Super Strength for Wan

A Wan video LoRA that's mostly about breaking stuff. Both High and Low noise LoRAs are required, but they're each about 150 MB, which is half the size of most Wan LoRAs.

This LoRA works well as either image-to-video or text-to-video.

This was the most challenging LoRA I've made yet, and it cost me a fortune in buzz to make enough source videos to build up a dataset with the different actions. So if anyone feels like throwing a few buzz my direction, I'd appreciate it.

Important: This model is 100% fictitious and is not based on images of any real person. The dataset was entirely AI-generated. No images or videos (neither actual nor AI-generated) of any real person was ever involved in the making of this model.

The actions are inspired by superhero movies and comics. It's already possible, especially with image-to-video, to make videos about people picking up and carrying very heavy things like school buses and such, but Wan is very bad at deforming and breaking objects. My goal was to teach Wan about bending and breaking things. I don't know how well Wan will generalize these lessons beyond the specific examples in my dataset, but we'll find that out together if people download this LoRA and get creative with it. My hope is that it will make it possible, or at least easier, to make action scenes where stuff gets broken in different ways.

The dataset included videos of:

• bending pipes, metal bars, thick pipe wrenches, metal baseball bats, etc.

• bending open metal bars as you'd see on a prison cell

• crushing metal objects like a metal skillet, a toaster, a metal refrigerator, a metal safe, a metal wardrobe

• crushing crumbly objects like concrete blocks, billiard balls, bricks

• pulling a car engine out of a car

• tearing/breaking sturdy things like heavy doors and manhole covers

• pulling doors and metal panels off of buildings and cars

My hope was to teach Wan how it looks when metal is crushed, when it bends, dents, and crumples. I also wanted better renderings of crumbly destruction with small fragments and dust falling. Without help, Wan doesn't seem to know how to bend metal and it crushes bricks and stones into unrealistic smaller bricks and stones. This LoRA is generated as a text-to-video LoRA, but it often works better in an image-to-video setting since the first frame can sometimes be made better with an image generator like z-image or flux or whatever. If you use this model in image-to-video, just be sure to describe enough of the important parts of the scene (the person, the object you want to crush, etc.) before you talk about the action you want the character to do. The base model for this LoRA is text-to-video and that works well, but all the sample videos are image-to-video because it was faster for me to get the samples I needed.

I also hoped to teach Wan about a few other crushing actions like a beer bottle full of beer and some other items made of different materials, but it may have been too many actions and variables in the model at once. In particular, adding in anything filled with liquid gave me a LoRA where nearly every object squirted out liquid when it was crushed, and I couldn't figure out how to get past that no matter how I changed the captions for the dataset. So unfortunately that kind of action is not included in this model. Maybe someday.

I'm sure this LoRA is far from perfect, but it's much better than what you would get with plain Wan 2.2 without the LoRA at all. If people use this LoRA and post their videos to this model page, I may be able to use some of those video clips to add to the dataset and improve the LoRA for a version 2 in the future.

For those who are curious about my dataset, I had the best luck using OpenAI's Sora engine to generate the source videos. All of these actions (including things like bottles of liquid) were possible in Sora with the right prompting. I had occasional good results with Kling, but it wasn't reliable. In every AI engine, including Wan even with this LoRA included, the biggest problem is deforming an object and having it STAY deformed. In many cases, the AI engine will bend or crush an object, then it pops back into shape, which is not what I was aiming for. In any case, you can make many of these videos with Sora, but if you want to combine the video with a character or scene that you like in Wan, or if you want something really risqué like a woman in a bikini, then you can't use Sora for that because it doesn't pass the censor.

I hope you have fun with this model and please don't forget to post your videos with this model tagged so I can see what kinds of creative scenes you all come up with.