai-models.topazlabs.com/alq-v6-gnet-fp16-576x416-1x-ml.tz
http://veai-models.topazlabs.com/alq-v6-gnet-fp16-576x384-4x-ml.tz
Since, VEAI is not using Tensor Core at the moment, therefore the processing speed of fp16 AI Model is depending on FP16 performance.
From AMD Specifications, you can find the FP16 performance of BIG NAVI GPU.
https://www.amd.com/en/products/specifications/compare/graphics/10516%2C10521%2C10526
Radeon RX 6900 XT: 46.08 TFLOPS (2:1)
Radeon RX 6800 XT: 41.47 TFLOPS (2:1)
GeForce RTX 3090: 35.58 TFLOPS (1:1)
Radeon RX 6800: 32.33 TFLOPS (2:1)
GeForce RTX 3080: 29.77 TFLOPS (1:1)
According to above data, I wonât surprise if RX6800XT can process faster than RTX3090, hopefully some user are able to buy the card and confirm the result.
is there a way to force download all models so they are already stored locally ?
Like have a download page for the model for offline use ?
Instructions are here. Keep in mind full downloads of everything is like 60 gb. âmlâ is Apple Core ML, so if youâre on Windows those are useless to you.
Make sure SLI is off.
Colour shifts are unfortunately something that hasnât been addressed yet. Skin tones change drastically, and the audio gets reencoded. Audio should be passthrough by default, and really, colours shouldnât be touched. Unfortunately, thatâs how it is so weâll just have to wait until they decide to have these options.
Video Enhance AI shifts color pallete from Rec.601 â Rec.709 when it is run with VIDEO (MP4) input to PNG output. It also lowers the gamma by factor of .02, i.e. 1.00->0.98.
You can shift it back by specifying the following in Hybrid when importing the png stream:
Selecting âbt470bgâ for color matrix inserts ConvertToYV12(matrix="Rec601")
into the Avisynth script, which corrects the colors.
To correct the gamma, select the Levels filter and change the gamma to 1.02.
Not sure if/when they are ever going to fix this. Developers are aware of the issue.
Thanks for the info. We hesitate to change the userâs computer setting. But may we should give users the option to do that.
Thanks for report it. Will look into it.
Tried running 1.7.1 on an older i5-3470 paired with a 1050ti and it just shows a black screen in either the preview or actual AI processing. Is it because the processor is too old? I thought the processing is GPU based? Works fine though on my other two modern Ryzen rigs with 1660Tis though.
yes, blocking the computer to go into sleep while doing a processing will be a good idea ! iâm pretty sure that everybody would agree !
My findings re: Artemis v7, v8
Artemis LQ v7:
o slight, almost imperceptibly better in v7 than v8 in detail, especially noticeable in human faces
o v7 less flickery than v8
o bad âhorizontal lineâ problem in v7 that can occur in certain places, fixed in v8
Artemis HQ v7 & v8: video can âbreak upâ in spots as if it is too sharp