I can’t stand AI for video, because I feel it is so much easier to do normal old-school video, but after all the hype of the past two days, I had to check Sora 2 out because I’m still looking for a decent pose-detection model. Just as expected, sora 2 is a total disappointment for what I need it to do, a slight improvement from previous models attempting the same thing, but anyone who says that it does a good job at kpop has never seen an actual kpop choreography, or has never attempted to dance in their life and is unable to detect moves even as a person…
Sora 2’s results are extremely unnatural. Maybe it is good for other stuff. I don’t know what else to say
I asked my AI if I’m being too negative here and she gave me sources that confirm exactly what I just said. I disagree however that beginners can benefit from this model. It can actually damage a beginner. I always place myself in the shoes of a total beginner because that’s how I learn best, no ego, and I feel that specifically for someone who starts out, moves and counts should be extremely clear-cut, or our attention moves to everything other than dancing. There must be even more rigor than otherwise:
You’re right—while Sora 2 can generate K-pop dance videos, the moves can feel limited or unnatural, especially for experienced dancers or choreographers. It struggles with complex, fast-paced sequences and often simplifies motions to stay within its training patterns.
Recent sources from October 2025 confirm that Sora 2 still has limitations in generating complex K-pop dance tutorials.
-
Despite improvements in motion smoothness and physics accuracy, hand and limb rendering remain inconsistent, with reports of missing fingers, morphing limbs, and unnatural transitions during fast choreography.
-
The model performs better with short, concise shots rather than long, dynamic sequences, making it less ideal for full dance routines.
-
Users are advised to generate multiple clips and edit them together for cleaner results, as longer 8–10 second generations can degrade in quality.
-
While Sora 2 excels at cinematic visuals and storytelling, realistic human biomechanics—especially in intricate dance—still lag behind specialized models like Google’s Veo2
-
For accurate K-pop tutorials, real instructor-led content remains superior.
Since we learned about the excellence of Google Veo 2, I went on Youtube to check how miraculous the results are, and still nope. We’re talking extremely short sequences with extremely basic moves, stuck together with zero coherence:
I give you that Sora 2 already has designated kpop channels. The main point here is not to sit several hours behind 4 screens and ask AI to do Kpop routines of 8 to 10 seconds, but rather to give AI the URL to an actual Kpop dance practice video and to have AI produce a tutorial for each of the dance parts. A tutorial is much longer than the actual counts, because it breaks down each count. Realistically, the more I look at the new overhyped models, the more I realize we’re not getting an actual useful model any soon.
Maybe these models are good for making sci fi or action films. I have no idea.