The Wan 2.1 Breakthrough

Horny TL;DR: Enjoy some HD dicks and tits ya goons!

Find 4 videos attached to this post and 1 here:

Releases\2025\004-Big Balled Lap Dance\Video\Up-Int

and 3 here:

Releases\2025\017-Mom Balls Vacation\Video\Up-Int

But, if dense technical reading makes you hard, read on:

Happy to say I've had a very few productive days back from vacation. We are cookin (my GPU that is)! Not only have the gen gods been with me, but I've been trying out new stuff and talking to people in the amazing AI community and everything so far has just been working! And working amazingly well. So, right now the new big things for Wan 2.1 are:

• GGUF quantizations (these are intelligently reduced sizes of the model, previously I was using the FP8 which is quite large even though it was the smallest quantization available). I've tried the Q6_K and the Q5_K_M and both work really well. Can't decide which is the sweet spot yet. But, with the reduced size they are both faster than the FP8 I was using and both get results that seem as good.

• Skip Layer Guidance (SLG) - folks seem to be discovering that turning on SLG to block the 9th and/or 10th layers of Wan 2.1 seems to provide better motion and less weirdness. My initial attempts at this tended to just create an incomplete image that looked like everything was on fire like it was a "this is fine" meme. But, this has been solved since I found an off-the-shelf community workflow that is really, really good.

• LoRAs - Low Rank Adaptations. You've probably heard me mention these in passing. These are custom trained models that plugin to larger models subverting their generation. For example, you might have a LoRA to teach your model about a sex position so you can generate images similar to the training images when using a specified key word. We are reaching the point where Wan 2.1 now has a ton of LoRAs for it. Which is great since the base model, while able to handle NSFW stuff serviceably, it could be a bit odd since it lacks training on particularly spicy stuff. I'm now pulling in LoRAs for breasts, penises, cum, jerking off, etc. Of course, to mixed success, not sure if it is the randomness messing me up, some of my settings, or just the LoRA I'm using sucks, or all of the above. Still, this is great. At some point I need to work on training my own LoRA for particular things, but that requires a huge time investment in tagging a catalog of source material for the AI to train on (and GPU time I could be using generating videos and images).

New to me is that I finally have a good working interpolation and upscaling workflow! Which is awesome since I've generated hundreds of Wan 2.1 videos at 480x832 @ 16FPS, now I can make the into 720x1248 @ 32FPS while maintaining their quality. This is timely since I'm now able to upscale all my nude clips for Big Ball Lap Dance so they don't drag down the quality of the non-nude clips. I ripped the nodes out of a all-in-one video gen-interpolation-upscaling workflow and tweaked them a bit and they work great.

I'm processing the remaining 20-odd clips for Big Ball Lap Dance. In total I have about 6 minutes of video I'd like to use and I've picked out the music. I'm wanting to also do sound effects and vocals for the first time. This one will be some work getting it all in the editor and putting it together. But, my plan is to generate a quick 1 min trailer in the next day or so to give you guys a preview.

Lastly, I finally got SageAttention and TeaCache working locally, this has cut my generation time in half for basically free. But, now that I'm upscaling and interpolating my videos and using a GGUF quantization, it works out to be roughly the same generation time as before, though it is more consistent: about 2,200 to 3,500 seconds (36 min to 1 hour) for a 5 second video. Before, it could occasionally spike to 4.5-5 hours :(

For now, I'd like to share some vids I've generated using all my new Wan 2.1 techniques and 1 of the previous videos upscaled and interpolated (the tank top swinging dick for Big Ball Lap Dance). Note, these are 960x1664 @ 32 FPS. I decided that was too big and 1.5x upscaling would be preferred since that'll more closely match my non-nude video res and also have less of an "airbrushed" appearance. Alternatively, I'm going to experiment with an aspect ratio that matches my source images, even if it is a little unconventional, we are talking 416x608 or 624x912 (and then 1.5x or 2x upscale). The reason being is that the current aspect crops the sides of the frame, even if it is a more popular portrait aspect ratio for video (and I'm using the most popular portrait aspect for AI still images, why they aren't the same, who knows!).