turnabout :: test/experiment (Patreon)
Content
Hi all - still working on the final act of the Lisette/Keyni video (I know, I know) but I ran a few tests this week and I thought this was worth sharing with you while you wait.
It's a brief (well, brief for me: 4:32 run length) vignette with Yenneca and Bann, kind of reminiscent of the earlier "worship" short; it's footage I initially made for the "tribute" video feature, but I think it'll end up getting spun out and elaborated into it's own short feature, demonstrating a bit more of the relationship dynamic between Bann and Yenneca.
I'm calling it an experiment or test because I'm trying out a couple of things ahead of the final export of Call of the Wild Act2. As I've mentioned before, I am rendering everything in 480p to start, and I have a whole long post outlining my upscaling workflow. Overall I've been happy with the results; just knowing that I can go from 480p clips to a 4K final video and still have people compliment the quality is insane to me. But I'm always looking for efficiencies - workflow is a big part of what I do for a living IRL - so I wanted to see if it was feasible to skip some or even most of those steps.
So for this video, I took some of the clips that make it up and skipped my initial upscaling process. The remaining clips used the normal process of upscaling 2x using Lifecast Upscaler prior to editing. In Davinci, I then used the "AI Super Scale" feature in Davinci to provide the upscaling - I used the Nvidia upscaler at 2x with Ultra quality for the clips that I had upsampled normally, and 4x Enhanced upsampling for those where I'd used the native 480p clips.
Can you tell which is which?
I then exported direct to 4k out of Davinci, and used Adobe Media Encoder to optimize and to export the 1080p versions.
To my eye the end result looks pretty good. These clips don't use as many effects as I have in the earlier videos, so no added grain or color grading or texture pop tweaks etc., just the raw clips upscaled entirely in Davinci.
As a couple of additional tests, one of these clips was generated using Wan 3.0 - first time I've used that model for something that I ended up releasing. Again, can you tell which one is Wan?
Some additional observations about Seedance for anyone that's read this far:
Seedance has a really weird aversion to cunnilingus. If I have reference images and a starting pose that clearly shows a vagina, it'll do it no problem as shown here. But as soon as the camera moves to not explicitly show the genitals, it reverts to belly or back licking 100% of the time. I am running into it with the Lisette video as well, and it also seems to apply to any sort of penetration - it can show explicit penetration, but it will not or can not do implied cunnilingus or penetration. Similar issue with posing - getting Yenneca to straddle Bann's muzzle is apparently impossible just using the prompt - no combination of words no matter how exactingly described will make her do anything other than kneel or sit on his chest. If I use a storyboard image in the reference showing the pose it'll work, but again as soon as the camera cuts it's back to the chest-sitting. Infuriating, and kind of weird. I also think that Seedance has some concept bleed issues, or it just really likes adding muscle to characters. The reference images for Bann are much less muscular and Yenneca is also not nearly this buff n any of the source images. Not sure what's going on to be honest.
OK, back to work on the Lisette video! Hope this was at least a little bit interesting.
Here's the direct link to the folder with the videos on Mega: 4K24 and 1080p24 versions.