A video generation AI model that runs on consumer GPUs has come out, and the quality is far better than existing video generation AI. It does not take as long as expected either, and above all 15-second generation works very well.
Even if you just give the prompt in Korean it understands it well and produces results. Before, I always had to redo them in English, but for now it seems to work well. Of course, English prompts may be needed to bring out the detail of the video, but at the level of merely dipping a toe in as I have, it is fine.
MiniMaxAI/MiniMax-H3
https://huggingface.co/MiniMaxAI/MiniMax-H3
The model I used is minimax_h3_fl2va_pruned_int8_convrot.safetensors, but you will need to download the VAE file as well. If you are not sure, just ask an AI to do it for you. ㅎ_______ㅎ;;
For reference, I ran it with ComfyUI. I am running an RTX 3090 24GB. And I made an mcp to connect to the Cursor agent, so this is a video I made by roughly throwing prompts at it through mcp from the desktop Cursor. This is more convenient and better than opening ComfyUI. Try it yourself. It is easy and convenient.
▶ Original source: https://huggingface.co/MiniMaxAI/MiniMax-H3
