本地AI视频模型对比展示

在Minimax H3中采用相同风格的提示词，这是一个在我个人电脑上本地运行的AI视频模型，分辨率仅为720p，生成时长为15秒。 我对提示词做了一些“自由发挥”来测试一些物理效果…… 现在，与Seedance 2.5或许多顶级的闭源模型相比，本地运行、经过量化和速度优化的Minimax H3在遵循提示词方面肯定较弱，但令人惊叹的是，我们拥有一个如此强大的开放权重模型，而不是某个闭源云模型。 它遗漏了几件事，比如狗叫声、手推车变成自行车，以及她在开头丢下袋子，还有一些重影问题，但已经相当不错了。 另外……我提到过它没有审查吗？

Same style prompt in Minimax H3, an AI video model running locally on my PC at only 720p resolution, 15-second generation. I took some "liberties" with the prompt to test some physics... Now, compared to Seedance 2.5 or many of the top closed models, Minimax H3 running locally, quantized and optimized for speed, is definitely weaker at following prompts, but it's just amazing that we have an open-weight model that's this powerful and not some closed-source cloud model. It missed several things, like the barking, the cart turning into a bicycle, and she drops the bags at the beginning, along with some ghosting issues, but it's pretty darn good. Also... did I mention it's uncensored?