A Comparison Of Character Coherence In Videos

I dressed one of my dancers in a sort of lamé bikini to test different video models and find out which one was best at maintaining the character's features, both facial and body, over time. The scene is set on the stage of a nightclub. Each video is made up of four or more clips stitched together, each created from a start image and an end image. The lamé bikini is enough to prevent the images from being rejected by models for nudity, even if, as you can see, the same models then have no problem showing off the dancer's rear end. So things work in the strange world of censorship algorithms. The first model I tested is KLING 2.5 Turbo. Here's the 60-second video I created, split into two parts due to the 50MB upload limit. As you can see, Kling's model is very attentive to the fluidity of movement and the realism of the flesh. However, it is not particularly reliable for the character's coherence, since it changes both the body features, making it less graceful and heavier, and the face, which at times seems to belong to another woman. These defects are due to the functioning of the model, since the start and end images of the various clips are consistent with each other in showing the same dancer. Then I tested, right here on BudgetPixel, SeeDance 1.5 model, and it was a pleasant surprise, first of all because, by deactivating the sound that I don't need, the cost is 50%, that is, 30 credits per second (at 720 p). Here is the 40-second video I got. The fluidity of the movements and the choreographic details invented by the model are remarkable, and the coherence of the dancer's body is perfect throughout the video. The consistency of the facial features is acceptable, although not perfect, but this is, in my opinion, the only flaw of this model. As I said, I was very satisfied with this model, even if I still have to test it with NSFW images (and I'm not sure it will accept them). Finally, here is the 50-second video (split into two parts) I created with the model I regularly use for NSFW videos. The model is called Real Motion 2.6 and is available in various versions of which I use two: Real Motion 2.6 Pro for 5-second clips and Real Motion 2.6 Remix for 10-second clips. I'm not 100% sure, but I think this model is unique to the Digen.ai platform. I think that for the fluidity of the movements and the coherence of the character's features this model has nothing to envy to other more widespread and well-known models. However, I must add, for anyone who might be interested, that on that platform there are always many users, and the waiting times are long, precisely because it allows the creation of NSFW videos. I hope this comparison can be useful. Maybe I'll do some testing with other video models when I have time. Take care! 😎