ModelTC/LightX2V

Supports multi-modal generation tasks (T2V, I2V, T2I, image-editing) with aggressive optimization techniques including CFG/Ulysses parallelism, FP8/NVFP4 quantization, step distillation, and memory-efficient offloading. Deploys across diverse hardware accelerators (NVIDIA, AMD, Intel, Ascend, Cambricon, MThreads) and integrates with HuggingFace models via LoRA checkpoints trained with reinforcement learning. Architecture prioritizes inference throughput through disaggregated deployment patterns and model compression, achieving 25-42× speedups on production workloads.

2,078 stars. Actively maintained with 50 commits in the last 30 days.