← sources

What is "the best" video model?

source: reddit (reddit:StableDiffusion)published 2026-09-02 15:38ingested 2026-09-03 02:40

retained text

I've tested a couple of models so far, and here's my findings:

My input prompt is something similar to "Create a mid 20s (nationality) ballerina slowly dancing in a judged performance"

LTX-2.5 nails the narrative and the nationality, but nationalities (as I've mentioned previously) all tend to blend unless I'm ULTRA descriptive about what a "french look" equates to, which is roughly 3 paragraphs in length.  LTX-2.5 ends up with tearing, facial and feature deformities, and on a couple of occasions has rendered a ballerina with missing legs.  It also has camera drift, which is a known problem with LTX-2.5

MiniMax-H3 on the other hand, nails it, even with the smaller prompt.  The prompt adherence in MiniMax-H3 is wonderful.

Wan 2.2 was a complete disaster.  Rendering problems, tearing problems, and when the ballerina would do turns, her head would remain in place while her body did the turn (which was funny, and also terrifying to watch.)

I've heard Cosmos3 can handle the movement, but can't handle rendering people.

Has anyone found an open weight model that can handle the "ultimate trifecta" - generate a person in the correct nationality parameters, someone who is more or less feature-accurate in generation, and won't spontaneously explode when performing a pirouette?  (I got so frustrated with a model generator once, that I did this, and to my surprise, it worked out well!)

document fetched 2026-09-03 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1w5dvwj/what_is_the_best_video_model