Open yuanxion opened 1 year ago
Hi, all. I wonder whether it is possible to compute a 3D consistency metric between different frames? Maybe this kind of metric needs us to propose in some special format.
Maybe one way is trying to use these frames and nerf to generate a 3D scene. And analysis whether the 3D scene is good?
3D reconstruction tasks are typically evaluated using reference-based metrics like Chamfer Distance (https://zhuanlan.zhihu.com/p/629086843):
From DreamFusion, they evaluate the CLIP R-Precision: https://arxiv.org/pdf/2209.14988.pdf.
I have a consideration: maybe we shouldn't choose Zero to be the base model for our second stage. The advantage of Zero is that it doesn't need any training. If we break this constraint to enhance motion coherence by finetuning it in some way, it may be meaningless in terms of innovation because there have been some models solving the motion coherence problem by finetuning a text-to-image model with unlabeled video datasets. If we indeed want to solve the coherence problem by 3D latent constraints, maybe we should choose some models which are designed to be finetuned to be our base model or comparison ones. However, for the accelerating task in our first stage, Zero is a good choice.
We can also generate 3D motion, if we only focus on human video.
paper : Generating Diverse and Natural 3D Human Motions from Text
I have a consideration: maybe we shouldn't choose Zero to be the base model for our second stage. The advantage of Zero is that it doesn't need any training. If we break this constraint to enhance motion coherence by finetuning it in some way, it may be meaningless in terms of innovation because there have been some models solving the motion coherence problem by finetuning a text-to-image model with unlabeled video datasets. If we indeed want to solve the coherence problem by 3D latent constraints, maybe we should choose some models which are designed to be finetuned to be our base model or comparison ones. However, for the accelerating task in our first stage, Zero is a good choice.
Yes, agree. We have powerful GPUs with Server now, so it is ok to change and train whichever models.
https://huggingface.co/blog/text-to-video In this web page, there are several model demos including Zero for us to try freely.
I have tried several T2V demos on huggingface and this one seems the best up to now. https://huggingface.co/spaces/NeuralInternet/Text-to-Video_Playground
Read papers to get more (innovation) ideas. Share learning of Diffusion model/ControlNet.
Idea: Make the video generation more smoothing.