HUST-Tencent SpecWM Turns World-Model Verification Into Refinement, Hits 1.63x Speedup on WBench Navigation
Summary
Huazhong University of Science and Technology and Tencent authors posted SpecWM, a training-free inference framework for streaming video world models that verifies draft chunks using the target's latent velocity field and reuses the verification signal as a correction. Two thresholds select Accept, Refine, or Reject, with one-step refinement improving perceptual quality and three-step refinement tightening camera-following and trajectory consistency. Reported 1.63x speedup over the target on WBench Navigation.
Originally reported by huggingface.co
Read the original article →Original headline: HUST-Tencent SpecWM Turns World-Model Verification Into Refinement, Hits 1.63x Speedup on WBench Navigation