Few-step distillation improves T2I models without real images or CFG trade-offs.
problem Challenges in accelerating T2I models with high-resolution and CFG.
method Score identity distillation (SiD) for few-step generation, with adversarial loss and new guidance strategies.
result State-of-the-art performance on SDXL at 1024x1024 resolution, robust to real images absence.
IMM generates high-quality samples in few steps with stable training.
problem Slow inference and instability in generating high-quality samples using diffusion models and Flow Matching.
method Inductive Moment Matching (IMM) is a new generative model for one- or few-step sampling with a single-stage training procedure.
result IMM achieves state-of-the-art 2-step FID of 1.98 on CIFAR-10 for a model trained from scratch.
ReDi improves few-step generation for discrete data models.
problem Slow sampling speeds in discrete flow-based models.
method Rectified Discrete Flow (ReDi) reduces factorization error by rectifying coupling.
result Empirically, ReDi reduces Conditional Total Correlation and enables few-step generation.
TVM improves generative modeling by matching terminal velocities.
problem Creating high-fidelity one- and few-step generative models.
method TVM generalizes flow matching, modeling transitions between diffusion timesteps and regularizing terminal behavior.
result TVM achieves state-of-the-art FID scores with minimal architectural changes and fused attention kernel.
Few-step protein backbone generators reduce sampling time by over 20x.
problem Computational bottleneck in diffusion-based protein generation models.
method Score distillation adapted for protein backbone generation, combined with inference time noise modulation.
result Significant reduction in sampling time (20+ fold) while maintaining comparable performance.
Improved sampling for Diffusion Models by accounting for covariance.
problem Sampling quality degradation in few-step Diffusion Models.
method Covariance-aware sampler using Tweedie's formula and Fourier-space decomposition.
result Consistently superior samples compared to state-of-the-art samplers.
For the efficient compensation of fiber nonlinearity, one of the guiding principles appears to be: fewer steps are better and more efficient. We challenge this assumption and show that carefully designed multi-step approaches can lead to better performance-complexity trade-offs than their few-step counterparts.
One-step diffusion samplers reduce sampling time and computational costs.
problem Efficient sampling from complex distributions.
method One-step diffusion, self-distillation, deterministic flow.
result Achieves competitive sample quality with fewer evaluations.
FLDD improves discrete diffusion models by learning a non-Markovian noising process.
problem Efficiency and quality of discrete diffusion models in few-step generation.
method Introduces a learnable non-Markovian forward (noising) process to match the target distribution.
result FLDD produces higher quality samples in fewer steps compared to conventional discrete diffusion models.
The theory of Hitchin systems is something like a "global theory of Lie groups", where one works over a Riemann surface rather than just at a point. We'll describe how one can take this analogy a few steps further by attempting to make precise the class of rich geometric objects that appear in this story (including the…
Develops a framework for distilling flow models from few steps.
problem Improving few-step sampling in diffusion models for better performance.
method Local approximation errors and dynamical amplification controlled through analytical tractability.
result Deep residual compositions efficiently approximate long-horizon transport with controlled global error.
Bilevel optimization has been recently revisited for designing and analyzing algorithms in hyperparameter tuning and meta learning tasks. However, due to its nested structure, evaluating exact gradients for high-dimensional problems is computationally challenging. One heuristic to circumvent this difficulty is to use t…
Unified framework reduces NFEs for inverse problems.
problem High computational costs and degraded reconstruction quality in existing LDM-based inverse solvers.
method Consistency Regularised Gradient Flows for posterior sampling and prompt optimization.
result Significantly reduced computational cost with state-of-the-art performance.
This paper is a further extension of the method proposed in Itkin, 2014 as applied to another set of jump-diffusion models: Inverse Normal Gaussian, Hyperbolic and Meixner. To solve the corresponding PIDEs we accomplish few steps. First, a second-order operator splitting on financial processes (diffusion and jumps) is …
DistillKac generates images quickly using damped wave equations.
problem Generating high-quality images efficiently.
method Uses damped wave equations and Kac dynamics for finite speed transport.
result Fast image generation with high quality and numerical stability.
Deep learning frameworks leverage GPUs to perform massively-parallel computations over batches of many training examples efficiently. However, for certain tasks, one may be interested in performing per-example computations, for instance using per-example gradients to evaluate a quantity of interest unique to each examp…
Before deploying autonomous agents in the real world, we need to be confident they will perform safely in novel situations. Ideally, we would expose agents to a very wide range of situations during training, allowing them to learn about every possible danger, but this is often impractical. This paper investigates safet…
Improved NiNo networks accelerate Adam training by up to 50%.
problem Accelerating neural network training with stable and efficient methods.
method Proposed NiNo networks that leverage neuron connectivity and graph neural networks to nowcast parameters periodically during Adam training.
result Accelerates Adam training by up to 50% in vision and language tasks.
LADD models improve discrete diffusion for faster language generation.
problem Practical discrete diffusion models ignore cross-token dependencies, degrading performance.
method Introduces a learnable auxiliary latent channel, diffusing over the joint (token, latent) space.
result LADD models yield improvements on unconditional generation metrics.
SPGD improves adversarial training efficiency and accuracy.
problem Improving adversarial training efficiency and accuracy with fewer steps.
method Adversarial-sample generation from a frequency domain perspective, extending PGD to the frequency domain.
result SPGD achieves greater adversarial accuracy compared to PGD with fewer attack steps.
SA-Solver improves stochastic sampling from DPMs.
problem Efficient sampling from Diffusion Probabilistic Models (DPMs) is time-consuming.
method Proposes SA-Solver, an improved stochastic Adams method for solving diffusion SDE.
result SA-Solver achieves improved or comparable performance compared to SOTA methods for few-step sampling.
In this work we study generalization of neural networks in gradient-based meta-learning by analyzing various properties of the objective landscapes. We experimentally demonstrate that as meta-training progresses, the meta-test solutions, obtained after adapting the meta-train solution of the model, to new tasks via few…
A scalable algorithm for sampling and fine-tuning models using Tilt Matching.
problem Efficient sampling and fine-tuning of generative models.
method Tilt Matching, arising from a dynamical equation, minimizes variance and inherits regularity from stochastic interpolants.
result Empirically verified to be efficient and highly scalable, providing state-of-the-art results.
Consistency models generate high-quality samples fast and without iterative sampling.
problem Slow generation in diffusion models.
method Direct mapping of noise to data, supporting fast one-step generation and multistep sampling.
result Consistency models achieve state-of-the-art FID scores and outperform diffusion models in one-step generation.
Efficiently solves inverse problems with diffusion and flow models in just a few steps.
problem Solving inverse problems like super-resolution, inpainting, or deblurring using diffusion or flow models.
method Conditional Conjugate Integrators framework that projects inverse problem dynamics into a more amenable space for sampling.
result Generates high-quality samples in as few as 5 conditional sampling steps, outperforming competing methods.
CCM improves context for Meta-RL by contrastive learning.
problem Improving context for Meta-RL to enable task generalization.
method CCM framework using contrastive learning for context encoding and information-gain-based trajectory collection.
result CCM outperforms state-of-the-art algorithms in benchmarks and sparse-reward environments.
VFMs use noise adapters to conditionally generate images in one step.
problem Conditional image generation with iterative models is slow and requires explicit sampling paths.
method Developed a variational flow map framework that learns noise distributions for conditional sampling.
result VFMs achieve well-calibrated conditional samples in a single forward pass.
SiD distills pretrained diffusion models into a fast one-step generator.
problem Efficiently distilling pretrained diffusion models into a fast generator.
method Reformulates forward diffusion processes as semi-implicit distributions and uses three score-related identities to create a loss mechanism.
result Achieves high FID performance and significantly reduces generation time.
SCALLOP improves likelihood flow maps for efficient Boltzmann generation.
problem Efficient estimation of model likelihood in flow-based generative models.
method SCALLOP introduces a Hutchinson-free likelihood distillation objective for scalable flow-based models.
result SCALLOP achieves up to 10x inference speedup while improving performance.
Paper proposes a method to speed up discrete diffusion models by distilling many steps into few.
problem Challenges in capturing dependencies between elements in discrete diffusion models.
method Proposes 'mixture' models and loss functions to distill many sampling steps into few.
result Effective in distilling pretrained discrete diffusion models across image and language domains.
The recent progress in neural architecture search (NAS) has allowed scaling the automated design of neural architectures to real-world domains, such as object detection and semantic segmentation. However, one prerequisite for the application of NAS are large amounts of labeled data and compute resources. This renders i…
LSD distills high-quality samplers for DDMs with fewer steps.
problem Inefficient sampling in DDMs leads to low quality and high computational cost.
method LSD employs a distillation approach to train fast samplers with learnable coefficients and time schedules.
result LSD+ achieves higher sampling quality with fewer steps compared to existing samplers.
CoVAE improves VAEs by reducing training steps and improving quality.
problem Two-stage training overhead and increased sampling times.
method Consistency training of VAEs with progressive latent representations and time-dependent β parameter.
result CoVAE generates high-quality samples in one or few steps.
Consistency models accelerate generation with theoretical guarantees.
problem Empirical success of consistency models without theoretical justification.
method Theoretical analysis of consistency models mapping inputs to arbitrary points.
result Achieve KL divergence of order O(ε2) with $ O\left(\log\left(\frac{d}{\varepsilon}
ight)
ight) $ iterations. The paper optimizes interpolation schedules in generative models to improve sampling accuracy.
problem Improving sampling accuracy in generative models with fewer resources.
method Minimizing the averaged squared Lipschitzness of the drift field, using transfer formulas.
result Designed schedules yield more accurate fine-scale statistics at fixed integrator budget.
M-L2O adapts fast to new tasks by self-adapting during test-time.
problem L2O optimizers struggle with out-of-distribution tasks.
method Meta-training an L2O optimizer to adapt quickly to new tasks.
result M-L2O converges significantly faster than vanilla L2O with only 5 steps of adaptation.
Improved algorithm finds second-order stationary points in non-convex optimization.
problem Minimizing non-convex objectives while preserving training data privacy.
method SpiderBoost framework with two gradient oracles: precise and less precise.
result Improved rates for finding second-order stationary points.
Paper tackles matrix estimation under arbitrary noise, achieving minimax optimality.
problem Noisy low-rank-plus-sparse matrix recovery under arbitrary dependence.
method Incoherent-constrained least-square estimator, novel energy spreading result.
result Achieves minimax optimality in estimating structured Markov transition kernels.
In Federated Learning, we aim to train models across multiple computing units (users), while users can only communicate with a common central server, without exchanging their data samples. This mechanism exploits the computational power of all users and allows users to obtain a richer model as their models are trained …
CCVFM uses coreset to improve generative models by refining residual flows.
problem Generating multimodal distributions from scratch is challenging.
method Augments hierarchical rectified flow with a data-informed source distribution using a coreset.
result CCVFM achieves competitive few-step generation without a learned noise-to-data map.
Paper proposes CARL for better control in RL from sensory data.
problem Efficient control of complex systems from high-dimensional sensory inputs.
method Control-Aware Representation Learning (CARL) for model-based RL.
result CARL improves control performance in benchmark tasks.
Directed latent variable models that formulate the joint distribution as p(x,z)=p(z)p(x∣z) have the advantage of fast and exact sampling. However, these models have the weakness of needing to specify p(z), often with a simple fixed prior that limits the expressiveness of the model. Undirected latent variabl…
New RL algorithm explains why deep learning works in stochastic environments.
problem Why deep RL algorithms perform well in practice despite using random exploration.
method Introducing SQIRL, an iterative RL algorithm that separates exploration and learning.
result Effective horizon explains why deep RL works in stochastic environments.
A benchmark evaluates ioUS-to-MR synthesis methods for brain tumor surgery.
problem Difficult interpretation of ioUS images for brain tumor surgery.
method Six generators trained under four inference regimes and two targets on public data.
result SynDiff-2.5D best preserved downstream segmentation (U_Dice=0.55).