diffusion

VideoMDM trains 3D human motion priors from 2D poses using a diffusion framework with 2D reprojection loss and 3D motion regularizers, achieving near-3D supervised performance without requiring 3D ground truth.

Google released DiffusionGemma, an open-weight text generation model (26B parameters, 4B active) under Apache 2 license, demonstrating high inference speeds via NVIDIA's NIM cloud API.

DiffusionGemma, a 26B MoE model based on Gemma 4, achieves over 1000 tokens per second using diffusion for text generation in 256-token blocks, fitting in 18GB VRAM with quantization, released under Apache 2.0.

New paper shows how to optimize flow matching actors for reinforcement learning by approximating the Jacobian of the flow denoising process with the identity matrix, making training feasible.

Google DeepMind releases DiffusionGemma, a 26B-parameter Mixture-of-Experts model that uses discrete diffusion for faster text generation, supporting multimodal inputs and a 256K token context.

An essay explaining the physical constraints on cell size, focusing on surface-area-to-volume ratio and diffusion limits that make cells small.

SwiftVR is a real-time one-step generative video restoration framework that achieves high frame rates on consumer GPUs using efficient attention mechanisms and a lightweight restoration-aware autoencoder.

MaskAlign proposes a token-subset representation alignment method that improves diffusion transformer training by reducing reliance on complete token sets and maintaining stable alignment under perturbations.

Introduces SelfBootTok, a self-bootstrapped tokenization method that separates global and local information, reducing generator computation by ~40% and achieving a new state-of-the-art gFID of 1.56 with only 64 tokens.

This paper introduces the Eisbach log-barrier, a parameter-free weight derived from the entropy of DiT output's spatial energy distribution, which when applied to LoRA fine-tuning of Stable Audio 3 improves musical diversity and thematic development without causing mode collapse.

Introduces Discrete-WAM, a unified discrete latent vision-action world policy that enables compositional causal reasoning and counterfactual reasoning in autonomous driving through aligned discrete tokens and a shared discrete diffusion framework.

PrismML releases Bonsai Image 4B, a family of compact image generation models using 1-bit and ternary weights, enabling high-quality diffusion inference on local devices like laptops and iPhones with significantly reduced memory footprint.

StressDream enhances video world models by steering diffusion-based imaginations toward high-impact yet plausible outcomes through optimized noise initialization with semantic and plausibility objectives, enabling robust policy evaluation and improvement.

Sakana AI presents DiffusionBlocks, a method that trains neural networks block-wise by interpreting forward passes as diffusion denoising, significantly reducing memory requirements compared to traditional end-to-end backpropagation.

NAVA proposes a native audio-visual alignment framework for joint audio-video generation using an Align-then-Fuse MMDiT architecture, achieving improved synchronization and controllability with 6.3B parameters.

PrismML released Bonsai Image 4B models in binary and ternary quantized versions, enabling text-to-image generation to run locally in a browser via WebGPU with only 3GB size, under Apache-2.0 license.

The author details attempts to locally train a Qwen 3.6 27B autoregressive-to-diffusion model on an Nvidia 5090 GPU using qlora and modifications from open-dllm and d3LLM, facing VRAM constraints and hardware issues while exploring one-shot diffusion techniques.

The paper introduces a novel task of fact generation for hyper-relational knowledge graphs (HKGs) and proposes KREPE, a generative representation learning method using masked discrete diffusion that unifies link prediction and fact generation, achieving state-of-the-art performance.

Introduces GARD, a diffusion-based framework that operates in the feature space of a feed-forward 3D reconstructor to jointly recover scene geometry and high-quality imagery from degraded inputs.

FlowLM introduces a flow matching language model derived from pre-trained diffusion models via efficient fine-tuning, enabling high-quality few-step text generation that rivals 2,000-step diffusion sampling with far fewer training epochs.