#diffusion

15 curated events

papersTODAY 04:00 UTC

Revisable CTMC Inference Stack Proposed for Guided Discrete Flow Matching TTS

A new arXiv paper introduces a mask-sample-revise inference pipeline built on continuous-time Markov chains for guided discrete flow matching in text-to-speech. The approach targets alignment-free non-autoregressive TTS systems that treat synthesis as conditional infilling over neural codec tokens, avoiding separate duration predictors and external aligners. It proposes letting the sampling process revisit and correct earlier token decisions during generation.

text-to-speech/diffusion/arxiv/codec-tokens

papersTODAY 04:00 UTC

TIDAL: Interleaved Diffusion and Action Loop for High-Frequency VLA Control

A new arXiv paper proposes TIDAL, a control scheme that alternates between diffusion-based planning and action execution to keep vision-language-action models running at high frequency. The authors argue that current VLA systems rely on a low-frequency batch-and-execute approach, and that the resulting mismatch between model inference speed and robot control rate creates gaps in which the agent cannot react. TIDAL aims to close that blind spot while retaining the semantic generalization of large VLA models.

robotics/vla/diffusion/inference-latency

papersTODAY 04:00 UTC

Backward SDE Approach Targets Physics Consistency in Diffusion Models

A new arXiv preprint proposes using backward stochastic differential equations to enforce physics and measurement consistency when score-based diffusion models are applied to inverse problems. The authors argue that existing methods, such as heuristic guidance, periodic projections, or task-specific conditional training, are less principled. The work positions the approach as a general way to keep pretrained diffusion priors aligned with physical constraints.

diffusion/research/physics/generative-models

papersTODAY 04:00 UTC

DiffAdapterVLA: Planner-Integrated Backbone for Driving VLMs

A new arXiv paper introduces DiffAdapterVLA, a method that folds continuous trajectory planning directly into the backbone of a pretrained driving vision-language model. The authors argue that while driving VLMs absorb rich visual, route, language and driving context, their training objectives stay disconnected from continuous planning, so prior work tends to bolt planning on separately. The approach targets native continuous trajectory generation rather than a detached planning head.

autonomous-driving/vlm/diffusion/arxiv

papersTODAY 04:00 UTC

DiffuTester: Accelerating Unit Test Generation for Diffusion LLMs via Mining Structural Pattern

Researchers have introduced DiffuTester, a technique to accelerate unit test generation using diffusion large language models. The method extracts structural patterns from code to guide parallel generation, aiming to make automated testing faster and more scalable. The work addresses the need for efficient large-scale software testing.

diffusion/unit-test/code-generation/llm

papersTODAY 04:00 UTC

AdaFlash: Adaptive Speculative Decoding with On-Policy Distilled Diffusion Drafters

A new arXiv paper proposes AdaFlash, a speculative decoding method that uses diffusion-based draft models distilled on-policy to speed up large language model inference. The approach adapts the drafting process rather than relying on a fixed draft model, aiming to improve acceptance rates during verification by the target model. It builds on prior work in this line, including DFlash, and appears as a revised submission.

speculative-decoding/diffusion/inference/research

papersTODAY 04:00 UTC

Discrete Beckmann Transport Models Target One-Step Language Generation

A new arXiv preprint introduces discrete Beckmann transport models, a framework for language modeling and reasoning that avoids the multi-step sampling usually required by discrete diffusion and flow approaches. The authors note that existing methods compress sampling steps only by distilling a pretrained autoregressive teacher, which limits the student to the teacher's performance. Their approach instead aims to generate text in a single step without relying on that distillation ceiling.

language-models/diffusion/research/inference

papersSEP 11 04:00 UTC

Prediction-Loss Alignment for Sampler-Robust Flow Matching Training

The paper looks at a widely used training recipe for diffusion and flow-matching models, where the network predicts a clean sample that is then converted into a velocity for the loss. That conversion amplifies errors near the endpoints of the noise schedule, making training unstable and tied to a particular sampler. The authors propose aligning the prediction objective with the loss objective so that training stays robust regardless of which sampler is used at inference.

flow-matching/diffusion/training/arxiv

papersSEP 10 04:00 UTC

Researchers propose manifold-aligned generative transport for low-dimensional data structures

A machine learning preprint on arXiv introduces a generative transport method aimed at datasets that concentrate near a low-dimensional structure embedded in a high-dimensional space. The approach seeks to limit probability mass leaking away from the data-supporting manifold while staying computationally practical, in contrast to the iterative sampling used by diffusion models. The paper was updated as a v2 cross-list replacement.

generative-models/diffusion/manifolds/research

papersSEP 10 04:00 UTC

FlowCPO paper unifies online RL and offline preference alignment for flow and diffusion models

A new arXiv preprint presents FlowCPO, a framework that interprets preference alignment for flow and diffusion models through a single divergence-based lens. The work connects online reinforcement learning techniques with offline preference optimization, approaches that had previously been studied in isolation. The authors also examine limitations of existing forward-process alignment methods within this unified view.

flow-models/diffusion/preference-alignment/reinforcement-learning

papersSEP 10 04:00 UTC

Jump-Diffusion Framework Introduced for Generating Irregularly Sampled Time Series

A research paper presents a method for training generative models on continuous-time data that is recorded unevenly and out of sync across sources. The approach builds on generator matching and can represent trajectories with sudden jumps rather than only smooth paths, backed by closed-form expressions for diffusion components. It could be useful in domains where measurements arrive at irregular intervals, such as healthcare monitoring or sensor networks.

time-series/generative-models/diffusion/research

papersSEP 12 04:00 UTC

Information-Theoretic Framework Unifies Generalization Bounds for VAEs and Diffusion Models

A new arXiv paper derives generalization guarantees for both variational autoencoders and diffusion models within a single information-theoretic framework. The analysis exploits the encoder-generator structure shared by the two model families, which earlier theoretical work had largely treated separately. The authors report bounds that clarify how the shared architecture affects performance on unseen data.

research/diffusion/generative-models/theory

papersSEP 12 04:00 UTC

ScaleResfusion combines residual rectified flow with residual vector fields for image restoration

A new arXiv paper introduces ScaleResfusion, a method that uses a residual vector field to guide residual rectified flow for real-world image restoration. The approach targets recovering high-quality images from complex, unknown degradations, a setting where diffusion-based methods have improved perceptual quality but still face remaining obstacles. The abstract excerpt available does not detail the full experimental results.

image-restoration/diffusion/research/generative-models

papersSEP 11 04:00 UTC

Study Argues Continuous Diffusion Can Scale Competitively for Language Modeling

A new arXiv paper revisits Plaid, a likelihood-based continuous diffusion model for text, to test the assumption that continuous diffusion scales worse than discrete alternatives. The authors report that with the right design, continuous diffusion can match discrete diffusion at scale. The work is a replacement version of a cross-listed machine learning preprint.

diffusion/language-models/scaling/research

papersSEP 12 04:00 UTC

Model-Aware Diffusion Schedules Derived via Optimal Transport

A new arXiv paper argues that the schedules controlling how signal and noise are mixed along diffusion and flow-matching paths can be optimized by minimizing a kinetic action borrowed from optimal transport theory. The authors show that making these schedules depend on the specific model, rather than using fixed hand-tuned coefficients, improves generation quality. The work offers a theoretical framing for why certain noise schedules perform better than others.

diffusion/optimal-transport/generative-models/arxiv