Diffusion Model Research
Runway continues to advance its diffusion model research, focusing on improving efficiency, controllability, and interpretability. Recent work includes adapting autoregressive models for parallel diffusion decoding (A2D) to leverage existing pretrained models for speed-quality trade-offs, developing dual-process distillation schemes to teach feed-forward image generators new tasks from vision-language models, exploring stochastic rasterization for 3D Gaussian splatting to remove sorting artifacts, and developing methods for progressive prompt detailing and latent space manipulation for domain generalization and concept steering. Research also addresses mitigating stereotypical biases and improving structure/content-guided video synthesis.
Timeline