Search
From Single-Step Image Generation to Streaming Video: Tsinghua Team Explores 'Sample Distribution Training'
2 day ago /
Read about 0 minute
Author：小编
Researchers from institutions including Tsinghua University have proposed MGFlow and Elastic Forcing technologies, achieving SOTA-level single-step image generation. MGFlow simplifies the image generation process from four-step iteration to a single step through unified distributed training, significantly improving efficiency. Elastic Forcing addresses the error accumulation problem in autoregressive video generation through a zero-cost parasitic GAN design. By combining these two technologies, a streaming video model with 14 billion parameters can be post-trained using just 8 GPUs.
-
C114 Communication Network -
Communication Home
7 X 24 Track global technological trends
