I have been experimenting with AnimateDiff for stable diffusion 1.5 in A1111.

Stability ai has recently launched Stable Video Diffusion (SVD) which might be better. I am always a generation behind on tech because of how fast it is moving.

This is a simple result with a simple prompt. I used prompt travel to make the subject smile. It is hard to get a more complex result with more movement. The subject's arms often melt off or transform into part of the body. These simple results that have little motion are easy to make. It makes sun spots move, grass wave, waterfalls fall, and streams flow. It will make subjects blink. Sometimes it is only a half blink which makes her look like a lizard person. Sometimes the model will be still and then suddenly move her arm or leg abruptly which makes her look even more like a lizard person. Sometimes the eyes twitch. It is very creepy and unnatural sometimes.

If anyone is interested in nude subjects covered in mud or colored goop, I might post a few more updates here. If anyone else is experimenting with Animate Diff, feel free to add.

Here is an example showing something which resembles fluidic motion on a butt. It shows that the model which I am using does have some concepts of fluidic motions, but it is not convincing that it is real fluid. In the prompt travel, I tried changing the substances and colors.

The smile is definitely convincing! Such small details make me love WAM content.

Also the fluidic motion is interesting. But I bet it will evolve quickly as we are experimenting with it. Like from "clothing that looks like slime" to "slime that flows actually over the clothes".

Thanks for posting these, I know it takes some effort to produce

I've made a few but I cant post them because to get slime to move I have to use a video with someone being slimed and then its literally reskinning someone else's video and making it worse...

I stopped using animatediff when the control net in the last attempt was so accurate you could even identify the model from the video the motion capture was from, not ok...

Got high hopes for Stable Video anyway lets see whats next.

I have such a difficult time with cloths hehe. I usually give up and just stick with fully nude images. I really want to be able to generate videos of substance over cloths and especially stuffed in cloths. Most of the time, the cloths turn into a substance instead of appearing on top.


It is amazing how fast trends change. I'm still stuck on 1.5 and stubborn to try switch to SDXL or learn comfyUI. There are some advanced workflows with controlnet which I haven't even began to try. Just as I learn and get used to animatediff, SVD comes out . I nearly gave up on it. Iterations are very slow. I tried upscaling one result which took several hours on my 1080ti. I made one gif which is 960x1080 which is 21MB.

I have like a backlog of things I want to try, and now SVD is on there.

It is really challenging to get good results. Either the eyes are twitchy, they stare, or do half blinks. I have no control over what animates. The slime could remain in place, sometimes it flows down, and I have even seem slime flow up. Here are 3 results where I did not make the gif repeat.

Likewise! Whole thing is frankly just too much of a time sponge, not to mention how long the compute itself takes! I've decided to limit myself to making punctuated workflow upgrades when something major changes, rather than adapting to every single incremental improvement. ComfyUI looks very good though I must admit.

I had some luck with using simple animations as the driver, taken from a well-known youtuber's channel. Obviously won't be showing those either out of respect for their copyright, but the point of interest is that the controlnet driving the model inference can be quite basic.

Nice, this looks impressive. I like this. The details look good, and it is a convincing camera motion. Plus who wouldn't want to bath in custard with candles.

Wow, amazing! Could you give some insights?
For example I would like to know how complicated it would be to add small actions like "blinking".

It's not possible to tell it what to animate so it basically picks something to move could be the body can be the fluid can be the background. Only just started playing with it but I have a feeling it's going to be a case of running an image hundreds of times through the process to get 1 animation which is really good.

Comfyui workflow here https://education.civitai.com/quickstart-guide-to-stable-video-diffusion/
Have a go