What is echo-infinity?

Echo-Infinity is a research project from a team at Joy Future Academy (JD) that tackles a specific challenge in AI video generation: how do you keep generating video continuously for very long durations (minutes, hours, or even a full 24 hours) without the memory cost exploding as the video gets longer? The approach the paper introduces is a learnable memory module. Rather than storing every frame of video history in full, the system learns to filter, compress, and abstract that history into a compact representation that stays the same size regardless of how long the video has been running. This makes it possible to generate video at real-time speed for arbitrary lengths, which is described in the demos as producing a continuous 24-hour video in a single pass. The project builds on top of existing open-source video generation models (specifically Wan2.1, which handles the base video generation, and techniques from Causal-Forcing and Self-Forcing for training efficiency). Training happens in two stages: first a short initialization stage, then a longer fine-tuning stage for extended video. Both stages require significant GPU hardware, with the default training configuration using 4 machines each with 8 GPUs. For users who just want to run inference, the README provides command-line examples covering short clips (5 seconds), medium clips (30 or 240 seconds), interactive mode where you switch between different text prompts mid-video, and the full hour-level and 24-hour streaming modes. Pretrained checkpoints are hosted on Hugging Face. The codebase is Python with a Conda environment and standard pip installation. This is an academic research release accompanying a paper on arXiv (arXiv:2606.04527) rather than a production-ready application.