Legato, a new training method for action‑chunking flow policies, teaches robots to generate native continuous motions, eliminating hesitation and improving task speed and trajectory smoothness across five real‑world manipulation tasks, as demonstrated in the RSS 2026 paper.
Kimi-Audio, a general‑purpose audio foundation model from Moonshot AI, integrates ASR, audio QA, automatic audio captioning, emotion classification and end‑to‑end speech dialogue within a single framework, detailing its mixed‑audio input, MiMo‑Transformer core, efficient synthesis pipeline, architectural strengths, limitations, and suitable application scenarios.
GLM-4-Voice is an end-to-end Chinese-English speech dialogue model that aligns discrete speech tokens with GLM-4-9B, uses VQ-based tokenization at 12.5 token/s, supports emotion, tone, speed and dialect control, and offers streaming inference with low latency, while detailing its architecture, advantages, limitations and suitable use cases.
The OmniXtreme architecture introduces a two‑stage flow‑matching and actuation‑aware post‑training framework that enables humanoid robots to reliably execute high‑dynamic, extreme motions in the real world by overcoming simulation scalability limits and physical hardware constraints.
PropMolFlow, a new flow‑matching model introduced by researchers from the University of Florida and NYU, generates property‑guided molecules up to ten times faster than prior SOTA methods while preserving chemical validity and achieving superior performance on benchmarks such as QM9.
Flow‑GRPO introduces online reinforcement learning into flow‑matching models by converting deterministic ODE sampling to stochastic SDE sampling and reducing denoising steps, raising SD‑3.5‑Medium's GenEval accuracy from 63% to 95%—surpassing GPT‑4o—and demonstrating strong gains in complex composition, text rendering, and human‑preference alignment across multiple generative tasks.
This article explains the over‑optimization problem in GRPO‑based flow models, analyzes why importance‑ratio clipping fails, and introduces GRPO‑Guard with RatioNorm and cross‑step gradient balancing, showing through extensive experiments that it stabilizes training and improves image quality across multiple diffusion backbones and tasks.
This article walks through the full implementation of a flow‑matching generative model in PyTorch, covering dataset creation, a small MLP that learns a time‑dependent velocity field, the flow‑matching loss, training loop, ODE‑based sampling, visualisation of the learned vector field, and a discussion of the method's limitations and possible extensions.
This article walks through the theory behind flow‑matching generative models, contrasting them with diffusion models, detailing the velocity‑field formulation, training objective, and sampling procedure, and includes visual illustrations of the core concepts.
This technical article provides a comprehensive comparison of diffusion models and flow matching, covering their intuitive explanations, underlying mathematics, training objectives, sampling efficiency, theoretical guarantees, practical examples, and code implementations to illustrate how each generative approach works.
Apple’s newly released SimpleFold model leverages flow‑matching and a pure Transformer architecture to eliminate costly MSA and triangular updates, achieving performance comparable to AlphaFold2 and RoseTTAFold2 on CAMEO22 and CASP14 benchmarks while dramatically reducing computational requirements, and a step‑by‑step tutorial lets users run it on HyperAI’s platform.
RFdiffusion2 eliminates residue enumeration and sequence indexing by using flow matching and stochastic centering, enabling atom‑level active‑site design; it succeeds on all 41 benchmark cases (100% success vs. 39% for RFdiffusion1) and is available through a one‑click tutorial on the HyperAI platform.
The article presents MeanFlow, a novel one‑step generative modeling framework that replaces instantaneous velocity with an average‑velocity field, achieving a record‑low FID of 3.43 on ImageNet 256×256 with a single function evaluation and outperforming both prior single‑step and multi‑step diffusion models.
Flow Matching reformulates generative modeling by learning a time‑dependent vector field that deterministically transports Gaussian noise to data, using a neural network trained with an analytically derived L2 loss, yielding simpler training, faster convergence, and deterministic sampling that matches or exceeds diffusion model quality.
This article explains how recent MIT research uses stochastic differential equations to model diffusion and flow processes, defines training objectives, explores conditional guidance, compares U‑Net and diffusion transformers, addresses memory challenges with latent diffusion, and surveys applications ranging from robotics to protein design.
FlashVideo introduces a two‑stage framework that first generates low‑resolution videos with strong prompt fidelity and then uses flow‑matching ODE trajectories to upscale to 4K quality in just four function evaluations, achieving top VBench‑Long scores while cutting generation time by up to five‑fold.
