Been a while since I posted anything AI related, but with good reason. Got into using ComfyUI and all the new custom nodes that kept being released.

One in particular you might have heard of is Animatediff (open-sourced txt2vid, but a bit more hands on than runaway or Pika labs), which when combined with IPAdapter (you can use an image as a prompt + text) and FizzNodes (scheduled prompt changes over time)....can create MAGIC.

Here's some example, tried to filter out all the weird ones, Animatediff has a habit of generating body horror

                               

Those are GIFs btw

I "might" take requests if anyone has interesting source images (ideally non-preggo, idadapter is a bit dumb when it comes to bumps).

Also this can handle non-morphs a lot better actually, Animatediff will simply add motion, but that's no challenge at all.

A few other recent ones:
                                                               

Those last two....yeah

I'll throw in some requests

                                       

Finally got back around this, although I started testing the new Stable Video Diffusion.

Right now it simply does img2video, but the test I just ran looks promising (and works pretty well with 8GB).

Source Image:
   

Resulting Video (Plus some face swapping and upscale):
   

Gonna see what else I can test

I'm quite surprised how well SVD works, for the bottom generations I took some random google image searches, and a few Dalle3 images I'd made.




   



These 3 are my favourites, from a single Dalle3 generation:

           

       


       

           


       

What gets animated is a bit random, but mostly depends on the content. 
In the ultrasound example it was sort of a given it would animate the hand, similiar with the finger in the navel.

Yes to reiterate, all of these videos came from single still images

@geara I borrowed some of your excellent Bing/Dalle3 generations to show how well this works (https://preggophilia.com/my-domestic-fan...ith-bing-t)

                       

Was testing whether adding a depth-of-field to images would direct the motion specifically, sort of worked

                   

There's actually belly physics

There's been some updates to the world of SVD (Stable Video Diffusion), a microsoft research team have finally got a working public implementation of motion control.

       
       
       
       
       
       
           

Yeah it's a bit jank right now, but given before it was totally random what you got, this is progress.
There's a demo that I won't link to; some folks here might "try" uploading.... questionable content, so yeah we'll just have to wait until it can be used offline.