Perflow-Shuai/LongLive-LoRA-nonAR-DMD-RGS3-iter1500
Text-to-Video • Updated • 14 •
Wan2.2-5B dual-LoRA: few-step A (RGS3, iter1500) + CFG (iter250). Four steps; few-step weight 1; inference CFG 1.
Note Few-step LoRA A: non-AR DMD, RGS3, iter1500, rank/alpha 128/128. Pair with the CFG LoRA in this collection. Keep few-step weight at 1.0; use 4 denoising steps and inference CFG 1.0.
Note CFG distillation LoRA: CFG5-to-CFG1, iter250, rank/alpha 64/64. Pair with few-step A for 4-step inference. Sweep CFG LoRA weight 0/0.5/1/2 (0 = few-step only); this scales the adapter residual, not inference CFG.