PRAP is the auto parameter estimation model not an upscaler, will have to add field and disable those. Are you using the jsons sent with the build or the ones in your model directory?
Can we get start and finish marks in the timeline for different models?, or even diferent parameters for the same model?, for instance, i want from minute 1 to minute 3 proteus 3 with a specific set of parameters, but to minute 4 and minute 5 gaia, or for minte 6 to minute 9 proteus with other parameters, also it would be nice to make proteus get the values for certain times in the video instead of just one frame, for example to get all the values and get an average from a minute sample instead of the single frame it provides, it would also be great to specify the sample from the timeline.
1 Like
Also scene and frame detection would be great, sometimes my videos have complete black frames and veai wastes time and resources applying the models to them when itās just pure black
also vram usage as cache for frames or something would be great, i only upscale to 1080 some of my videos i use proteus mainly and i have 12gb of vram, proteus only uses 1.5gb roughly, and what about amd rdna2 support?
Tensor Cores have been supported by VEAI for over a year now. Since the launch of RTX 30 series GPUs.
1 Like
Post or vote for feature requests here:
RDNA2 GPUs are already supported by VEAI.
Yes the current beta/release version VEAI only use ffmpeg to merge audio back. So it uses the publicly available ffmpeg x86_64 builds.
The new version will be custom built for both M1 and x86_64 with the VEAI filters.
1 Like
Great news, thank you :).
I have only one request, but it is a big one:
The ability for end-users to perform their own training on existing models to refine them, and to share the resulting models on a model sharing platform provided by Topaz that allows other users to rate and comment on the community-made models. All community-made models would be freely available to any licensed VEAI user.
The user has to provide their own dataset of paired videos, where each pair of videos consists of one video that is degraded (LR) and another which is the ground truth (HR). The user picks which existing VEAI model to use as a starting point (pretrain), and waits patiently as the model trains for at least a week. While training, the software periodically outputs a short test video to validate that the model is still improving (checkpoint).
i didnāt specified, iāve been using one xd, what i meant is full utilization of an rdna2 gpu since i only get 50% usage overall, since there are spikes of 99% usages then 0 utilization
yes with the multiprocess arch move, it should be close to 100%. Try the alpha and let me know if it doesnāt reach 100% usage with multiple instances of the ffmpeg alpha running.
thereās a bug with running multiple instsances of veai in rdna gpuās, it crashes, iām not sure if itās a driver issue or a veai issue, i will try to replicate the problem and provide info
Will try on the W6800 we have here. Do you have enough VRAM on your RDNA GPU?
This is quite exciting. I have a few video snippets that fails with current version that I had already submitted last year (interlaced, stutter), Please let me know if you want me to post them again. Will be happy to retest those as well.
Also thinking of going with Mac M1 Pro or Max, would M1 Max make a difference vs. M1 Pro with 3.0?
thanks!
Iām looking forward to testing the first v3 Mac GUI version as I have a M1 Max for (workā¦ahem) testing purposes!
I did run beta and alpha together with Gaia (W6800, ECC on), there was no drop in performance with the alpha but with the beta.
I had no crash.
I did set the fan target to 2800 rpm since the hotspot on the card can reach 110Ā°C.
I did also write AMD because of its hardware sheduler.
When running VEAI (Stable) and three other apps (each close to 100% GPU load) at the same time one of them did produce an error.
VEAI 2.6.3 that will release soon, speeds up M1 Pro/Max performance by a lot compared to 2.6.2, unfortunately we only have M1 Max machines. So I would wait for people to report M1 Pro results before deciding between the two.
Good to know. May be the 6800 or 6900XT behave differently. The Pro cards are supposed to be more power efficient, so reaching 110 is still very hot.
I need to inquire about which component it is.
Maybe voltage converter or the memory.
The GPU itself is at 75 - 80Ā° C.
The 2800 rpm target prevents the GPU from throttling, it is audible but I donāt mind.
The hotspot moves abruptly between 70 and 90 Ā°C with peaks up to 107 Ā°C, with lower fan rpm the temperature oscillates further up.
update:
According to igors Lab, it is the memory on the back.
Lenovo workstations have fans for both the GPU and CPU memory.
I will probably install a fan for the back of the GPU.
In summer, the temperatures can be another 5Ā°C higher.
My Radeon VII would consistently hit hotspot temperatures of 105Ā°C on stock cooling and it would BSOD often. Since watercooling it I havenāt seen it go above 85 degrees and no more crashes. Tjunction now stays around 65 degrees full load.
