First semi longform project to color in Flamer rather than Resolve or Mistika (25min doc). Footage came in as ProRes 4444 render to scene detect, at 4K DCI. Nothing special. Single track timeline, with Image effect on each clip. Relatively basic node layout - Primary, a ColorMgmt clip loading a LUT, and then some selectives with Mastergrade. I don’t think there were any unusual nodes or matchboxes anywhere.
Anyway, I had colored it a few days ago, no problem at all. Today I did a final trim pass and tried to make a review render which turned into a multi-hour ordeal to get a usable file.
Essentially there must be a VRam leak somewhere, because every so often Flame would die with a ‘out of memory’ error and VRAM was pegged. This is with an A5000 GPU with 24GB. And I’ve done plenty of short-form jobs, big batches, everything. Never had a problem.
In the end I had to break it into 8 reels I rendered at HD resolution, restarting Flame between most of them, depending on how red resource manager was. Finally made it through that, stacked the reels on a track above, and then rendered that out.
Is this normal? I guess maybe memory leaks don’t show up as much in short form work and Batch, but then when you have a few hundred image effects to get through, boom, need extra coffee.
Once or twice the message about not using BFX or timeline FX comes up.
Sorry to hear about your shitty day.
Some of us have been there before, and don’t do things that way any more.
OK. So how do you color a medium length doc in Flame if you’re not using TL-FX? Just use Resolve like 3M other pixel pushers? Doesn’t seem like the answer.
And I’ve delivered a sizeable number of projects with color in TL-FX without any issue like this. It generally works really well, and is part of why I use Flame.
This is not a shot based timeline. There are no comps, no node trees, no detailed versions. Has no business for batch at all. Would just turn this into a time and money looser.
When you do a pure color job, you stay in the timeline, work right on the clips, use TL-FX explorer and and fly through the timeline. That’s it.
There was a time, where Smoke existed for this particular purpose.
And it works perfectly, unless you GPU is leaking like a sieve.
Software has bugs, they should and do get fixed. The existence of bugs doesn’t kill the associated features, or there would be no software worth using.
But I’ll take your wisdom. That’s it for me and Flame then. It was fun while it lasted. There is always the essentially free Resolve that will save the world.
Long form has always been an issue—even in the Smoke days before Anniversary edition and more so after. If you poke around here you’ll find a handful of post just about playback on long form not working. Only thing I can say is 12 minute reels for me was the sweet spot—anything longer was flirting with disaster.
I think that seems to be the right approach, and is manageable. That’s how I got the render to work. It also keeps the storyboard and storyboard reel navigable, though I did discover some undocumented keyboard shortcuts on the reel. Ctrl+PgUp/PgDn helps with navigation, as well as Ctrl+Home/End.
I submitted all the crash reports, lets see if we can get to the bottom of this. I’m not deterred (and the screenshot was not serious, of course).
That does not sound normal to me. I haven’t done any med or long form pieces in a bit. But I have had timeline loaded with Image nodes and some with BFX in addition to Image nodes hand haven’t run into the issues you describe.
Pre-Anniversary edition I had cut a 45min image/sales/doc for a government agency and smoke/flame held up well for me.
That’s been my general experience as well. For some reason this project is very different for no apparent reason. We’ll see if support can glean anything from the crash reports. It definitely smells like a memory leak, because itdoesn’t fail at the same point, but after a similar amount of processing but different places, and restarting clears it.
But the nodes used aren’t the usual candidates for a memory leak, hence my surprise.
flame is a lovely way to handle timelines - simple ones, complex ones…
seriously, multi track, multi version, open clip timelines, shared instantaneously between workstations is a joyous thing to behold, and an easy thing to maintain and improve.
last year i did an hour long multi-track ACES timeline in flame 2023, hundreds of shots, some bfx, lots of action TLFX nodes, tons of pattern browsed shots, layers of GRFX. did not use image effects once. Running an RTX 6000 with 24gb - had to make sure i kept saved desktops closed in the library but besides that - solid as a rock. it’s gotta be something with the image effect.
sorry to hear you had to scramble like that. worst feeling in the world. def been down that road.
I don’t have any troubleshooting tips to offer, but will say I’ve done a bunch of similar longer form work with 100% timeline color and basic BFXs here and there…never really had that issue and I’m living dangerously with a 3090 gpu.
I assume you weren’t using any ML selectives? Those seem to bite me the following day.
That fits my general picture. While there are always a few minor issues here and there, generally this is a solid solution, and I have several other projects without issue.
No, there are no ML selectives. Only a few basic diamond keys.
There are two things I don’t usually do - but both seem immaterial at the surface. I tried a different node structure, that relies more heavily on the ‘Primary Grade’ rather than just selectives. In part since I learned that all the selective keys use the primary grade as input, not their serial predecessor. So doing basic balance in Primary Grade has benefits, even though it comes at a cost.
As a result though I parented two additional nodes from the Primary surface, other than the pre-existing Mastergrade.
There’s a color management node to load a 3D LUT (film print emulation). But a 3D LUT processing should be negligible. The other one is a the channel_matrix Matchbox, for a bit of channel mixing magic in the look. But it’s a plain vanilla Logik matchbox, which is hard to see how it would create a memory leak in the GPU.
This morning I’ll clone the timeline and remove some stuff to see if I can isolate it. One is to remove that Matchbox, and double check the Image nodes to see if there’s something there I forgot about.
I totally believe you there is a GPU mem leak. Ive been experiencing similar things in 2023.3.2. Minimal shot work. If I forget to check VRAM levels before whipping out a NeatVideo, boom, crash.
But about the splitting into reels part… would it be easier to render chunks with In/Outs turned on? Then splice those together?
I took ~5K frames, set in/out, rendered and then placed on top, rinse repeat, then took the top layer and rendered that out. It’s a bit slower, since with marks set, Flame creates a temp timeline, renders each frame, and then writes the file, instead of writing the file as it goes. But at least I made it through.
That said, Flame also crashed on me when I watched the timeline down in the effects tab with range set to timeline. So I had to watch in timeline mode and then switch to effects tab selectively to fix thing during the watch down.
It’s definitely triggered by a specific render process that piles up GPU memory.
So setting up a project with reels may still be beneficial to safeguard against working in the timeline crashes, rather than just renders.
Just me opening the project to take this screenshot and moving the cursor around in a few places resulted in this, normally the red max on the right would be more 10-20%.
funny you should mention this @allklier on the same long from project i mentioned above Search&Select in timeline was an absolute life saver to locate specific shots. Never used it before as i live mostly in TVC land.
i wish it had a UI component where it lists everything that is a potential match and allow you to select the one you want and then take you to that, but trying to find a needle in a 1 hr haystack it’s invaluable.
