Diffusion models can create high-performing neural network parameters by coupling an autoencoder with a diffusion model to generate latent representations and decode them into new parameter subsets; these models match or exceed trained networks with low extra cost and do not memorize training data.
A survey of reinforcement learning–driven diffusion models, detailing how RL post-training reshapes diffusion processes for goal-oriented generation across domains, outlining methods, experiments, bottlenecks, and a roadmap for robust, interpretable RL-enabled diffusion systems.
Rolling Diffusion introduces a sliding window denoising approach that assigns more noise to later frames, capturing increasing future uncertainty; it outperforms standard diffusion on complex temporal data, shown on Kinetics-600 video prediction and chaotic fluid dynamics forecasting.
The article explains how masked diffusion language models refine several text positions at once, unlike GPT-style models, which generate one token at a time. It describes masking as a way to adapt diffusion to discrete text and points to remaining challenges, including token dependencies and key-value caching.
Diffusion Policy presents a reproducible robotics learning framework with state- and vision-based environments, complete experiment artifacts, and instructions for setting up the environment and reproducing results. It includes data, configs, logs, and guidance for scaling experiments using Ray and optional cloud infrastructure.
ControlNet adds extra-condition guidance to diffusion models, enabling precise, spatially consistent control over Stable Diffusion image generation.
The paper examines speculative diffusion for LLMs, showing how jointly tuning diffusion effort and acceptance leniency can extend the speed-quality tradeoff frontier beyond single-step methods, by enabling lossy yet faster draft sampling while preserving the verifier distribution properties in traditional speculative decoding.
This document provides a quick, code-style guide to run the Diffusers diffusion model lodestones/Kromoss, including installation, imports, model loading, and image generation from a given prompt.
MultiDiffusion offers a way to generate controllable, high-quality images from diffusion models without extra training, by optimizing a shared set of parameters across multiple diffusion processes to follow user controls such as aspect ratio and spatial guides.
ORL updates diffusion-language-model training by replaying current-model residuals and using entropy contraction to refine predictions, enabling scalable, efficient parallel decoding and higher throughput across multiple models.