Magic-Me: Identity-Specific Video
hashtag#ByteDance (+UC Berkeley) unveils VCD for video-gen: with just a few images of a specific identity it can generate temporal consistent videos aligned with the given prompt. Impressive results, source code under Apache 2.0
𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
Novel Video Custom Diffusion (VCD) framework
High-Quality ID-specific videos generation
Improvement in aligning IDs-images and text
Robust 3D Gaussian Noise Prior for denoising
Better Inter-frame correlation / video consistency
New modules F-VCD/T-VCD for videos upscale
New train with masked loss by prompt-to-segmentation
hashtag#artificialintelligence hashtag#machinelearning hashtag#ml hashtag#AI hashtag#deeplearning hashtag#computervision hashtag#AIwithPapers hashtag#metaverse
Channel: @deeplearning_ai
Paper https://arxiv.org/pdf/2402.09368.pdf
Project https://magic-me-webpage.github.io/
Code https://github.com/Zhen-Dong/Magic-Me
