>>15483
As someone who tried to generate some stuff, must say this is simply amazing.
Can you share some technical details on what tools were used to drive the WAN model? Was it ComfyUI? If so, what was the workflow, at least approximately? Especially the part with sucking the nipple.
Any LORAs?

>>15615

Thanks for the compliment.

Yes, I'm running ComfyUI. I tried to get started a few weeks ago playing with ComfyUI using the graphics I used to have, but that was hopeless. I got myself a refurb card with an NVIDIA RTX 3090. That's doing a decent job, although rendering is still a bit slow compared to the online rendering services I had experimented with before deciding to create my own setup.

I didn't use any special LoRAs for this video, just a couple that are a pretty basic part of a Wan 2.2 workflow. (I've attached a picture of that workflow.)

Oddly enough, it's not too hard to convince the basic Wan 2.2 to do nipple sucking. It definitely helps if you start with an image where your characters are positioned to make that an easy action. You probably need to run multiple attempts to get it right, however. (I'd have preferred, for instance, in the case of this last video, to get slower and more sensual sucking.)

I tried an "NSFW" workflow with a pair of NSFW LoRAs (Wan 2.2 often requires a matched pair of LoRAs for some things), and that made things worse. This particular pair of LoRAs where obviously trained on bouncy, flirty, hip-grinding porn with even less concept of violence than the standard Wan 2.2 model, and had a tendency (likely intentional, if not merely due to the training material) to morph younger characters into adult characters.

There are, of course, many more NSFW LoRAs for me to try someday.

Pretty much anything else I can think to tell you is in my original post.

What I really need to figure out how to set up and use next is a workflow for Wan 2.2 that uses something called SVI. SVI apparently helps overcome, to some extent, how brief each Wan 2.2 video clip is, stitching multiple clips together more or less seamlessly, and helping to maintain continuity between the multiple clips.

>>15624
Thanks so much for the detailed reply.

I am scratching my head a little bit looking at this workflow, as it is very different from image generation (or at least it appears different), but I will figure it out.
Did you notice VRAM utilization during generation? I have a 16 GB GPU and a little bit worried. Maybe I can start with a lower res...

>>15713
I haven't been checking my VRAM usage, but with 24 GB on my new card I haven't had to worry about it.

>>15713
Also using RTX 3090 with 16 GB, wan 2.2 hybrid does a great job with the basic setup and 5s videos, loras on the other hand are a pain I don't recommend anyone using them until they upgrade to a better card, something I will do very soon, thanks for the info

>>15483

just out of curiosity
have you ever considered doing dolcett videos?

Great work.. how do you get characters to stay dead with WAN 2.2 Maybe it's easier with decapitated heads, but intact bodies have a nasty habit of coming back to life!

>>16044
> just out of curiosity
> have you ever considered doing dolcett videos?

Not in the sense of, say, copying particular multi-panel stories, but I certainly like many of the same themes.

>>16069
> Great work.. how do you get characters to stay dead with WAN 2.2 Maybe it's easier with decapitated heads, but intact bodies have a nasty habit of coming back to life!

Just a matter of trial and error. The video here was made by assembling several indivual clips using DaVinci Resolve. Each clip was typically the outcome of many retries with more or less one general prompt for that clip, adjusting the wording until both the wording and shear luck got me an acceptable result.

I haven't had much success yet trying to use SVI for generating longer, consistent clips, but figuring that out is a work in progress. I'm also just now starting to experiment with creating my own LoRAs, in an effort to be able to have consistently rendered characters across clips, but wow, that's a time-consuming process -- not just the compute time of many hours, but the prep work to make 20-30 different poses of the same character as training data.