I have it running as an ONNX.
, that the ONNX file fails in Flame because of input values and input channels. It changes the build script to what version of flame and GPU are you on (and how much Vram)
Latest and greatest Flame 2026.2.2. 48GB M4 Max.
But I did the conversion on a gaming PC with 128GB RAM
Every day flame gets more like a west coast custom car meet, and less like a bunch of engineer/prison wardens round a sad little coffee percolator in a dark corridor next to the fire escape.
Hereās a 1024px version too, runs much faster of course.
The way everyone is using ActionVFXā¦ Awesome. I can stop working on this lol.
Can you upload it to the portal? That way everyone can download it easily.
Well, your chance to join the herd of model-influencers. This forum reads more like IG for VFX AI lately than perspective grid tips, or how best to do a projection paint job. The days when extended bicubics were a cool thing.
PS: you still need them to polish the turds off the AI results, but nobody wants to hear you talking about it anymore.
Not totally wrong. I mean AI is reshaping all our jobs faster than we can look, and sharing insights and helping each other is what this group is about. If Logik becomes a bit less Flame and more support group for VFX artists in 2026 thatās cool.
Thereās a bit of race on who can get somewhere faster, rather than more coordinated effort. A fight or flight reaction to the current times. But plenty of sharing and so itās in the well established spirit.
Flame on!
2 Likes
Thereās definitely some truth to that. I have been trying to fill holes that align with bigger projects, while I wait for 2027 and @philmās FU suite. In much the same way folks had rallied around Logik Projekt I hope the FU suite will become a bedrock point of reference for us all.
Then maybe we can start making a check list of shit we can dev together. Later we can braid each otherās hair.
for the old baldies among us, these skills are surplus to requirementsā¦
I managed to bumble my way through the JSON file creation and correct the order / colour of the inputs and outputs etc. Iāve submitted the full resolution version for now. If thatās all good maybe let me know and I can upload a 1024px version too? I think if all you want is the edge colours and youāre doing say an IBK for the actual key then running the model at 1024 becomes a bit more performant.
FINALLY, my pybox implementations are now live as well here.
Features:
Native MLX on Apple/ Cuda on Linux
Input size control between 2k/1k processing
Quantized option on Apple
Simple colorspace conversion to srgb for processing
ā¦and whatever the fuck the despeckling controls from the oringal do.
Simple demo video below (which was recorded driving flame with a mouse)
I had some time today so I put the pybox implementation against EZ-CorridorKey on a M4 w/ 64GB.
Sequence: 2048x1080, 100 frames. Linear.
Pybox: 9min
EZ: 1min 30sec
Pybox seems to be Pyboxingā¦
That said, how EZ-CorridorKey manages the processed frames is annoying. It dumps everything into itās own project folder meaning manual work to copy it to where you want afterwards.
I still donāt understand the use of corridor via pybox, where the matte is extracted from. Iām a bit out of touch with the updates, but if itās for clipping and you already have a good clipping, whatās the purpose of the corridor? Thanks to the developer.
@kyleobley Using EZ Corridor, one would really need to think of some script to automate the return to Flame. Something more or less like what Iām using in Sammie.
@wiltonmts Do you find VideoMama better than matanyone2 even at lower rez?
I find VideoMaMa more temporally consistent compared to MatAnyone2. Itās very hard to get a hole in the middle of the mask or have it bleed into some other region if the segmentation tracking is good. Since itās a diffusion model, it depends heavily on the quality of SAM2 segmenting across all frames. MatAnyone2, on the other hand, uses a memory propagation approach, processing frames sequentially ā which makes it significantly faster. VideoMaMa also produces slightly softer/blurrier edges compared to MatAnyone2. I ran several tests trying to address this in post-processing, but at the modelās native resolution the result is the same. Itās worth noting that while MatAnyone2 accepts any resolution at inference time, its training dataset VMReal was built primarily at 720p ā so both models have their own resolution-related trade-offs.