KlingAIResearch / DecMem

What is DecMem?

DecMem is a video generation system that produces minute-long consistent videos by combining a diffusion model backbone with a novel decoupled memory architecture. The core insight is that standard attention mechanisms struggle with long sequences, so DecMem splits memory into two components: a sparse global memory for efficient retrieval across the full history, and an anchored local memory that keeps short-term transitions smooth. The system trains on gameplay footage paired with action and pose data, enabling controllable long video generation where camera movements and character actions stay coherent across extended clips. It's built in Python on top of the Wan2.1 video diffusion model and uses custom CUDA kernels for the sparse attention operations.

Why is it gaining traction?

The hook is simple: open-source minute-long video generation with consistency is rare. Most diffusion-based video models handle short clips; DecMem specifically targets the long-horizon problem that causes flickering, drift, and broken physics in extended generations. The decoupled memory design is the technical differentiator--by separating what gets attended to globally versus locally, it avoids the quadratic cost explosion of dense attention while maintaining fidelity. The research team (Kling/Kuaishou) has a track record in video generation, which lends credibility even to this early release.

Who should use this?

This is for researchers and engineers working on world simulation, game content generation, or robotics planning where you need extended video sequences that maintain visual and physical consistency. If you're evaluating video generation for training data synthesis or creative tools, DecMem addresses a real gap that standard short-clip models leave unsolved. Be warned: the full pipeline requires H100/H200/H800 GPUs with custom CUDA kernels, so plan your hardware accordingly.

Verdict

DecMem solves a legitimate problem and ships with working inference code, but with 10 stars and a 1.0% credibility score, this is clearly early-stage research code rather than production-ready tooling. The setup is involved--multi-stage training, custom kernel compilation, specific hardware requirements--and documentation, while present, assumes familiarity with video diffusion systems. Worth exploring if you're pushing the boundaries of long video generation, but approach with realistic expectations about the maturity level.