What is raven?

RAVEN is the code release for a research paper from Imperial College London that tackles a specific problem in AI video generation: keeping quality steady when a model streams video in real time, chunk by chunk, instead of generating the whole clip at once. Streaming models predict each new chunk from the video they already generated, but during training they normally only see clean, ideal past chunks, while at actual run time those past chunks are a bit noisy and imperfect. That mismatch causes quality to drift the longer the video runs. RAVEN closes that gap by training the model on a mixed sequence of clean and noisy past states that better matches what it will actually see when it is generating video for real. The project also introduces a second technique called Consistency-model Group Relative Policy Optimization, or CM-GRPO, which applies reinforcement learning directly to the video generation process itself. Reinforcement learning is a method where a model improves by getting rewarded for good outputs and penalized for bad ones. The authors report that RAVEN outperforms other recent streaming video methods on quality, how well a video matches its text description, and how much motion it contains, and that CM-GRPO adds further improvement on top of RAVEN. This is a research codebase, not a polished app. Setup involves creating a specific Conda and Python environment, installing pinned versions of PyTorch, Transformers, and Diffusers, and building custom attention libraries from source, all targeting a high end Nvidia GPU. Users need to separately download the base model checkpoint and the released RAVEN weights before running anything. The repository includes training scripts, sampling scripts for generating video from prompts, and an evaluation pipeline using the VBench benchmark suite to score results, along with several other published methods vendored in for side by side comparison. This project is aimed at AI researchers and practitioners working on video generation who want to reproduce the paper's results, experiment with the training method, or benchmark their own models against RAVEN and other streaming video baselines. It is written in Python and requires substantial GPU hardware and machine learning experience to use.