Need info Davinci Resolve Fusion - openEXR sequence workflow

Good day.

I love working with EXR's. I like having everything I need contained in one file, and I find the options available in fusion to be intuitive. Need motion blur? assign red channel on the vector motion pass to velocity X, Green channel to velocity Y, toss in the motion vector blur node and voila! Works as intended.

Where I work we do some pretty basic compositing. I tend to model complex machinery, then highlight various components on it as the camera orbits around using a multitude of alpha channels. It's not uncommon to have 20-30 alpha channels assigned to specific pieces of the 3D model.

I feel like openEXR SHOULD be ideal for this. I can import my sequence, create a timeline, hop into fusion, and do my initial setup, then duplicate the media node as I see fit and simply swap out to the appropriate alpha channel.

But things get REALLY bogged down. GPU memory full errors, crashes, insanely slow playback (like 1 frame/15s slow)

My most recent venture has had me frustrated. The total length of the sequence was 60 frames. 3 of those frames were isolated and extended as the camera stayed there for a minute while different components were highlighted, before continuing on with the orbit.

Checking the properties of the folder where the sequence was located showed me the entirety of the folder was 1.58GB. I have 32GB RAM, 4GB of memory on the card. I run an i75820K and my GPU is an older GTX 970. I know this isn't the most modern workhorse, but I'm also not compositing a blockbuster here. I feel like my specs should be entirely adequate for fusion to handle 1.6GB worth of data. I'm not doing any color correcting in Resolve beyond a colorspace transform from Linear-sRGB.

Has anyone reading this worked with EXR sequences in Fusion with positive results? Is there a workflow that I'm not understanding? Are there resources that you know of that I could look up that would help me with performance?

Any tips would help!

Thank you

-J

I feel like my specs should be entirely adequate for fusion to handle 1.6GB worth of data.

I'm not an expert on this, so I might be wrong, but 1.6GB worth of data on your disk in no way reflects how much data you actually have to work with. Once you load in your data it's uncompressed in memory, so if you say you extend three frames for a minute, you no longer have three frames, but you have 1440-9000 frames that need to be in memory depending on your frame rate. And if those involve lots of layers (20-30 alpha channels) that effectively means you're working with 1500x20 = 30,000 frames in memory at least. That's no longer 1.6 GB worth of data.

If those frames are extended as-is, just holding them, and there is any reasonable working cache, they should all evaluate to the same cache frame. Only after some animation or masks or whatnot are added, change is introduced and each frame becomes unique. But source should still read in very fast, from same cache.

Switching to Fusion Studio would help with the GPU memory full issues and probably with a lot of other crashes as well.

Maybe rendering an extra cryptomatte instead of rendering all the alphas into one exr could also simplify your workflow (maybe even improving).

Other than that, I'm working with exrs in Fusion (Studio) all day and yes, sometimes there is so much data that not everything fits into RAM, but with the right caches that is usually fine. Or using a flip book/RAM preview if I have to. But in those cases it's usually better to use a saver instead so you don't loose the preview.

I tried using the Fusion page for big(ish) shots before and had a similar frustrating experience. I didn't try too long since I am a Fusion Studio users anyway, so maybe there are ways to make it work but I'd suggest using the standalone version.

Edit: also using multipart instead of multilayer exrs can potentially decrease I/O times, but that's not really something that's up to you. Either your rendering engines renders one or the other.

Once you're in Fusion Studio, the other thing you can do is turn off GPU rendering. It might help with everything being passed through your cards till you figure out what's going on. Here's a screen grab of a project I'm doing that uses 1920x1080 32bit Multi-image EXRs. A normal pass is around 12 - 16 active loaders. This one is less so I quadrupled the loaders and trees just to see what would happen. Inside of Resolve 17.1.1 it using about 7.8GB of RAM. If you're really using 32 distinct alpha loads at 32 bits per file then you might have an issue with you GPU being large enough.

It might be a pain but you can also take all of your alphas and do a pack like you do for a video game asset. Run each alpha into a separate channel using a Boolean. You can pack 4 alphas into each Boolean meaning you should only have eight Matte EXRs that you then use to distribute the 32 unique mattes. It might help take the usage down.

It's a tough call for studio standalone as I like the environment in Resolve, but that may be the route I have to take.

When it comes to channel booleans, that's how I've done it in the past and performance was definitely better, since I would save out my final color (Foundry Modo terms here) as 16bit tifs, while saving out the separate alpha channels as 8bit png's, since I really don't need extra bit depth for them. Any aliasing can be fixed if needed. Performance was definitely better, but then file management becomes a bit more confusing. Pros and cons, like everything else in life.

The thing is layered EXR's are just so damned convenient and I would REALLY like to be able to reliably make use of them. I do have a 3080 on pre order which I should receive this summer (hopefully!). This will likely help a bunch. For the time being, though, I can only use what I have.

Modo unfortunately doesn't save out multipart EXR's. Which is really dumb since the same company has been spearheading it's use with Nuke. Hopefully in some future versions support will be added.

This project is done and I've got about a week of downtime before the next one kicks off, so I'll make use of that time rebuilding the project using cryptomattes.

Thanks for your input everyone. It looks like I have a couple things to try out. In the meantime if anyone else has insight I'm all ears!

Regards,

-J

Usually this is a bigger problem for swapping from Resolve to Fusion and back again, but since you're using EXRs, you can copy and paste back and forth between the two. That way you can work on a problem using "scratch comps" in Fu and paste them into reFu. Fu is more forgiving to play around in.

And of course there's still VFX connect if I'm not mistaken, which will allow you to work in Fusion SA from within Resolve. It will be a little less interactive, but if your system supports it, the gains can be substantial.