Runway Research | Research Papers

This post introduces several research papers advancing diffusion models. 'Autoregressive-to-Diffusion Vision Language Models' adapts autoregressive models for parallel diffusion decoding to unlock speed-quality trade-offs without retraining. 'Dual-Process Image Generation' proposes a distillation scheme where a VLM rates generated images and backpropagates gradients to update image generators, enabling new control tasks. 'StochasticSplats' combines 3D Gaussian splatting with stochastic rasterization for sorting-free rendering, achieving faster rendering times. 'Progressive Prompt Detailing' introduces SCoPE to refine prompts progressively for better text-to-image alignment. 'What's in a Latent?' leverages diffusion latent space for domain generalization by discovering latent domain structures. 'Concept Steerers' uses k-sparse autoencoders for controllable generation and concept manipulation. 'Revelio' interprets and leverages semantic information in diffusion models using k-sparse autoencoders. 'Mitigating stereotypical biases' proposes finetuning on synthetic data to improve fairness. 'Structure and Content-Guided Video Synthesis' presents a diffusion model for video editing that retains structure while editing content, using monocular depth estimates and a novel guidance method for temporal consistency.