Actually, the Pyramid Flow 768p miniflux model dropped 15 days ago. It is far from perfect. I have generated hundreds of test videos (IMG2VID) because most iterations are imperfect, and it takes trial and error to get an exceptional 10 second clip. If you generate sterotypical images, it can work well first time. As soon as the scene in the source image is complicated, or when human figures raise their hands, the morphs can become extremely ugly and a total waste of time. Finding that right balance between how much the characters depicted may move, and the quality of those motions is quite hard to achieve for complex scenes. I find subtle movements or balance between motion and stillness works best. I still use it daily as I believe there will be corrective apps released in the future to fix flawed videos. On a 3090 it takes between 600-700 seconds per video. I have also been testing LTX video and it is very fast, but I do not like the quality versus the more stylized videos produced in Pyramid Flow.
Actually, the Pyramid Flow 768p miniflux model dropped 15 days ago. It is far from perfect. I have generated hundreds of test videos (IMG2VID) because most iterations are imperfect, and it takes trial and error to get an exceptional 10 second clip. If you generate sterotypical images, it can work well first time. As soon as the scene in the source image is complicated, or when human figures raise their hands, the morphs can become extremely ugly and a total waste of time. Finding that right balance between how much the characters depicted may move, and the quality of those motions is quite hard to achieve for complex scenes. I find subtle movements or balance between motion and stillness works best. I still use it daily as I believe there will be corrective apps released in the future to fix flawed videos. On a 3090 it takes between 600-700 seconds per video. I have also been testing LTX video and it is very fast, but I do not like the quality versus the more stylized videos produced in Pyramid Flow.
Thanks for sharing your experience, it's definitely a work in progress!
Super interesting! I'm still really new to this, accroding to you which generative model/framework are used by sites like hailuo or pixverse?