3 ms·
Hmm, I thought deep fakes consume lots of computational power both at training and inference time. Is there some tech/science breakthrough here or they have jus
by kavalg 7y ago
Hmm, I thought deep fakes consume lots of computational power both at training and inference time. Is there some tech/science breakthrough here or they have just scaled it up to more hardware?
- mavhc 7y agoI think it's the training of the model that requires most of the time, once that's done it just needs to decide what to swap with swap. So I'd expect it to only work on the 1 person in the demo, with the 4 famous people depicted.
- jorgemf 7y agoTraining yes, but for inference a commercial GPU can do that at 14fps with a small latency (very noticeable in the video)