LightX2V

Step-distilled and CFG-distilled acceleration of Wan2.1-T2V-14B by the LightX2V team — four-step inference with classifier-free guidance removed, via Self-Forcing bidirectional distillation. Nineteen cards here carry a recipe, down to a 12 GB floor, and what changes across them is the quantisation path: int8 offload on Ampere, native fp8 on Ada and Blackwell, and a GGUF plus 4-step LoRA route on ROCm for the two Radeon boards. Apache-2.0.