New open source 360° video diffusion model (CubeComposer) – would love to see this implemented in ComfyUI

Tencent ARC has launched CubeComposer, a significant open-source contribution to the generative AI space focused on immersive media. The model specializes in creating 360° panoramic video using a novel cubemap diffusion technique. Instead of generating a single, distorted equirectangular projection, CubeComposer creates a video by independently generating and then composing the six faces of a cube (front, back, left, right, top, bottom) with spatio-temporal consistency. This approach, detailed on the project's Hugging Face page, allows for higher-resolution outputs and more coherent video generation across the entire panoramic field of view, which is critical for believable VR experiences.

The release includes full model weights and code, positioning it as a research pipeline rather than a consumer-ready tool. Currently, it runs as a standalone system, but its open-source nature has sparked immediate interest within communities like ComfyUI, a popular node-based interface for AI workflows. Developers are already discussing the potential for creating custom ComfyUI nodes to integrate CubeComposer, which would allow users to convert generated perspective frames into 360° cubemaps and slot the model into existing video generation pipelines. This could dramatically lower the barrier to creating high-quality, dynamic environments for virtual reality, architectural visualization, and immersive storytelling.