Decart Launches LSD-v2 – New Realtime Video Model (twitter.com)

Decart has announced LSD‑v2, a next‑generation real‑time video model intended for live video synthesis and editing. Framed as an evolution of latent‑space diffusion approaches, Decart positions LSD‑v2 to deliver low‑latency, temporally coherent frame synthesis suitable for streaming and interactive applications. The company emphasizes practical deployment: real‑time inference, API/SDK access for integration into live production pipelines, and support for multimodal conditioning (text prompts, reference frames, and motion cues) to steer output. Technically, LSD‑v2 reportedly moves heavy computation into a compact latent space and combines temporal attention or cross‑frame modules to preserve motion consistency while reducing per‑frame compute. Decart highlights system‑level optimizations — model distillation/quantization, tiled or streaming decoding, and efficient memory management — that enable high throughput on GPU/edge hardware. For the AI/ML community this matters because it lowers the barrier to real‑time generative video use cases (live VFX, AR/VR, telepresence) while sharpening operational and safety questions: potential for convincing deepfakes underscores the need for provenance, watermarking, and detection tools. Researchers will be watching for released benchmarks, architecture details, and open tooling to validate LSD‑v2’s latency, quality, and robustness claims.