how much better is the cloud version of h3?
The cloud i hear does true 2k resolution where as this watered down version does 768p. Plus a smaller sized transformer 33b. Is there a big quality jump you guys are noticing for the cloud vs open source? 4090 gpu 15 seconds 0.6 megapixels image to video took 15 minutes on my system. 05 megapixels 11 second duration takes 8 minutes. For those with a 3090. I have noticed a 3090 takes almost twice as long as a 4090 gpu.
There are 2 models, look at their architecture explanation. The one here is the 768p one, which should be the same as the cloud version if you generate only in 768p.
There are 3 models: the video generation model, H3-Base, which is the open sourced one, and it can generate videos up to 2K. They haven't open-sourced the other 2 models, H3-Context-IR (which I think is for prompt expansion) and H3-Regenerate-2K, which is essentially an upscaler from 768P to 2K.
They said they will open source H3-Regenerate-2K here: https://huggingface.co/MiniMaxAI/MiniMax-H3/discussions/39
They talk about the H3-Context-IR here in Chinese: https://huggingface.co/MiniMaxAI/MiniMax-H3/discussions/14