Blackmagic Forum

I installed V20 Public Beta on my desktop workstation (specs on my signature) and the first thing I noticed is the Neural Engine Optimization process was very slow and crashd Resolve when the process reached about 95%. Then I opened Resolve again and it continued the Neural Engine Optimization process for few seconds then crashed Resolve again.

I opened Resolve for the 3rd time, and this time I skipped the Neural Engine Optimization process, so I can try Resolve without it. The first thing I tried was Magic Mask 2. The masking is great but the tracking was very slow comparing to Magic Mask in Resolve Studio 19 where it reached 34 FPS. In V20 the tracking speed reached a maximum of only 6 FPS.

For comparison, it is much faster in my laptop where it reachs 45 FPS.
Here is the spec of my laptop:
ProArt Studiobook 16 OLED (H7604)
- 16-inch 3.2K 120 Hz
- Windows 11 Pro
- Intel Core i9-13980HX 24 cores
- NVIDIA GeForce RTX 4070 Mobile 8GB VRAM
- 64GB RAM
- 8TB NVMe Storage
- Resolve Studio 19.1.3

For the record, before I Install Resolve 20 Beta, Windows did install alot of updates, and I also updated the nVidia driver. So some variables happened before installing Resolve 20!

Update: I rebooted my computer, started Resolve v20 for the 4th time, I let it continue the Neural Engine Optimization process whish successfully finished in around 2 minutes. I tried Magic Mask 2 again but didn't see any improvements with the tracking speed, 6 FPS like if Neural Engine Optimization is disabled!

Without the optimization of week run slower.

Check in the settings, there is a optimization checkbox you cab tick to run it again at the next restart

waltervolpatto wrote:Without the optimization of week run slower.

Check in the settings, there is a optimization checkbox you cab tick to run it again at the next restart

I just updated my thread.. After rebooting my computer, Neural Engine Optimization finshed without a crash. However tracking speed remain the same at 6 fps!

I am not sure if this is a bug related to my specific GPU driver or something else!

If you change back to legacy magic mask in R20 is the performance the same as R19?
RTX A2000 ain't a particularly fast GPU and the 6GB could also be limiting if that's the version you have. An RTX4070 is much faster.

shebbe wrote:If you change back to legacy magic mask in R20 is the performance the same as R19?
RTX A2000 ain't a particularly fast GPU and the 6GB could also be limiting if that's the version you have. An RTX4070 is much faster.

I tried the legacy Magic Mask and it is much faster at 20 FPS vs 6 FPS with Magic Mask 2.

However, with Resolve 19, Magic Mask reachs 34 FPS on the same machine.

For the record, my desktop GPU is the 12 GB version.
Is it possible that the "mobile" version of RTX4070 8GB is faster than a desktop RTX A2000 12GB?

I am doing a couple of performance tests with magic mask in v20.

First, in v19 i tracked my baseball cap in a clip and it tracked at about 8fps with a total of 13 minutes to track the whole clip.

Then I upgraded to v20 and upgraded the project and repeated the tracking and got 3fps for a total of about 30 minutes.

This is on my M1 Max with UHD material on an UHD timeline.

I'll repeat this shortly with the legacy approach in v20 to see if it gets back to the v19 performance.

I should note that I have 'Faster' set in both v19 and v20 and the actual mask in v20 is much better with this new Magic Mask.

Stay tuned...

Steve Alexander wrote:I am doing a couple of performance tests with magic mask in v20.

First, in v19 i tracked my baseball cap in a clip and it tracked at about 8fps with a total of 13 minutes to track the whole clip.

Then I upgraded to v20 and upgraded the project and repeated the tracking and got 3fps for a total of about 30 minutes.

This is on my M1 Max with UHD material on an UHD timeline.

I'll repeat this shortly with the legacy approach in v20 to see if it gets back to the v19 performance.

I should note that I have 'Faster' set in both v19 and v20 and the actual mask in v20 is much better with this new Magic Mask.

Stay tuned...

So this confirm that tracking in Magic Mask 2 is indeed slower than legacy Magic Mask.

I hope that BMD will improve Magic Mask 2 and make it at least as fast as legacy Magic Mask.

That would be nice but then again, it's doing more so I'm not surprised that it's slower.

Steve Alexander wrote:That would be nice but then again, it's doing more so I'm not surprised that it's slower.

Yes I understand that it's doing more and the masking quality of Magic Mask 2 is better, but still it does not justify the big difrence of tracking speed! We are talking about 45 fps vs 6 fps .. that is more than 7 times in diffrence!

Yasser Saeed wrote:

Steve Alexander wrote:That would be nice but then again, it's doing more so I'm not surprised that it's slower.

Yes I understand that it's doing more and the masking quality of Magic Mask 2 is better, but still it does not justify the big difrence of tracking speed! We are talking about 45 fps vs 6 fps .. that is more than 7 times in diffrence!

This is a BETA, so it's possible that there are speed issues ranging from bugs to debugging code, or not-yet-optimized code. It's also possible that certain PC/Nvidia combinations are having issues.

Do you have an actual basis for your "does not justify" statement? Like experience making or using large neural networks? Larger ones can indeed be an order of magnitude slower.

wfolta wrote:

Yasser Saeed wrote:

Steve Alexander wrote:That would be nice but then again, it's doing more so I'm not surprised that it's slower.

Yes I understand that it's doing more and the masking quality of Magic Mask 2 is better, but still it does not justify the big difrence of tracking speed! We are talking about 45 fps vs 6 fps .. that is more than 7 times in diffrence!

This is a BETA, so it's possible that there are speed issues ranging from bugs to debugging code, or not-yet-optimized code. It's also possible that certain PC/Nvidia combinations are having issues.

Do you have an actual basis for your "does not justify" statement? Like experience making or using large neural networks? Larger ones can indeed be an order of magnitude slower.

Yes I know this is still a beta and such probelms are expected, and I hope BMD will fix it in the final release of Resolve Studio 20.

My reply was directed to Steve when he said that its (Magic Mask 2) is doing more, thus doing more "does not justify" to be 7.5 times slower than the previous genration

For me Magic Mask 2 Performance is also very slow.

About 2 FPS on 4K material on a MacBook Pro M3 Max with 64G RAM.

Just be sure neural engine optimisation is finished guys. It does crushed 5-6 times while doing it but also continues when i start-up again with wherever it stuck. Just not disable it

I'm pretty fine with magic mask v2 speed with rtx5090 GPU btw.
Here is a small test with 4K timeline

Magic Mask V2 - 30 fps tracking speed ( better option)
35 fps on (faster option)

Older model is also extemely fast 53fps on Legacy-Faster
49fps on Legacy-Better ( I just tested that Resolve 19 without 5xxx series non-supported and it does produce 7-8fps with that )

I’m having an experience on my MacBook Pro M3 max with 128 gigs of round. Down from about 15 frames a second on version 19 to 4 fps.

That said, the mask is excellent and requires a very little tweaking if anything. Because I don’t have to tweet, it is actually faster for my workflow.

RyanFlynn wrote:I’m having an experience on my MacBook Pro M3 max with 128 gigs of round. Down from about 15 frames a second on version 19 to 4 fps.

That said, the mask is excellent and requires a very little tweaking if anything. Because I don’t have to tweet, it is actually faster for my workflow.

Yes Magic Mask 2 is better and I don't mind if its normal to be slower, which I doubt it. However, Legacy Magic Mask is also very slow comparing to Magic Mask in Resolve 19 and that is the problem.

Hi, I'm down to 2fps on my MacBook M3 Pro with 18GB memory, 11 core CPU, 14 core GPU.
is there anything I can do to speed this up?

Thanks in advance!

tunnytunes123 wrote:Hi, I'm down to 2fps on my MacBook M3 Pro with 18GB memory, 11 core CPU, 14 core GPU.
is there anything I can do to speed this up?
Thanks in advance!

I have M2 Ultra (64gb RAM, 60-core GPU)... and I have likу 6 fps on officially released DVR20. So... you're not alone

Yasser Saeed wrote:

Steve Alexander wrote:That would be nice but then again, it's doing more so I'm not surprised that it's slower.

Yes I understand that it's doing more and the masking quality of Magic Mask 2 is better, but still it does not justify the big difrence of tracking speed! We are talking about 45 fps vs 6 fps .. that is more than 7 times in diffrence!

Maybe that's about right?

According to the AI, your GPU is only 70watts, the 20GB Vram is great but 70watts isn't, It's most similar to a RTX 4070 as far as tensor and cuda cores but the 4070 is allowed to draw 200 watts and has higher memory bandwidth. 6fps might be about right if we were to assume the technology didn't get better only the precision got better so the tracking speed got worse. Compare to my 4090 at 4K , peaks about 20fps (better) and a 5090 user above said 30fps(better) and that's a 575W GPU.

From benchmarks that heavily utilise tensor cores(like MM2) it looks like your GPU is often 30-40% slower than an RTX 4070

You could always go pack to manually rotoing the object by hand and get 0.00001 fps and it would be a bubbling mess for the first 3 iterations.

Most people are not using the mask for an entire video and just a shot so 3x slower is not bad especially if the superior result means less tweaking, The reality is 40fps plus tweaks = 5-20fps in reality, If we do less or no tweaks we come out with similar speed from v1 to V2. I am speculating and assuming it is more accurate but I have not had a chance to use it yet.

I upgraded from the free program to the Studio version. Like others on this post I am getting at best 2 fps during tracking. I am running on an M1 pro Macbook. We make a lot of green screen videos in the two-minutes range and was hoping to replace the green screen with the. magic mask, but that is not really appear to be feasible.

On a small test clip I could not achieve the quality of a good old-fashioned green screen using Magic Mask 2. Aside from the slow tracking time, would that / should that, be expected?

Yasser Saeed wrote:the Neural Engine Optimization process was very slow and crashd Resolve when the process reached about 95%. Then I opened Resolve again and it continued the Neural Engine Optimization process for few seconds then crashed Resolve again.

How come? I thought your expensive workstation was immune of crashes!

Is Magic Mask 2 only useful for a super computer. Tracking a i minute clip at well over an hour at 1 or 2 frames per second does not seem to fit the narrative that you see in YouTube reviews that claim that the MM2 is amazing and so useful.

mhcstudio wrote:Is Magic Mask 2 only useful for a super computer. Tracking a i minute clip at well over an hour at 1 or 2 frames per second does not seem to fit the narrative that you see in YouTube reviews that claim that the MM2 is amazing and so useful.

That's a reasonable comment, at 4K on 4090 I get about 24fps(better), a 5090 user said 30fps(better), but the thing is I found MM1 (Legacy MM) mostly useless, it gave capcut vibes as far as how bad it was, anything I produce I don't want it to be thought of as coming out of capcut, MM1 is capcut quality auto masking, but MM2 is so much more precise at the cost of compute and being able to paint out the inaccuracies in(hopefully) not that many frames means I"m able to use it for the first time.

Other AI functions have increased in speed by 2x - 4x (on Nvidia anyway) so maybe it will be faster in the future.

mhcstudio wrote:I upgraded from the free program to the Studio version. Like others on this post I am getting at best 2 fps during tracking. I am running on an M1 pro Macbook. We make a lot of green screen videos in the two-minutes range and was hoping to replace the green screen with the. magic mask, but that is not really appear to be feasible.

On a small test clip I could not achieve the quality of a good old-fashioned green screen using Magic Mask 2. Aside from the slow tracking time, would that / should that, be expected?

tunnytunes123 wrote:Hi, I'm down to 2fps on my MacBook M3 Pro with 18GB memory, 11 core CPU, 14 core GPU.
is there anything I can do to speed this up?

Thanks in advance!

Yes, that’s to be expected. MM2 relies on Generative AI, which demands powerful GPUs. A MacBook M* is essentially just a smartphone with a larger screen and a keyboard—its AI capabilities are severely limited (to be polite). I don’t mean to be harsh, but sometimes Apple users need a reality check on what ‘good hardware’ truly means. You can’t keep imagining that a 45W GPU will work magic. There’s a reason people invest in 600W beasts equipped with 21.760 CUDA cores and 680 Tensor Cores dedicated solely to AI computations. And no, those are not 'supercomputers'—they're simply a step beyond the gadgets you’re used to. If you’re serious about getting real work done, it’s time to move on from underpowered Apple devices which are mainly just for showing off in cafés. To be completely honest, I will never understand why people, who are supposedly on a tight budget, still buy this overpriced sh*t from Apple, when you can spend money in much better hardware. It must be that "Pro" label ...

ha! the MBP is an amazingly decent machine until you run into's is limitations, says me writing on my M2 MBP, but owning workstations with real world horsepower

MagicMask2 is real world useful on a workstation, nothing but pain on the M2

like bringing a screwdriver when the job needs powerdrill

battery life on the M2 is pretty astounding tho, there's tradeoff's...

mhcstudio wrote:On a small test clip I could not achieve the quality of a good old-fashioned green screen using Magic Mask 2. Aside from the slow tracking time, would that / should that, be expected?

I'm not confident AI generative tools will ever replace 'old-fashioned' keying. A high end key is deterministic and consistent across all shots from the same scene. It can only produce errors by human error (but that goes for anything even AI tools) and will never be temporally unstable.

The only thing that AI would be able to improve is a sort of assisted keying mode where it can predict settings needed to get the best quality and/or provide a fast core matte. A generative model cannot be as good in detail and transparency because every frame is evaluated and invented on an individual basis rather than using clever math to manipulate existing pixels only.

Maybe in the future it'll be 'usable'. Until then it highly lends itself for mask data for grading or harder mattes for quick roto to put text behind object for example. For those tasks MM2 is great so far.

Admittedly Im running an older system - Mac mini 2019 intel i7 64gb ram with 3 x Blackmagic egpus, NVME drives etc. It generally performs well with my 4k requirements in Raw and PR. I can run full frame full res with most grades except NR. Render times are reasonable.

Tried Magic Mask 2 to isolate a 15 min continuous seated interview clip. Using the better setting it repeatedly wanted three days or more to track the subject of the interview!....I tried I love the quality of the mask but who can use it?

PTuser wrote:Tried Magic Mask 2 to isolate a 15 min continuous seated interview clip. Using the better setting it repeatedly wanted three days or more to track the subject of the interview!....I tried I love the quality of the mask but who can use it?

I think many users expect way too much from a tool that isn't necessarily built for what they use it for. Just because it's "AI" doesn't mean it should blast through thousands of frame and fix whatever it is your trying to fix. I would be curious to know why you want to use the tool to mask 15 minutes of footage to begin with... maybe other simpler solutions can do the job just as fine and be real-time or close to while retaining flexibility of adjustment without recalculations needed. Whatever it is, I would never rely on the use of this tool as part of planning and shooting expecting great results in post.

This tool works well for short spans of maybe couple seconds for shots where you know you won't need a lot of adjustments of the generated mask and need the detailed results over more coarse masking with lots of feather. Even in that state implementation wise it does still has some issues revolving around caching and reliability but can be somewhat mitigated by not having too drastic recalculate runtimes.

shebbe wrote:

PTuser wrote:Tried Magic Mask 2 to isolate a 15 min continuous seated interview clip. Using the better setting it repeatedly wanted three days or more to track the subject of the interview!....I tried I love the quality of the mask but who can use it?

I think many users expect way too much from a tool that isn't necessarily built for what they use it for. Just because it's "AI" doesn't mean it should blast through thousands of frame and fix whatever it is your trying to fix. I would be curious to know why you want to use the tool to mask 15 minutes of footage to begin with... maybe other simpler solutions can do the job just as fine and be real-time or close to while retaining flexibility of adjustment without recalculations needed. Whatever it is, I would never rely on the use of this tool as part of planning and shooting expecting great results in post.

This tool works well for short spans of maybe couple seconds for shots where you know you won't need a lot of adjustments of the generated mask and need the detailed results over more coarse masking with lots of feather. Even in that state implementation wise it does still has some issues revolving around caching and reliability but can be somewhat mitigated by not having too drastic recalculate runtimes.

Masking, as far as I can see, is the accepted way to alter skin tones in post with tracking if the subject moves. Just check out the number of tutorials on YouTube. grading choices in post often necessitate some correction to foreground. I dont understand why you think it's not built for it....

PTuser wrote:Masking, as far as I can see, is the accepted way to alter skin tones in post with tracking if the subject moves. Just check out the number of tutorials on YouTube. grading choices in post often necessitate some correction to foreground. I dont understand why you think it's not built for it....

I did not think it is not built for use in color grading. I said it is not built to crunch through a 15minute clip effortlessly.

Skin adjustments can be done in may ways that are faster and easier to manage. A proper white balance is a start. A proper management of the hues present in the shot is a second. Qualification of the skin hue range with the Qualifier tool is a third. If all are performed in that order and it's still not possible to get a good result the shot must have severe issues if MM2 is deemed to be the only solution. I would never consider it's use as first order of operation.

shebbe wrote:

PTuser wrote:Masking, as far as I can see, is the accepted way to alter skin tones in post with tracking if the subject moves. Just check out the number of tutorials on YouTube. grading choices in post often necessitate some correction to foreground. I dont understand why you think it's not built for it....

I did not think it is not built for use in color grading. I said it is not built to crunch through a 15minute clip effortlessly.

Skin adjustments can be done in may ways that are faster and easier to manage. A proper white balance is a start. A proper management of the hues present in the shot is a second. Qualification of the skin hue range with the Qualifier tool is a third. If all are performed in that order and it's still not possible to get a good result the shot must have severe issues if MM2 is deemed to be the only solution. I would never consider it's use as first order of operation.

Shebanjah, on most machines, the V1 of the magic mask is substantially faster in setting up and executing that v2...

that is one of the main complaints....

The first time you use Magic Mask, it has to 'wake-up' the feature, so the first click takes a bit longer.
Going forward, it's quite fast at making selections.
The tracking speed is about 50% slower, but a reasonable trade-off for the significant increase in quality imo.

And for 15 min clips, I've always found it's a lot easier to work in smaller chunks, like MM on a 2 min clip seems long even.

capthook wrote:The first time you use Magic Mask, it has to 'wake-up' the feature, so the first click takes a bit longer.
Going forward, it's quite fast at making selections.
The tracking speed is about 50% slower, but a reasonable trade-off for the significant increase in quality imo.

And for 15 min clips, I've always found it's a lot easier to work in smaller chunks, like MM on a 2 min clip seems long even.

"every time" I click it takes forever to "wake up".

I can tick three strokes and i have to wait a long time before i can tap again...

Ha. Yea - it's not instant. But much faster than that first click.
Bound to get better in time and optimization.

Interesting you mention 3-clicks. That's my 'normal' # of clicks to isolate a subject.

waltervolpatto wrote:Shebanjah, on most machines, the V1 of the magic mask is substantially faster in setting up and executing that v2...

that is one of the main complaints....

This is totally understandable. The desire to be able to work fast is always a high priority. But the legacy MM1 was imo a joke when it comes to quality and stability of the generated mask and never found a use in my work at least. The model was simply too early to be of use for any high quality work. I'm really certain that a few papers down the line we can have much faster AI models for these tasks but the ones currently out there in the field are simply this heavy. It is not fully on BMD that this performance is 'bad'. It's just the way it is and the quality makes up for it enough for my work at least. But we know things grow really fast with AI so in a few months there may be already a new model doing the same thing 10x faster with equal or better quality.

It's only up to BMD to decide when they think it's time to implement a new model. The problem with these tools is that each model requires new implementation with lots of added data and the necessity to keep the old data for backwards compatibility. Unless they decide to separate this from the main install and allow users to choose which models are available for any AI tool in the software things will start to stack up w.r.t install size and option bloat in the tool. I would hope they think this through in a clever way.