Mocha is way more fun than Wan Animate.

For the last couple of days, I've been messing around a bit with Mocha, the newest development in video character replacement, but as we've seen from across the internet, not just character replacement.

When Wan animate first came out, don't get me wrong, it did (and does) a great job when used correctly (like everything else). The thing about Wan Animate that I didn't like is that the set up felt "clunky". There were a lot of things to think about when generating with Wan Animate - it doesn't just work out of the box.

Beyond that, in my experience, you can't always predict what you're going to get. While that's the same for just about any AI model, I've found Mocha's results to be really fantastic and exactly what I was looking for or envisioning with an extremely simple set up.

Reference Images:

Original - Whiplash

Granted, the model gets a bit of a boost in the sense it's introduced so late into the Wan saga where so many users already have the necessary models to run this. On top of that, with the Kjai model, you're only looking at about 14g for the model which is great! If you have >14gb VRAM, you should be able to run this locally!

On the other side of the coin, running this in Runpod with a higher end GPU really let's you expand in the sense you can generate longer frames. For example, one sequence here is 260 frames long!

I'm really impressed with the expressiveness and realism that Mocha provides with just a very simple reference image. The references I gave to it were intentionally not good, and while not every generation came out perfect, 9/10 generations were more than usable, especially when run through some very light post production.

Beyond that, like we've seen with Wan Animate, this model works really well when using I2I to change the first frame of your video into your desired character and/or style. The lip-sync is also fantastic without any control net or depth maps of any kind! I will say it did struggle a bit to focus on replacing characters on the edge of the frame, but a re-render typically solved the issue.

Overall, if I had to choose between Wan Animate and Mocha, so far, I'm going Mocha every time. Especially after some upscaling in Topaz, the results it gives are just way too good and, at scale, could totally be used to create some really engaging stories. We definitely intend for Mocha to be accessible in Silent Snow AI, but for now see our attached workflow here! We won't be explaining this one, but if we feature Mocha in one of our upcoming videos, we'll deep dive it further!

Have you tested Mocha? Which would you use?

Kjai Mocha Model: https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/blob/main/MoCha/Wan2_1_mocha-14B-preview_fp8_e4m3fn_scaled_KJ.safetensors