Hongyang-Du / VideoGPA

What is VideoGPA?

VideoGPA is a Python framework that enhances 3D consistency in video diffusion models by automatically aligning generative videos with dense structural preferences from geometry foundation models. It eliminates deformation and spatial drift without human annotations, using DPO to distill these priors into LoRA adapters for models like CogVideoX. Developers get plug-and-play inference scripts for text-to-video or image-to-video generation, plus full training pipelines.

Why is it gaining traction?

It stands out by delivering temporally stable videos from standard diffusion models via self-supervised geometry alignment, skipping costly 3D supervision. Simple CLI scripts download checkpoints and run inference on 5GB VRAM, while DPO training supports preference-based fine-tuning on scored video pairs. This makes consistent generative video accessible without retraining from scratch.

Who should use this?

Video generation engineers tuning CogVideoX for apps needing rigid 3D motion, like AR/VR content or simulation visuals. Researchers evaluating diffusion consistency metrics on custom datasets. Teams prototyping image-to-video extensions where drift ruins usability.

Verdict

Worth forking for video diffusion experiments—MIT-licensed with arXiv paper and project page—but at 27 stars and 1.0% credibility, it's early research code needing more tests and examples before production. Start with inference scripts to test fit.

(198 words)