The looping here is the lovely.

This has got to be the first ever AI generated animation i’ve seen that’s actually coherent, and not a timelapse acid trip lmao

zegreenstardood said:
This has got to be the first ever AI generated animation i’ve seen that’s actually coherent, and not a timelapse acid trip lmao

It's not worth calling animated, it's just puppet-warping between different angles in the feature-temporal space. It's done by creating a batch and then interpolating between the similar weights. It's temporally consistent here because the motion module is overfitted to this exact animation. The trouble with MM tensors is they're limited to small clips before running out of memory, they're like 3D bitmap proportions of heavy data processing.

Animation has timing and flow to it. Just because it's moving like a jerry-rig doesn't mean it's animated, it looks like its dragging MPEG artifacts around in an over-compressed video. The motion aspect isn't really a part of the neural network either so you don't get actual control over the video logic at all.
AI animation just doesn't exist yet. Or well it does actually but current AI anim is so low res it looks like a mess of noise that barely differs from the stock trained video. Read the Tune a Video paper, it's not generating any original movement but you can use ControlNet and TD to do vid2style pretty well now. I've not seen a single so-called AI video that had original movement, the most creative it can do is just add a bit of head/arm movement to an otherwise still image. Not seen any papers yet either. Besides it's better to generate mocap data and then rotoscope that with AI. 2D copy buffer AI isn't the future of AI anims, gaussians is. Seeing these puppet animations is cringe, it's not worth the GPU cycles at all. Even NFT AI videos from 3 years ago are 100x better than these head-turn-arm-wiggle anims. Just see LeiaPix and other crap if you want cheesy parallax effects.

sagefurry said:
It's not worth calling animated, it's just puppet-warping between different angles in the feature-temporal space. It's done by creating a batch and then interpolating between the similar weights. It's temporally consistent here because the motion module is overfitted to this exact animation. The trouble with MM tensors is they're limited to small clips before running out of memory, they're like 3D bitmap proportions of heavy data processing.

Animation has timing and flow to it. Just because it's moving like a jerry-rig doesn't mean it's animated, it looks like its dragging MPEG artifacts around in an over-compressed video. The motion aspect isn't really a part of the neural network either so you don't get actual control over the video logic at all.
AI animation just doesn't exist yet. Or well it does actually but current AI anim is so low res it looks like a mess of noise that barely differs from the stock trained video. Read the Tune a Video paper, it's not generating any original movement but you can use ControlNet and TD to do vid2style pretty well now. I've not seen a single so-called AI video that had original movement, the most creative it can do is just add a bit of head/arm movement to an otherwise still image. Not seen any papers yet either. Besides it's better to generate mocap data and then rotoscope that with AI. 2D copy buffer AI isn't the future of AI anims, gaussians is. Seeing these puppet animations is cringe, it's not worth the GPU cycles at all. Even NFT AI videos from 3 years ago are 100x better than these head-turn-arm-wiggle anims. Just see LeiaPix and other crap if you want cheesy parallax effects.

i ain’t reading allat

sagefurry said:
It's not worth calling animated, it's just puppet-warping between different angles in the feature-temporal space. It's done by creating a batch and then interpolating between the similar weights. It's temporally consistent here because the motion module is overfitted to this exact animation. The trouble with MM tensors is they're limited to small clips before running out of memory, they're like 3D bitmap proportions of heavy data processing.

Animation has timing and flow to it. Just because it's moving like a jerry-rig doesn't mean it's animated, it looks like its dragging MPEG artifacts around in an over-compressed video. The motion aspect isn't really a part of the neural network either so you don't get actual control over the video logic at all.
AI animation just doesn't exist yet. Or well it does actually but current AI anim is so low res it looks like a mess of noise that barely differs from the stock trained video. Read the Tune a Video paper, it's not generating any original movement but you can use ControlNet and TD to do vid2style pretty well now. I've not seen a single so-called AI video that had original movement, the most creative it can do is just add a bit of head/arm movement to an otherwise still image. Not seen any papers yet either. Besides it's better to generate mocap data and then rotoscope that with AI. 2D copy buffer AI isn't the future of AI anims, gaussians is. Seeing these puppet animations is cringe, it's not worth the GPU cycles at all. Even NFT AI videos from 3 years ago are 100x better than these head-turn-arm-wiggle anims. Just see LeiaPix and other crap if you want cheesy parallax effects.

using puppets are actually an animation technic, unless you are thinking about animated filters on real actors