Tag
This paper proposes DiffDiff, a diffusion framework for probabilistic time series forecasting that embeds predictability asymmetry into the diffusion trajectory, outperforming six diffusion baselines on seven benchmarks across four prediction horizons.
This paper unifies twelve continuous-time generative models under mean-field game theory via a cost tuple, introduces MFGLab (a PyTorch library that auto-shares training loops and solvers), and proposes DI-Flow with differentiable entropy for better mode coverage.
SeT-Diff proposes the first foundation model for HPC telemetry, using diffusion conditioned on semantic sensor descriptions to enable zero-shot generalization across tasks like imputation, forecasting, and virtual sensing, achieving an MAE of 0.0470 on reconstruction.
Introduces C-VCE, a diffusion framework that builds an interpretable concept bottleneck layer into the generative model, enabling human-guided visual counterfactual explanations without relying on external noise-robust classifiers.
Parallel Decoding Distillation (PDD) is a trajectory-based distillation method that accelerates image and video generation by predicting multiple denoising steps per network evaluation, achieving state-of-the-art performance with 4-8 NFEs on models like LTX-2.3, Wan14B, and Qwen-Image.
A systematic review of diffusion-based methods for medical image inpainting, covering architectures, applications, datasets, and evaluation strategies, with a proposed taxonomy and identification of challenges such as lack of standardized benchmarks.
Repackaged model files for ComfyUI from the Microsoft Mage-Flow model, including diffusion models, text encoder, and VAE.
This paper develops a statistical framework for optimal noise-level allocation in diffusion model training, showing that the optimal schedule is atomic in the coupled regime and follows a square-root entropy proxy in the independent-learner regime, with experiments confirming these predictions.
This paper proposes Spectral Alignment (SPA), a lightweight guidance-based method that reduces exposure bias in diffusion models by calibrating the power spectrum of intermediate predictions, showing consistent improvements across pixel-space, latent, and flow-matching models with minimal computational overhead.
Nunchaku, a 4-bit diffusion inference engine based on SVDQuant, is now natively integrated into Hugging Face Diffusers, enabling fast and memory-efficient loading of quantized diffusion models with a simple from_pretrained() call.
This paper resolves the paradox that diffusion models can accurately recover mixture weights even when the score function is insensitive to them, introducing the Diffusion Score Sensitivity Index (DSSI) and showing that intermediate noise levels provide informative signals for weight recovery.
Proposes DiFA, a training-free framework that reframes inference-time data prediction refinement as sequential state estimation using Kalman filtering, significantly improving generative fidelity on CIFAR-10 and ImageNet.
NVIDIA NeMo Automodel integrates with Hugging Face Diffusers to enable scalable distributed fine-tuning of diffusion models for image and video generation, supporting models like FLUX.1-dev, Wan 2.1, and HunyuanVideo.
FVAttn is a training-free sparse attention system that uses runtime load balancing to improve distributed execution efficiency of adaptive sparse attention under multi-GPU sequence parallelism for video generation, achieving up to 4.41x attention speedup and 2.11x inference speedup over FlashAttention on step-distilled Wan2.2 I2V while maintaining competitive video quality.
This paper introduces a unified conceptual framework for discrete diffusion models, analyzing their design space through tokenization, state space construction, and highlighting trade-offs in training, inference, and scaling.
Introduces Self-Correcting Coupled Markov Jump Processes (SC-CMJP) and a training-free sampler CO2Jump for concurrent image understanding and generation, achieving state-of-the-art joint performance on editing, maze, and nonogram tasks.
Google Research shows that the creativity of diffusion models is a mathematical consequence of neural network regularization causing score smoothing and interpolation, demystifying their ability to generate novel data rather than just memorize.
This paper presents ReDiTT, a retrieval augmented conditional diffusion transformer for asynchronous time series prediction. The model retrieves structurally similar latent sequences as reference conditions to improve long-horizon forecasting and sample diversity, achieving state-of-the-art performance on seven real-world datasets.
This paper proposes a diffusion-based framework for learning adaptive mesh discretization conditioned on observed PDE dynamics, using spectral guidance and physics constraints to allocate resolution where needed. The method achieves competitive or superior performance across five PDE regimes.
This paper develops conservation laws for diffusion models using generalized extrinsic information transfer (GEXIT) functions, showing that the cross-entropy can be characterized as an integral of local information-theoretic derivatives along the noise path, unifying likelihood characterization for discrete and continuous diffusion.