Back to Project
W

Wave Forcing

Project Nov 3, 2026

A research framework for continuous video generation with released 1.3B few-step checkpoints and a multi-GPU WaveRT streaming runtime.

4 sources attached
Author-supplied generated video frames compare a reading scene across Causal Forcing, Rolling Forcing and four-step and five-step Wave Forcing from zero to 95 seconds.
Generated samples selected by the Wave Forcing authors for their long-video comparison. These are research-project examples, not outputs from the community Space or an independent AITrending test. Source: Wave Forcing authors / DENG Lab @ SJTU
multimodalself-hosted

What happened lately

Wave Forcing timeline

Wave Forcing combines chunk-wise causal video generation, a three-stage training procedure and a runtime called WaveRT. Its paper was submitted on October 31, 2026. The authors release four-step and five-step WaveForcing-1.3B transformer checkpoints under Apache-2.0; base Wan2.1 text and decoding components are downloaded separately.

The reported 134.8 end-to-end fps result uses eight H200 GPUs at 832 by 480 resolution, excluding model loading and file encoding. It does not describe the performance of Akhaliq’s separate single-GPU community Space. The repository distinguishes the Apache-2.0 WaveRT code from training code under a separate academic-use license. Its 14B runtime experiments use randomly initialized transformer weights and should not be read as quality results for a released 14B model.