Helios

Why we included this project

Helios targets teams that want long-form synthetic video but can't afford the usual multi-minute render queues. It's a 14B diffusion model that produces minute-scale clips at roughly 19.5 FPS on a single H100, fast enough for live, interactive generation rather than batch work. The repo ships base, mid, and distilled checkpoints along with training and inference code, and it handles text-to-video, image-to-video, and video-to-video tasks, so one setup covers a fair range of production jobs. The speed comes without common long-video crutches like keyframe sampling or KV-cache; instead it leans on architectural and training optimizations that also bring memory use down. Because the project integrates with Diffusers, SGLang, and vLLM and includes a Gradio demo, you can judge its coherence and quality yourself before committing to a workflow.