40s generation time for 10s vid on a 5090 using custom runtime (ltx 2.3) (closed project, will open source soon)

An independent developer has achieved a major performance breakthrough in AI video generation by creating a custom Rust-based inference engine tailored specifically for the LTX-2.3 architecture. The system generates 10-second, 1920x1088 resolution videos at 24 fps in under 40 seconds on an RTX 5090 GPU—a significant improvement over standard implementations. The key innovation involves bypassing Python entirely on the "hot paths" (the most computationally intensive parts of the generation pipeline) and implementing aggressive memory management techniques in Rust.

The engine is explicitly designed for LTX's specific 3D attention blocks, allowing the developer to hardcode the computational graph and eliminate dynamic dispatch overhead. A custom 3D latent memory pool in Rust perfectly matches LTX's tensor shapes, resulting in zero VRAM fragmentation and no allocation overhead during the step loop. Additional optimizations include zero-copy loading of safetensors directly to the GPU. The current setup uses the LTX-2.3-22B model with a distilled LoRA and spatial upscaler, running 15 steps in stage 1 plus 3 refinement steps in stage 2.

While the project is currently closed-source, the developer plans to open-source it soon and will publish a technical breakdown explaining the architecture. This work demonstrates how specialized, low-level optimization can dramatically improve the performance of cutting-edge generative AI models, potentially making high-quality video generation more accessible and practical for real-time applications.