Generating animatable 3D model of body and face from 2D photos

[CitcyckynibmlmbmmcMlmbmmc]

As someone who doesn't have any technical knowledge in this field, I am curious to know what the state of the art for AI deepfake porn is. And yeah, just for my own amusement. Is it yet possible to:

1) Generate a realistic 3D nude model of a person's body and face using multiple 2D clothed average-quality photos? Their posture might not be standing straight, and their facial expression might not be relaxed. And then...

2) Animate the generated 3D model using AI and text prompts?

By realistic 3D nude model of the person's body, I mean is there an AI software that can estimate the person's body type, cup size, based on what their clothed photo looks like?

How high quality do the photos have to be? Can multiple lower-quality screenshots of someone's face (different facial expressions) from a VHS tape be sufficient?

 

[CitcyckynibmlmbmmcMlmbmmc]

Further to my post, the reason I'm wondering about generating a 3D model first before generating a final realistic video is because I'm worried that a text-prompt-to-video might have physical mistakes and uncanny-valley.

Also, has AI been used to improve the realism of 3d model animations of humans and their facial expressions and reduce uncanny-valley?

 

[TMBV]

You can generate a basic 3D model of anything using AI, you can then export such model as OBJ and import into a 3D software, add a rig and animate it, but it will still be just a basic model, with basic texture, bad topology (the mesh that makes up the mode, probably triangles, for 3D animations you need clean topology, with good edge flow, quads), animating that will be hard, you won't be able to do a really high quality animation, as for letting AI do the animation, there is research into using AI for animation, but so far it's all usually either improvements to existing animation (adding micro facial movements, blending different animations together/merging them, training characters to walk, follow paths, etc) but nothing that can take a prompt and generate a realistic 3D animation, use your generated model and give you something high quality, this still requires a lot of work (fun fact, before I started doing deepfakes, I was learning Blender to make overwatch porn, but it didn't work out lmao).

 

[dubsaves88]

AI = model, not AI, ie. AGI, neural network

tbh its easier to just take a video of someone similar and then faceswap for better results

I would guess there are production grade tools available, ie. Disney, that have the ability to fully faceswap and render and you have to realize that there are teams out there that go by frame for touchups or video comps as in my experience training a model for a profile view is slow/old tech as opposed to front facing

i think what you are looking for is possible, its just on a production level, why not do mocap + texture map - like you don't need a 3D model, you could just take video in front of a greenscreen and then per frame Stable Diffusion will just output - obviously you've seen the clips of renders of people walking etc. and then just edit out the variance

 

[dubsaves88]

someone made a tool that was somewhat similar to what you were thinking: