I was just wondering: Will the plugin in the future support dual GPU’s?
Second question: I notice that with the mp4 animation rendering I hit a wall at around 6 seconds, anything over that animation wise and it will error out on me (out of memory, despite me having dual GPU’s) my question is simple: I would love if the final version of the AI render plugin would also support exporting in animation frames. That way it can render animations for as long as you want, since it will release after every frame anyway (so no more out of memory issues) and then you can just turn the rendered frames in to an animation later on. (or maybe even just have FFMPG do it, this would be handy too if the plugin did that after rendering all the frames)
Anyway, lemme know what you think. All this being said I feel that AI rendering for iclone is nothing short of brilliant. You guys are very early to jump on this, the first I think even, and its awesome to be able to just combine your own character animation with AI.
Also I suspect my questions are nothing new, and you guys probably already thought of this. Still, I have to ask.
Animations are limited by the AI model. These are made by other companies like Alibaba, Tencent or LTX, not by Reallusion. WAN for example is trained on generating 81 frames. That will technically give you about 5 seconds at 16 frames per second. More frames will degrade the resulting video. You might be able to squeeze out 100-160 frames, if you reduce the resolution enough to fit in VRAM. That’s why all these AI slop videos cut every five seconds. Longer generation is usually not possible.
How much VRAM do you have? The models in the Reallusion workflow are already highly reduced in size. For the original models you would need an RTX 6000 Pro or H100.
There are ways to create “endless” animations via start/end-frame generation. I haven’t tested if the current Reallusion workflow supports this.
I think you should be able to use multiple GPUs with this extension here:
But I think this will only be useful to save time for loading/unloading because everything can stay in VRAM instead of swapping at each stage.