Skip to content

🎨 Image Generation

🧠 NeurIPS2026 · 206 paper notes

📌 Same area in other venues: 🎞️ ECCV2026 (335) · 📷 CVPR2026 (495) · 🔬 ICLR2026 (355) · 💬 ACL2026 (5) · 🧪 ICML2026 (141) · 🤖 AAAI2026 (80)

🔥 Top topics: Diffusion Models ×71 · Text-to-Image ×18 · Alignment/RLHF ×11 · Image Editing ×10 · Layout & Composition ×10

ABC: Any-Subset Autoregression via Non-Markovian Diffusion Bridges in Continuous Time and Space

ABC models continuous-time, continuous-space stochastic processes via a single non-Markovian diffusion bridge SDE whose time and volatility adapt directly to physical elapsed time, trained simulation-free via path-dependent denoising score matching to support arbitrary causal and non-causal conditioning subsets.

Accelerating Rectified Flow Models via Trajectory-Aware Caching

This paper proposes TACache, a training-free cache acceleration framework for Rectified Flow models that orthogonally decomposes discrete velocity acceleration into parallel magnitude and orthogonal direction components, using dual cumulative-variation thresholds offline to build a safe skip schedule and recursively re-orthogonalized historical directions online to reconstruct skipped velocities without extra model evaluations, achieving up to 4.14× text-to-image and 2.11× text-to-video speedups.

Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules

Addressing the core limitation that standard flow and diffusion pre-training merely matches observed data distributions and remains blind to out-of-distribution valid regions, this paper introduces Active Flow Expansion (ActFlow), an active continued pre-training framework driven by representation-space verifier uncertainty that establishes the first reachability-based statistical guarantees for OOD generative discovery and substantially expands valid molecular and biological coverage.

Align & Invert: Solving Inverse Problems with Diffusion and Flow-based Models via Representation Alignment

This paper introduces test-time representation alignment between internal diffusion hidden states and a self-supervised visual encoder (DINOv2), using dynamic proxy targets and theoretical local contraction guarantees to substantially boost perceptual fidelity and reduce discretization steps in solving inverse problems without ground-truth supervision.

All Roads Lead to Rome: Flow-driven Multi-Anchor Exploration for Open-Environment Active 3D Mapping

The method generates multiple long-horizon exploration anchors with conditional flow matching, then combines obstacle-aware planning, exploration-mode clustering, and hierarchical path selection to improve mapping coverage in same-difficulty AiMDoom evaluation, difficulty transfer, and AiMDoom-to-MP3D transfer, without guaranteeing generalization to arbitrary open environments.

An exact information theory of generalization phase transitions in Bayesian diffusion models

This paper introduces analytically tractable Bayesian Information Restricted Diffusion (BIRD) models, rigorously proves via the statistical physics Random Energy Model that the memorization-generalization phase boundary occurs exactly when the mutual information between restricted observations and true data equals the log training set size (\(I(\varphi;\mathcal{C}_{x,t})=\ln|\mathcal{D}|\)), and demonstrates that spatially local BIRD models accurately predict individual outputs of early-training UNets and DiTs (\(r^2 \sim 0.85\text{--}0.93\)), revealing how dynamic information restriction enables diffusion models to circumvent the curse of dimensionality.

Analytical Correction for Subsampling Bias in Drifting Models

Addressing the \(\mathcal{O}(1/n)\) centroid-contraction bias inherent in the softmax-weighted minibatch estimator of drifting models, this paper introduces Analytical Bias Correction (ABC)—a plug-in adjustment that requires only two extra lines of code, preserves convex-hull containment, causes no first-order variance inflation, and slashes the residual bias down to \(\mathcal{O}(1/n^2)\).

Are Text-to-Image Models Inductivist Turkeys? A Counterfactual Benchmark for Causal Reasoning

This paper introduces CF-World, a three-level progressive counterfactual benchmark, alongside the CF-Eval pipeline equipped with Prior Resistance Rate (PRR) and Reasoning Retention Rate (RRR) metrics, revealing that current text-to-image (T2I) models suffer sharp performance degradation under counterfactual premises due to deep entanglement between objective rules and visual priors.

Are we really tilting? The mechanics of reward guidance in flow and diffusion models

This paper theoretically proves that reward hacking and mode-selection failure in reward-guided flow and diffusion models stem from the finite-particle plug-in approximation of the Doob \(h\)-function, proposing a zero-overhead closed-form reward damping schedule to eliminate within-mode collapse and clarifying how Best-of-\(n\) sampling compensates for cross-mode selection.

Asymptotically Exact Negative Guidance of Diffusion Models via Positive-Unlabeled Learning

Proposes PU Steering, a principled framework that recasts negative guidance of diffusion models as sampling from an explicit avoidance target density, coupling an online Positive-Unlabeled (PU) posterior classifier with Sequential Monte Carlo (SMC) resampling to achieve asymptotic exactness and eliminate heuristic local repulsion bias.

Balancing Frequencies and Pixels in Flow Matching

To address the spectral bias in pixel-space flow matching models where the natural \(1/f^2\) power spectrum causes optimization to prioritize low frequencies and delay high-frequency texture learning, this paper proposes a Focal Log-Frequency Loss (\(\bm{f}\)-loss) combining logarithmically compressed spectral residuals with maximum-residual normalization, and introduces a Sigmoid schedule (\(\bm{fv}\)-loss) that emphasizes spectral balancing early in training before transitioning to spatial \(\bm{v}\)-loss for phase and edge refinement, accelerating convergence by up to 40% without architectural changes and achieving 1.83 FID and 323.4 IS on ImageNet \(256\times 256\) in 500k steps.

Barycentric Guidance: Turning Foundation Image Editors into Continuous Affective Controllers

This paper introduces Barycentric Guidance, a training-free framework that constructs a 2D simplex mesh of discrete emotions over the valence-arousal (VA) space and reformulates classifier-free guidance in latent space, turning pretrained foundation image editing diffusion models into continuous, accurate, and wide-coverage affective controllers.

Beckmann Transport Models: From Autonomous Flows to One-Step Maps

Addressing the reliance of traditional flow matching on time-conditioned networks and multi-step ODE integration, this paper proves that when the target distribution is supported on a lower-dimensional manifold with codimension at least two, a time-independent autonomous flow achieves exact pushforward measure transport via divergence balance, and introduces a kernel-free stop-gradient objective to learn one-step generative maps directly from samples.

Bernoulli Flow Models: Self-Consistent Generative Modeling for Binary Data

This paper proposes Bernoulli Flow Models (BFM), which predefine a unified continuous global Bernoulli probability flow path in the binary parameter space and derive analytical closed-form posterior transitions over arbitrary time intervals, eliminating the structural training-inference mismatch of discrete diffusion under low-NFE cross-step sampling and reducing 16-step FID on LSUN Churches \(256\times 256\) from 204.10 (BLD) to 9.22 without distillation or retraining.

Best-of-\(N\) Guidance for Test-time Diffusion Alignment

Best-of-\(N\) Guidance (BoNG) is a rollout-free test-time alignment framework that internalizes the Best-of-\(N\) selection principle directly into the reverse diffusion trajectory, leveraging online-selected Tweedie denoised estimates as asymmetric guidance signals to dramatically boost both single-output peak metrics and multi-output average fidelity with negligible inference latency.

Best-of-N Guidance for Test-Time Diffusion Alignment

Addressing the inefficiency of Vanilla Best-of-\(N\) (which evaluates rewards only post-hoc) and the prohibitive rollout overhead of existing sample-based guidance methods, Best-of-\(N\) Guidance (BoNG) performs online Best-of-\(N\) selection on one-step Tweedie denoised estimates and injects an asymmetric inter-particle guidance signal during reverse diffusion, achieving superior single-sample and multi-sample alignment without requiring additional score network forward or backward passes.

Beyond Trajectory Matching: ReFlow with Marginal Distribution Alignment

Addressing the theoretical under-determination where trajectory matching in ReFlow cannot uniquely determine endpoint marginal distributions, this paper introduces a marginal-alignment regularizer (MA-Reflow) without auxiliary networks, integrating an augmented ODE log-density tracker with frozen teacher score evaluations to achieve high-quality 4-step image synthesis.

BitDance: Scaling Autoregressive Generative Models with Binary Tokens

BitDance introduces a scalable autoregressive (AR) image generation framework built on high-entropy binary visual tokens (with a vocabulary size up to \(2^{256}\)), combining a continuous-space Binary Diffusion Head with Next-Patch Diffusion parallel prediction to achieve a state-of-the-art 1.24 FID among AR models on ImageNet 256×256 and over 30× inference speedup in 1024×1024 text-to-image generation.

Boosting Text-to-Image Diffusion Models via Core Token Attention-Based Seed Selection

This paper reveals that early-stage cross-attention on prompt core tokens strongly predicts final image generation quality, and accordingly proposes Attention-Based Seed Selection (ABSS), a training-free, plug-and-play early seed screening method that consistently boosts text-image alignment and human preference scores across U-Net and DiT/Rectified-Flow backbones at only 71–73 NFE per image.

Bridge Graphical Models: Coupling, Projection, and Current-Preserving Dynamics for Generative Modeling

This paper introduces Bridge Graphical Models (BGMs) to decouple continuous-time generative modeling into endpoint coupling, bridge kernel, Markovian projection, and current-preserving dynamics representation, alongside the "Markovization gap"—a minutes-scale pre-training diagnostic that measures the irreducible conditional velocity variance under Markov compression and accurately ranks coupling and bridge designs in alignment with downstream flow-matching loss and FID.

BridgeTwist: Twisting Schrödinger Bridges for Training-Free Conditional Sampling

To address the lack of training-free conditional samplers for pretrained Schrödinger bridge generative models, BridgeTwist introduces a tempered twisted sequential Monte Carlo framework combining closed-form velocity–score endpoint prediction with a bridge-shaped diffusion schedule, achieving asymptotically exact, high-fidelity, and diverse conditional generation.

CAB: Accelerating Flow and Diffusion Sampling via Rectification and Corrected Adams-Bashforth

Addressing the severe sample distortion and instability of flow and diffusion models under tight budgets (6–20 NFEs), CAB unifies the reverse-time ODE into rectified noise-to-signal coordinates to remove drift, combining explicit Adams-Bashforth multistep integration with an extrapolation-defect correction term to achieve high-order convergence and high-fidelity generation with strictly 1 NFE per step.

ColorConceptBench: A Benchmark for Probabilistic Color-Concept Understanding in Text-to-Image Models

This paper introduces ColorConceptBench, an expert-annotated benchmark comprising 1,281 implicit concepts and 6,584 designer-colored sketches that replaces traditional deterministic single-point color matching with probabilistic color distribution distances (such as EMD) to systematically evaluate how nine mainstream text-to-image (T2I) models understand abstract visual states and emotional color semantics.

Concurrence of Symmetry Breaking and Nonlocality Phase Transitions in Diffusion Models

This paper empirically unifies "symmetry breaking" and "nonlocality" phase transitions in modern diffusion transformers (DiT-XL and SD3 Medium) through instantaneous score probes and integrated sampling probes, demonstrating that semantic guidance and global cross-patch attention activate concurrently within the same critical time window to provide a principled diagnostic for pruning redundant computation.

Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation

This paper proposes CoFlow, a training-free, inference-time adaptive step selection framework that extracts lightweight semantic and structural prompt features and employs a neural contextual bandit (NeuralUCB) guided by an unsupervised velocity-variation surrogate reward to select integration step budgets online, achieving over \(2.5\times\) speedup across image generation, image editing, and video generation while preserving output quality.

Continual Learning in Modern Hopfield Networks with an Application to Diffusion Models

By leveraging the implicit energy equivalence between modern Hopfield networks (MHNs) and diffusion models, this paper defines unsupervised forgetting after a task shift as the rise in Hopfield energy (intrinsic forgetting), theoretically proves that high-energy outlier samples reside in sharper energy basins and are thus forgotten more severely, and derives a high-energy-prioritized memory replay rule that accurately predicts both per-sample forgetting and replay gains in Stable Diffusion and pixel-space DDPMs.

Coreset-Induced Conditional Velocity Flow Matching

This paper proposes Coreset-Induced Conditional Velocity Flow Matching (CCVFM), which constructs a zero-parameter closed-form conditional velocity source distribution via an entropic Sinkhorn coreset and a low-rank Gaussian mixture model (GMM), allowing the inner correction flow to learn only a within-mode residual from the surrogate to the target rather than transporting isotropic Gaussian noise across modes from scratch, achieving strong few-step generation at 51 NFE on MNIST (\(\text{FID}_{50k}=0.75\)), CIFAR-10 (\(6.35\)), ImageNet-32 (\(8.76\)), and CelebA-HQ 256 (\(\text{FID}_{28k}=4.17\)).

CSFlow: Aligning Flow Matching with Human Contrast Sensitivity

This paper establishes the first mathematical connection between the spectral evolution of reverse flow matching and the human Contrast Sensitivity Function (CSF), estimating the spatial frequencies recovered within each denoising interval and weighting them by human perceptual sensitivity to improve both visual realism and quantitative metrics (lowering FID by 4.7%, increasing IS by 2.2%, and boosting GenEval by 2.5%) via training-free inference step modification and short fine-tuning.

Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models

Inspired by random organization and mechanical annealing in periodically driven disordered solids, this paper introduces stroboscopic cyclic denoising—iterating partial forward noising and full reverse denoising at controlled amplitudes—to expose a yielding-like transition and ultrastable memorized attractors in diffusion models, enabling fully unconditional, prompt-free, and training-data-agnostic extraction of memorized training images using only sampler-level control.

D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models

This paper discovers that modern diffusion models equipped with large language model (LLM) or vision-language model (VLM) encoders naturally inherit in-context learning capabilities, and proposes D-OPSD, an on-policy self-distillation framework where the same model acts as both a text-conditioned student and a multimodal-conditioned teacher to align velocity predictions along the student's own few-step rollout trajectories without external reward models.

D2D: Detector-to-Differentiable Critic for Improved Numeracy in Text-to-Image Generation

D2D converts non-differentiable object detectors into gradient-friendly counting critics using high-curvature Sigmoids and logit-distance scaling, guiding a lightweight Latent Modifier Network at test time to boost text-to-image numeracy by up to 13.7% across diverse diffusion backbones with minimal computational cost.

DC-DiT: Adaptive Compute and Elastic Inference for Visual Generation via Dynamic Chunking

DC-DiT replaces fixed patchification in Diffusion Transformers with a learned encoder-router-decoder scaffold that adaptively compresses the 2D latent grid into a compact token sequence via spatial predictability, enabling single-checkpoint elastic inference and unsupervised spatio-temporal compute allocation.

Delta-Adapter: Scalable Exemplar-Based Image Editing with Single-Pair Supervision

Delta-Adapter extracts a content-decoupled visual feature "Semantic Delta" from a single source-target exemplar pair and injects it into a frozen FLUX diffusion model via a Perceiver-based adapter, enabling scalable exemplar-based image editing and single-pair test-time adaptation without textual guidance or pair-of-pairs supervision.

Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View

Starting from continuous-time path-space importance sampling and value-gradient variance reduction, this paper unifies reverse-trajectory policy gradient methods (e.g., Flow-GRPO) and forward-noising matching methods (e.g., AWM, DiffusionNFT) into a single five-parameter design space, and introduces a rollout-reusing multi-sample KDE value-gradient estimator paired with scale-bounded timestep weights to achieve significantly faster RL convergence and higher rewards on SD3.5-Medium and Qwen-Image.

DetailAnywhere: Fashion Detail Generation via Cross-Modal Feature Alignment Distillation

Addressing cross-branch semantic misalignment and detail hallucination in multimodal diffusion transformers for apparel close-up generation, DetailAnywhere proposes a dual-branch feature alignment distillation framework via a lightweight discriminative teacher alongside a multi-dimensional consistency reward model, achieving state-of-the-art prompt-free detail synthesis on the newly released FDBench.

DiDE:Direct Injection with Color-Texture DEcoupling for 3D Stylization

DiDE reveals that texture information saturates only a small subset of high-variance channels in the structured latent space of image-to-3D models, enabling training-free, decoupled modulation of color and texture for 3D assets via branch-specific preprocessing and self-attention channel partitioning.

Diff-CA: Separating Common and Salient Factors with Diffusion Models

Diff-CA first trains a low-noise-dependence flow-matching generative model conditioned on DINOv3 cross-query tokens and a color token, and then decomposes the conditioning sequence into uniquely identifiable common and salient factors using only dataset-level binary supervision, enabling precise attribute swapping and interpolation without sacrificing image fidelity.

Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL

Didr identifies terminal reward domination in one-step generator alignment, proves the minimizer equivalence of Integral KL to clean-image RLHF, and propagates optimal reward signals across all diffusion noise levels via a differentiable short-step proxy, achieving Pareto-optimal preference alignment without training images.

DiffRatio: Training One-Step Diffusion Models Without Teacher Supervision

DiffRatio eliminates the dual-score estimation bias in traditional score-based distillation by directly estimating the score difference between student and data distributions via the gradient of a learned time-conditioned log density ratio, achieving SOTA teacher-free one-step generation on CIFAR-10 and ImageNet alongside test-time parallel importance resampling.

Diffusion Guidance Is a Controllable Policy Improvement Operator

This paper rigorously proves that classifier-free guidance in diffusion and flow-matching models is mathematically equivalent to a reinforcement learning policy improvement operator, introducing CFGRL to dynamically tune the exploration-optimality tradeoff at test time without retraining, and dramatically enhancing goal-conditioned behavioral cloning without learning any value function.

Diffusion Models Observe Only Gradients: A Geometric Perspective on Score Matching Errors

By applying a weighted Helmholtz-Hodge decomposition to score matching errors, this paper proves that only the gradient component affects reverse marginal Fokker-Planck dynamics while the solenoidal component is structurally invisible, thereby deriving a strictly tighter KL divergence upper bound and a sample-free dual Sobolev estimator that correlates substantially better with image generation quality (FID).

Diffusion Models without Classifier-free Guidance

Presents Model-guidance (MG), an innovative training objective that internalizes implicit Bayesian posterior conditioning into a single-pass network, completely eliminating Classifier-free Guidance (CFG) at inference, doubling generation throughput, and achieving a state-of-the-art 1.34 FID on ImageNet 256×256.

Direct Product Flow Matching: Decoupling Radial and Angular Dynamics for Few-Shot Adaptation

Addressing angular velocity distortion and radial confidence loss caused by radial-angular coupling in flow matching for few-shot adaptation, DP-FM formulates a warped product Riemannian framework with a constant-warping metric that strictly decouples cross-modal alignment into independent radial evolution and constant-speed spherical geodesic transport, combined with hidden-state classifier-free guidance.

Discrete Diffusion Models Exploit Asymmetry to Solve Lookahead Planning Tasks

By systematically comparing Autoregressive (AR) Transformers and Discrete Diffusion Language Models (dLLMs, Non-Autoregressive NAR) on the Star-Path lookahead planning benchmark, this paper shows that while both paradigms achieve 100% accuracy when trained with full-sequence gradients over the graph prefix, dLLMs spontaneously exploit topological asymmetry by decoding backwards from the target leaf to the center source, reducing complex \(\ell^{\text{th}}\)-order lookahead to \(1^{\text{st}}\)-order local neighbor transitions and converging with exponentially fewer training examples.

Discrete Flow Matching: Convergence Guarantees Under Minimal Assumptions

For flow matching models on discrete state spaces (Discrete Flow Matching, DFM), this paper derives non-asymptotic KL divergence bounds for the early-stopped target distribution and the first explicit total variation (TV) distance convergence guarantees to the true target distribution under minimal score approximation assumptions, using nearest-neighbor and uniform random walk reference processes with DMPM time discretization.

Disentangled Representation Learning via Flow Matching

Casts disentangled representation learning as learning factor-conditioned flows within a compact latent space, decomposing the transport velocity field via output attention and non-overlapping orthogonality regularization to achieve high-fidelity generation and fine-grained semantic control.

Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning

To prevent collateral damage caused by latent feature entanglement in conventional sparse autoencoders, SAEParate explicitly disentangles the latent space through cross-attention-guided spatial pooling, latent-space contrastive learning, and a nonlinear encoder, achieving surgical diffusion concept unlearning while preserving retained attributes.

Disentangling Dual Image References in Frequency Aware Diffusion Models for Personalized Generation

This paper proposes Dual-FDM, a training-free frequency-aware diffusion framework for dual-reference personalized image generation that separates foreground and background representations by inverting singular values of cross-attention key projection matrices, replaces mid-frequency structure or low-frequency color bands in the Discrete Cosine Transform (DCT) domain, and adaptively modulates early-to-late denoising timesteps via an information entropy ratio to eliminate text misalignment and style leakage.

DisRFM: Polar Riemannian Flow Matching for Structure-Preserving Graph Domain Adaptation

DisRFM introduces a geometry-aware graph domain adaptation framework that decomposes graph embeddings on a constant-curvature Riemannian manifold into geodesic polar coordinates—calibrating topology-sensitive radial scales via 1D Wasserstein alignment, preserving scale-normalized angular class semantics via confidence-gated radially weighted prototypes, and replacing adversarial domain confusion with metric-corrected polar Riemannian flow matching—consistently outperforming 20 baselines across node-density, edge-density, and feature shifts.

Do Image Editing Models Understand Lighting?

This paper introduces the 3D-anchored Light Probe (3DLP) benchmark built upon 1,000 real-world indoor HDR image pairs of a light probe being physically turned on and off, alongside two affine-invariant pixel-level physical metrics—Standardised Intensity Error (SIE) and Low-Frequency Error (LFE)—that factor out automatic white-balance, exposure, and bulb intensity shifts to rigorously quantify the physical light transport capabilities of modern image editing models.

Drag as Evidence: Motion-Grounded Latent Recomposition for Drag-Based Editing

By treating pixel-space warping as coarse motion evidence rather than an intermediate edited product, MoRe-Drag combines region-aware latent recomposition with generation-stage adaptive conditioning to achieve high-precision spatial drag control and natural generative synthesis on modern multimodal editing backbones.

DrawingsDreamer: A Unified Multi-View Engineering Drawings Generation Model

DrawingsDreamer formulates multi-view engineering drawing generation as a pure sequence modeling task via hierarchical postfix tokenization, eliminating raster visual encoders and unifying localized topological repair and multi-view generation within a single LLM under a task-aware curriculum schedule.

Dream-Cubed: Controllable Generative Modeling in Minecraft by Training on Billions of Cubes

This paper introduces Dream-Cubed, a large-scale Minecraft dataset comprising over 2 million \(32^3\) block chunks (tens of billions of cubes), and systematically compares discrete masked diffusion (MD4) against continuous embedding diffusion (DDPM) using a unified 280M-parameter 3D Diffusion Transformer, demonstrating that diffusing directly over native discrete block tokens not only matches or outperforms real terrain in render-based FID (59.26) and human preference win rate (67.1% against real chunks), but also enables zero-shot inpainting, biome blending, user-authored block prompting, and multi-chunk sliding-window world outpainting.

Drift Flow Matching

Drift Flow Matching (DFM) bridges one-step Drift Models and multi-step Flow Matching by learning a mean-velocity-parameterized two-time transport map between arbitrary intermediate marginals \((p_t, p_r)\) under group-wise distribution-level drift supervision, achieving 1.52 FID at 1 NFE and scaling to 1.31 FID at 10 NFE on ImageNet \(256\times 256\) with a 463M backbone without requiring JVP or self-distillation.

DyCoRM: Dynamic Criterion-Aware Reward Modeling for Text-to-Image Generation

This paper proposes DyCoRM, a two-stage dynamic criterion-aware reward model for text-to-image (T2I) generation that first grounds task-relevant fine-grained criteria from the prompt-image context and then predicts criterion-conditioned pairwise preferences, supported by the 23,378-pair DyCoDataset-20K, the DyCoBench-1K benchmark, and the DyCoPick selection framework to substantially outperform baselines including HPSv3 and GPT-5 across single-criterion, multi-criteria, and overall preference evaluations.

DynT2I-Eval: A Dynamic Evaluation Framework for Text-to-Image Models

To eliminate prompt-set overfitting and benchmark contamination inherent in static evaluation, DynT2I-Eval constructs a structured dynamic prompt space from long-form descriptions and sustains a self-evolving online text-to-image leaderboard through micro-batch pairwise comparisons, UCB scheduling, and confidence-weighted Bayesian updates.

Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing

To tackle long-context intent dilution and state contamination in multi-turn image editing, Edit-R2 reconstructs session intent via in-context chain-of-thought and jointly optimizes discrete text reasoning with continuous latent flow matching under trajectory-level RL and prefix-valid advantage refinement.

Edit the Bits, Diff the Codes: Bitwise Residual Editing for Visual Autoregressive Models

Proposes BitResEdit, a training-free editing framework that exploits two native algebraic structures of bitwise-residual visual autoregressive generators, coupling decision-time closed-form Bernoulli-KL trust-region bit guidance (BitEdit) with combination-time mask-gated continuous-code residual injection (ResEdit) to achieve superior prompt alignment and exact background preservation.

Efficient Adjoint Matching for Fine-tuning Diffusion Models

By reformulating the stochastic optimal control framework with a memoryless linear base drift and a rederived terminal cost, this paper eliminates the expensive forward SDE simulation and backward adjoint ODE simulation in Adjoint Matching, achieving up to 4x faster training convergence while strictly preserving the identical optimal path measure.

Efficient bias mitigation in T2I diffusion models using Concept Graphs

The paper proposes CO-ALIGN, a framework that extracts concept neuron graphs from both the text encoder and the denoiser cross-attention layers, and jointly aligns their relational concept topologies via differentiable ranking loss to mitigate demographic bias without semantic incoherence while boosting concept unlearning robustness.

Efficient Image Synthesis with Sphere Latent Encoder

This paper introduces Sphere Latent Encoder, which decouples reconstruction and generation via a pretrained representation autoencoder, executing all iterative denoising steps within a spherical latent space and achieving high-fidelity few-step image generation at substantially reduced computational cost.

Emergence of Distortions in High-Dimensional Guided Diffusion Models

Using dynamic mean-field theory and the Random Energy Model (REM) from statistical physics, this paper proves that classifier-free guidance (CFG) distortions inevitably emerge in high dimensions when the number of classes scales exponentially with dimension or is continuous, and proposes a negative-guidance window schedule that simultaneously boosts class separation and sample diversity.

Enabling Preference-driven Unlearning in Few-step Distilled Text-to-Image Diffusion Models

Addressing the failure of conventional noise-prediction preference optimization in few-step distilled diffusion models, this paper introduces CePU, a consistency-error-driven preference unlearning framework that directly erases unsafe concepts from distilled models in 15 minutes on a single GPU without expensive re-distillation while preserving general utility.

EPIC: Efficient Predicate-Guided Inference-Time Control for Compositional Text-to-Image Generation

EPIC casts compositional text-to-image inference-time refinement as predicate-guided search, compiling prompts once into fixed typed visual programs and leveraging modular perceptual evidence to drive targeted editing and resampling, achieving dramatic gains in prompt-level accuracy while drastically cutting model execution and token overhead.

Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models

Equilibrium Matching (EqM) replaces the time-conditional non-equilibrium dynamics of diffusion and flow models with a time-invariant equilibrium gradient field over an implicit energy landscape whose gradient vanishes on the data manifold, turning generative sampling into gradient descent optimization that achieves 1.90 FID on ImageNet 256×256 and natively handles partially noised denoising and OOD detection.

Ergodic Trajectory Design by Learned Pushforward Maps: Provable Coverage via Conditional Flow Matching

To overcome the heavy online re-optimization costs and brittle constraint adaptation in spatial ergodic coverage, this paper proposes a two-stage framework decoupling ergodicity from density matching: an analytic trajectory guarantees exact uniform ergodicity on a singularity-free annular domain, while an optimal-transport conditional flow matching (OT-CFM) map learned offline pushes the latent distribution onto the target density, yielding an amortized, multi-agent reusable trajectory with provable \(O(1/\sqrt{K})\) convergence and modular soft-penalty constraint integration.

Expanding the Role of Diffusion Models for Robust Classifier Training

Expanding the role of diffusion models beyond mere synthetic data generation in adversarial training, this paper introduces Diffusion Representation Alignment (DRA)—distilling intermediate denoising representations from frozen diffusion models into discriminative classifiers via a lightweight projection head to significantly enhance clean accuracy, AutoAttack robustness, and feature disentanglement without any inference overhead.

Feature Information Dynamics in Diffusion

Grounding feature emergence in classical I-MMSE relations, this paper introduces Feature Information Dynamics (FIDy) and a hierarchical chain decomposition that quantifies mutual information growth rate via the gap between unconditional and conditional minimum mean squared errors (MMSE), precisely mapping how latent representations (Pixel, SDVAE, VAVAE, RAE) govern the chronological order of visual feature synthesis across continuous signal-to-noise ratios.

FiRe: Fine-grained Multimodal Reasoning for Enhanced Image Generation

Addressing the lack of fine-grained self-reflection and localized editing in unified MLLMs for text-to-image generation, FiRe introduces a six-step closed-loop reasoning pipeline comprising prompt summarization, semantic tuple decomposition, tuple-level VQA, precise feedback generation, and localized correction, complemented by FiRe-GRPO with step-specific rewards and step-level advantage estimation, substantially outperforming existing diffusion and autoregressive baselines on compositional benchmarks.

Fixed-Point Masked Generative Modeling

This paper introduces CoFRe, a fixed-point masked generative modeling framework that replaces intermediate fixed-depth transformer stacks with a parameter-shared fixed-point solver for adaptive depth, combined with cross-step consistency loss and token-aware three-state reuse to drastically cut parameter count and VRAM while substantially improving low-budget generation quality.

FlashControl: One-Step Controllable Generator via Distillation-Friendly Single-Stream Teachers

Addressing the severe optimization instability and control failure when distilling dual-branch architectures like ControlNet into one-step generators, this paper proposes DeltaControl—a single-stream teacher fine-tuned selectively on feed-forward network (FFN) layers—paired with AutoAdapter condition alignment and a Latent Sensitivity Preservation objective to achieve high-fidelity controllable synthesis in a single forward pass (1 NFE).

Flow-DPPO: Divergence Proximal Policy Optimization for Flow Matching Models

Flow-DPPO uncovers the inherent estimation bias in ratio clipping caused by single-sample Gaussian noise in high-dimensional continuous latent spaces, and introduces an exact, closed-form KL divergence mask derived from velocity field mean displacements to enforce strict trust-region policy optimization at zero additional computational cost.

Flow Matching for Count Data

Addressing the challenges of discrete, non-negative, high-range count data that resist continuous relaxation, this paper proposes count-FM—a flow matching framework based on a continuous-time birth-death process with local unit jumps—which matches marginal transition rates without simulation and achieves state-of-the-art generation and transport on single-cell genomics and neural spike data with substantially fewer parameters.

FLUX: Geometry-Aware Longitudinal Flow Matching with Mixture of Experts

To tackle both curved low-dimensional manifolds and latent dynamical regime switches in unpaired longitudinal snapshot sequences, FLUX builds data-dependent geometry-aware paths and decomposes the continuous velocity field using a sparse mixture-of-experts router, enabling accurate multi-marginal transport alongside unsupervised regime discovery.

Follow the Mean: Reference-Guided Flow Matching

By identifying that the velocity field in flow matching is governed by the conditional endpoint mean, this paper introduces a new controllable generation paradigm via reference-guided mean shifts, delivering training-free control on frozen models (RMG) and an adaptive semi-parametric architecture with swappable banks (SPG).

FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance

FoMo presents a fully automated, annotation-free data generation pipeline that leverages the forking moment along a diffusion denoising trajectory as a pointwise perceptual distance proxy, supervised by a batch-level global ranking objective to substantially outperform human-annotated baselines across diverse benchmarks and architectures.

Gen-Searcher: Reinforcing Agentic Search for Image Generation

Presents Gen-Searcher as the first multimodal deep search agent trained via agentic reinforcement learning for knowledge-intensive image generation, introducing a dedicated data curation pipeline, the KnowGen benchmark, and a dual-reward GRPO objective combining text and image feedback to improve Qwen-Image by 16.54 points on KnowGen.

Generalized Wasserstein Flow Matching: Transport Plans, Everywhere, All at Once

Extends flow matching to the space of metameasures (probability measures over probability measures) via coupled outer and inner transport plans, and resolves the nested transport computational bottleneck using sliced and lazy linear approximations for fast, straight-line point cloud generation.

Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation

Decouples one-step sample generation in an autoencoding reconstruction latent space from distribution matching in a compact frozen self-supervised learning (SSL) feature space, achieving a 39x FID reduction on ImageNet via Sinkhorn divergence with classifier-free guidance while exposing metric hacking.

Generation Navigator: A State-Aware Agentic Framework for Image Generation

Reframing text-to-image synthesis into a state-conditioned sequential action-making process, this paper presents Generation Navigator, a closed-loop multi-turn agent, and PRE-GRPO, a trajectory-level reinforcement learning objective that jointly optimizes peak discovery, quality retention, and turn efficiency.

Generative Modeling via Drifting

This paper introduces Drifting Models, a novel generative paradigm that shifts the distribution pushforward process entirely to training time via an anti-symmetric drifting field, achieving state-of-the-art native one-step (1-NFE) image generation without iterative ODE/SDE sampling or adversarial optimization.

GenEvolve: Self-Evolving Image Generation Agents via Tool-Orchestrated Visual Experience Distillation

GenEvolve reformulates open-ended image generation as a tool-orchestrated multi-turn visual trajectory learning task, distilling five-slot structured visual experience extracted from best-worst trajectory pairs into a privileged teacher branch to enable self-evolving generation agents via on-policy self-distillation and GRPO.

GenScale: A Benchmark for Relative Object Scale in Image Generation and Editing

To address the pervasive failure of modern image generation and editing models in respecting real-world size proportions, this paper introduces GenScale, a pairwise benchmark and calibrated ordinal evaluation protocol, along with Rescale, a model-agnostic scale-correction agent.

GeoFidelity-Bench: Evaluating Block-Conditioned Geographic Fidelity in Text-to-Image Street-View Generation

This paper introduces GeoFidelity-Bench, the first street-view text-to-image benchmark evaluated at named street-block granularity across 25 global cities, revealing that open-weight diffusion models struggle to separate target blocks from same-city alternatives and that standard CLIP text alignment is largely uninformative about local geographic fidelity.

GMO-E²dit: Grounded Multi-Operation Editing for E-Commerce Images

Addressing the brittleness and lack of rollback in one-shot compositional e-commerce image editing, GMO-E²dit introduces an agentic framework where a VLM decouples planning and reflection from mask-conditioned rendering, achieving closed-loop, auditable multi-operation completion.

Guiding Visual Autoregressive Models through Spectrum Weakening

Introduces Spectrum Weakening Guidance (SWG), a training-free guidance framework that attenuates feature channels via discrete Fourier transform truncation and energy renormalization, enabling robust unconditional and conditional guidance in visual autoregressive models without requiring external negative prompts or architecture modifications.

GVCC: Zero-Shot Video Compression via Codebook-Driven Stochastic Rectified Flow

By converting pretrained rectified-flow video generative models into marginal-preserving stochastic differential equations, GVCC transmits per-step codebook atom indices and signs to directly steer the decoder's generation trajectory, achieving training-free zero-shot video compression at ultra-low bitrates.

HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing

To bridge the embodiment gap and avoid physical interaction disruption when transferring first-person human demonstrations to robot policies, HandEdit introduces the first large-scale embodiment-aware editing benchmark featuring 200M instances across 26 URDFs with a multi-dimensional metric suite.

Heavy-Tailed Flow Matching via Random Clocks

HTFM casts heavy-tailed source distributions as mixtures of clock-conditioned Gaussians, preserving tractable affine flow matching targets while enabling cross-family tail calibration and superior low-NFE generation via truncated logsignatures.

HERO: Improving the Reliability and Sensitivity of Generative Model Evaluation Using Historical Data

To tackle the core dilemma in generative AI evaluation where expert gold labels are scarce and costly while crowdsourced silver labels are noisy and biased, this paper proposes HERO, a framework that leverages sparse gold annotations embedded in historical evaluation rounds to calibrate silver labelers' sensitivity and specificity via EM, and incorporates prompt-level historical covariates as score-level control variates to drastically suppress estimation bias and variance.

Hierarchical Variational Policies for Reward-Guided Diffusion

Addressing the heavy computational bottleneck of test-time optimization in reward-guided diffusion, this work recasts guidance as a Hierarchical Variational Policy (HVP) that amortizes both initial noise and per-step stochastic controls into lightweight DiT networks, delivering over 5x faster inference (AHVP) while maintaining competitive perceptual quality, alongside an optional semi-amortized refinement regime (SHVP).

How Diffusion Models Memorize

Revealing that memorization in diffusion models is fundamentally driven by early overestimation of training samples amplified by classifier-free guidance, which prematurely suppresses initial noise variance and forces latent trajectories to collapse onto the memorized image.

HyperTransport: Amortized Conditioning of T2I Generative Models

HyperTransport trains an ultra-lightweight hypernetwork using 1D Wasserstein optimal transport loss to directly map concept representations from a frozen encoder (CLIP) to affine activation-steering parameters in one-step text-to-image generators, amortizing per-concept optimization into a single forward pass (speedup of 3600–7000×) and naturally unlocking zero-shot cross-modal image conditioning.

Ideal: In-DEpth ALignment Makes A Discrete Representation AutoEncoder

Addressing the severe detail-loss dilemma in representation autoencoders caused by the semantic abstraction of deep vision foundation model (VFM) features, IDEAL fuses shallow appearance cues with deep semantic representations via cross-attention before vector quantization and optimizes dual-head deep/shallow feature alignment alongside a DINO discriminator, establishing state-of-the-art results on ImageNet with 0.61 rFID reconstruction and 1.89 gFID autoregressive generation.

Improved Baselines with Representation Autoencoders

RAEv2 simplifies and supercharges Representation Autoencoders (RAE) by adopting parameter-free multi-layer feature addition, uncovering a complementary semantic-spatial mechanism between RAE and REPA, and repurposing the REPA head into a zero-overhead \(x\)-prediction self-guidance module, yielding over 10× faster convergence and a state-of-the-art 1.06 gFID in just 80 epochs on ImageNet-256.

Improved Techniques for Fine-Tuning Flow Models via Adjoint Matching: A Deterministic Framework

This paper introduces a deterministic adjoint matching control framework for fine-tuning flow-based generative models, integrating a truncated adjoint scheme with higher-order polynomial regularization to slash per-update wall-clock time from 345s to 32s while preserving sample diversity and preventing reward hacking.

Improving Flexible Image Tokenizers for Autoregressive Image Generation

To eliminate the downstream autoregressive generation bottleneck caused by information over-concentration in early tokens under naive nested dropout, ReTok introduces redundant token padding and hierarchical semantic regularization to activate tail tokens and deliver high-fidelity progressive generation.

Improving Function Space Flow Matching with Kernel Optimal Transport

Addressing the large velocity variance and geometry neglect caused by independent endpoint pairing in functional flow matching, this paper proposes kernel Functional Flow Matching (kFFM), which incorporates kernel-induced entropic optimal transport endpoint coupling without altering the underlying neural operator backbone, significantly boosting generation fidelity and training stability across path and PDE benchmarks.

Inline Critic Steers Image Editing

To tackle heterogeneous difficulty across spatial regions and cases in instruction-based image editing, this paper embeds a learnable critic token directly into the forward pass to predict intermediate generation errors and steer latent representations without external evaluators or extra sampling steps.

Keep It CALM: Analyzing the Limits of Global Unsafety in Text-to-Image Generation

From a geometric perspective, this paper exposes the coverage-selectivity dilemma of global unsafe subspace removal in text-to-image generation and introduces CALM, a training-free framework that leverages paired counterfactual anchors for adaptive prompt routing alongside dual-space minimal corrections in text tokens and denoising residuals.

L2P: Unlocking Latent Potential for Pixel Generation

L2P introduces a resource-efficient Latent-to-Pixel transfer paradigm that discards the VAE in favor of large-patch tokenization, freezes intermediate DiT backbone layers, and trains only shallow layers with LDM-generated synthetic data on 8 GPUs, preserving rich semantic priors while natively unlocking 4K ultra-high-resolution generation.

Learning a Maximum Entropy Model for Visual Textures using Diffusion

This paper introduces the first principled framework for unsupervised learning of visual texture statistics within a maximum entropy probability model via diffusion denoising, generating texture samples comparable or superior to the state-of-the-art Gatys model while reducing statistics from ~177k to 512.

Learning Domain Trajectories with Flow Matching for Gradual Domain Adaptation

FlowGDA reformulates gradual domain adaptation as continuous transport on a representation-level statistical manifold via conditional flow matching, replacing unstable adversarial training with simulation-free velocity-field regression and eliminating error accumulation through source-anchored semantic consistency and two-stage curriculum optimization.

Learning to Generate Multiple Objects from Dense and Occluded Layouts

Addressing catastrophic instance merging and count failures in text-to-image diffusion under dense overlapping layouts, this paper introduces a depth-ordered ownership-aware layout attention bias alongside an amodal-aware instance-balanced loss (AIBL) that dynamically amplifies gradients for heavily occluded objects without adding inference-time mask overhead.

Learning to Sample From Diffusion Models via Inverse Reinforcement Learning

By framing diffusion sampling with a frozen denoiser as a discrete-time finite-horizon MDP, this paper introduces an inverse reinforcement learning framework that directly matches state occupancy measures via f-divergence minimization, replacing costly hyperparameter grid search with a single training run.

Learning Unbiased Permutations via Flow Matching

PermFlow establishes conditional flow matching directly on the affine subspace of matrices with unit row and column sums, employing an algebraic closed-form tangent-space projector to maintain strict constraints along trajectories and a nearest-target coupling to eliminate entropic bias and recover multimodal permutation distributions.

LoRAtorio: An intrinsic approach to LoRA Skill Composition

Low-Dimensional Adaptation of Rectified Flow: A Diffusion and Stochastic Localization Perspective

By establishing an equivalence between Rectified Flow (RF) and diffusion models via stochastic localization, this work introduces a theoretically grounded U-shaped non-uniform time schedule and an adaptive stochastic sampler, proving for the first time that RF adapts to the intrinsic low dimensionality \(k\) of data to achieve \(\tilde{\mathcal{O}}(k/\varepsilon)\) iteration complexity.

MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation

MaskSense identifies and tackles the "visual exploration trap" in masked image generation, where confidence-based greedy decoding prematurely collapses subject exploration into background textures; by combining semantic anchored routing sampling with global entropy transition anchoring, it focuses policy updates on decisive steps and pushes GenEval accuracy from 0.54 to 0.86.

MegaStyle: Constructing Diverse and Scalable Style Dataset via Consistent Text-to-Image Style Mapping

MegaStyle leverages the consistent text-to-image style mapping capability of advanced generative models to curate MegaStyle-1.4M (170K style prompts across 1.4M images), enabling style-supervised contrastive learning for the MegaStyle-Encoder and paired DiT training for the MegaStyle-FLUX transfer model.

Metropolis-Adjusted Diffusion Models

Addressing the non-vanishing time-discretization bias introduced by Unadjusted Langevin Algorithm (ULA) corrector steps in diffusion model sampling, this paper proposes Metropolis-Adjusted Diffusion Models (MADM), reformulating the intractable target density ratio as a line integral of the score function and presenting both an exact two-coin Bernoulli factory correction and an efficient \(\mathcal{O}(h^{5/2})\) Simpson's rule approximation.

Midpoint Generative Models

Midpoint Generative Models (MGM) leverages the midpoint symmetry of Flow Matching where the drift field vanishes for identical endpoints, introducing random time flips and symmetric stochastic interpolants to construct a definite generalized Midpoint Divergence and a tractable variational minimax objective for training one-step generators directly from data without teachers.

MIND: Monge Inception Distance for Generative Models Evaluation

Addressing the sample inefficiency and vulnerability to moment-matching attacks of the classical FID caused by high-dimensional covariance estimation, this paper proposes the Monge Inception Distance (MIND) based on sliced Wasserstein distance, which matches FID-50k discriminative fidelity using only 5k samples while running two orders of magnitude faster.

MiniDog: Dog-Breed Text-to-image Flow Matching from Pretraining to Fine-Tuning on 4 Consumer GPUs

MiniDog establishes an efficient text-to-image flow matching pipeline that pretrains a 22M LightningDiT in 55 minutes and fine-tunes it in 15 minutes on 4 consumer RTX 3090 GPUs via cached representations, E2E-INVAE, and iREPA with EUPE features.

Multiscale Supervised Unbalanced Optimal Transport Flow Matching

Proposes Multiscale Supervised Unbalanced Optimal Transport Flow Matching (MUST-FM), a simulation-free dynamic UOT framework that leverages hierarchical annotations and biological priors to infer cellular trajectories and mass variations, eliminating fine-level coupling computation via within-block independent lifting.

NAMVIS: Next-Scale Autoregressive Multi-View Image Synthesis

NAMVIS introduces the first diffusion-free, geometry-conditioned next-scale autoregressive framework for sparse-view multi-view image synthesis, leveraging multi-scale projective pose encoding and dual-path conditioning to generate geometrically consistent target views over \(3\times\) faster than representative diffusion baselines.

On the Limits of Latent Reuse in Diffusion Models

This paper rigorously characterizes the irreducible score-estimation error lower bounds of freezing and reusing source latent representations under distribution shift in diffusion models, proves that the error is governed by principal-angle subspace misalignment and time-amplified ambient noise, and establishes the necessary geometric conditions for mixed representation learning to eliminate signal truncation mismatch.

On the Robustness of Watermarking for Autoregressive Image Generation

This paper systematically assesses the adversarial vulnerability of watermarking schemes for discrete and bitwise autoregressive image generators, introducing three novel attacks (VQ-Regen, LatentOpt, and Frequency Injection) that achieve effective removal and single-reference forgery without model parameters or secret keys, and proving that adjusting detection thresholds cannot resolve the forgery dilemma.

One View Is Enough: In-the-Wild Monocular Pretraining for Novel View Generation

OVIE uses a frozen depth model to turn 30M single images into partially valid pseudo-view training pairs, learns observed regions through masked reconstruction and perceptual losses, and completes unknown regions through a full-image adversarial prior, ultimately generating a novel view from only a source image and target pose in one forward pass at 116 FPS on an H100, with multi-view fine-tuning improving in-domain performance.

Opening the Black Box of Classifier-free Guidance via Information Bottleneck

This paper frames the diffusion sampling trajectory as a time-dependent information bottleneck, proving that conditional guidance becomes effective only after a critical Fisher-information balance stage, and introduces a training-free adaptive CFG activation rule driven by online score statistics.

Optimal Convergence Analysis of DDPM for General Distributions

This paper establishes a sharp convergence theory for the DDPM sampler under a relaxed non-uniform Lipschitz condition, proving a total variation convergence rate of \(\widetilde{O}(\frac{\sqrt{d}\min\{L,\sqrt{d}\}}{T})\) (which reaches \(\widetilde{O}(\frac{\sqrt{d}}{T})\) for common distributions where \(L \le \mathrm{poly}(\log(dT))\)), demonstrating for the first time that DDPM achieves the same \(\sqrt{d}\) dimension dependence as deterministic DDIM alongside matching theoretical lower bounds.

ORCA: Hunting Compositional Failures in Text-to-Image Diffusion

ORCA introduces a training-time auxiliary alignment loss that dynamically parameterizes an orthonormal projection basis via the residual between T5 and CLIP text embeddings, aligning diffusion latents with a frozen low-rank DINOv2 PCA target to eliminate compositional binding failures at zero inference cost.

OverLay++: Dense-Overlap Layout-to-Image Generation Dataset

Addressing the frequent failure of Layout-to-Image (L2I) generation in dense occlusion and complex interaction scenes, this paper introduces OverLay++, a large-scale dataset of ~500K images curated via MLLM prompting, local bounding-box overlap induction, patch-crop closed-loop verification, and dual-granularity hierarchical captioning, increasing object density to 6.6 objects/image (a 1.67× gain), boosting nearest-neighbor IoU to 0.31, and substantially enhancing spatial alignment, attribute fidelity, and training convergence across state-of-the-art L2I models.

Panoptic Scene Program Diffusion Transformer

PSP-DiT turns a panoptic scene program containing instances, attributes, relations, and counts into a latent variable jointly denoised with the image, uses ownership and cycle-consistency supervision to constrain its visual realization, and improves GenEval 2 from 32.8 to 38.2 over a matched internal baseline while reporting approximately 1.12 times the inference latency.

Parallel Rollout Approximation for Pixel-Space Autoregressive Image Generation

To resolve high single-step prediction error and train–inference exposure bias in pixel-space continuous autoregressive generation, this paper proposes Parallel Rollout Approximation (PRA), which learns compact context-aware intermediate states end-to-end and reconstructs inference-like pixel inputs in parallel, establishing a new state of the art of 1.94 FID on ImageNet-256 while preserving a native pixel-in, pixel-out interface.

ParetoSlider: Diffusion Models Post-Training for Continuous Reward Control

ParetoSlider introduces a multi-objective reinforcement learning post-training framework for flow-matching diffusion models that conditions a single network on continuous preference vectors with per-reward late-scalarization advantage normalization, enabling smooth real-time traversal of the complete Pareto front at inference time without maintaining multiple checkpoints.

Path-independent Flow Matching for Multi-parameter Generative Dynamics

Path-independent Flow Matching (PiFM) extends continuous-time flow matching to multi-parameter probability paths over high-dimensional temporal parameter domains, enforcing path-independent and commutative distribution transport via Lie bracket integrability conditions or affine probability paths while recovering Wasserstein barycenters.

PDE-SSM: A Spectral State Space Approach to Spatial Mixing in Diffusion Transformers

Extends 1D state-space models to continuous multi-dimensional domains via a learnable convection-diffusion-reaction partial differential equation solved analytically in the Fourier domain with \(\mathcal{O}(N\log N)\) spatial complexity, replacing self-attention in Diffusion Transformers to achieve superior image generation fidelity with up to \(40\times\) training speedup at small patch sizes.

Permutation-Invariant Spectral Learning via Dyson Diffusion

Addressing the Weisfeiler-Leman blindness of standard graph diffusion models caused by forcing permutation equivariance into neural architectures, this paper proposes the Dyson Diffusion Model (DyDM), leveraging random matrix theory to decouple permutation-invariant spectral dynamics via Dyson Brownian motion and model orthonormal eigenvectors via Lie algebra score matching.

PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion

PiD reformulates latent-to-pixel decoding as conditional pixel diffusion, unifying latent decoding and spatial upsampling into a single generative module, and leverages noise-aware gating and DMD2 distillation to achieve 4-step ultra-fast 2K/4K pixel-level high-quality image reconstruction.

PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers

Addressing the severe quality degradation of conventional block sparse attention caused by truncating unselected blocks under high sparsity, PISA introduces a training-free piecewise sparse attention framework that combines exact computation for critical blocks with block-wise Taylor expansion for unselected blocks within a unified online Softmax, delivering 1.2×–2.57× lossless end-to-end acceleration across Wan2.1, Hunyuan-Video, and FLUX.

PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining

PixelART dispenses with latent autoencoders and pretrained text-to-image foundation models, demonstrating that image-to-layer decomposition can be formulated as a pixel-space assignment-and-completion inverse problem and solved with a from-scratch rectified-flow Transformer paired with terminal-boosted timestep sampling.

PixelDense: Dense Prediction as Representation Alignment for Pixel Diffusion

PixelDense resolves negative-transfer interference when aligning pixel diffusion transformers with multi-task dense foundation models by decomposing teachers into factored semantic (DINOv2, SAM2) and geometric (Depth Anything v2, Metric3D v2) projection streams regularized by weight-space orthogonality, achieving superior compositional generation and geometry preservation at zero inference overhead.

PixelDiT2: Representation-Grounded Pixel Diffusion Transformers

PixelDiT2 re-encodes the current noisy image at each pixel-denoising evaluation, using frozen DINO patch representations and a timestep adapter as spatial conditioning without an autoencoder, achieving FID 1.46 on ImageNet-256 after 600 epochs and 1.48 on ImageNet-512 after 680 epochs.

Precise: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models

Addressing unstable exploration and discretization bias in online RL post-training of flow-matching models, Precise couples a logSNR-derived exploration schedule with an SDE-consistent closed-form transition under a frozen clean-latent posterior mean, eliminating excess noise and cutting training time by 13.1%–53.2%.

Preconditioned Flow Matching

By identifying an optimization bottleneck where ill-conditioned intermediate covariances trap gradient descent on low-variance directions, this paper proposes a precondition-then-match framework that maps data into an isotropic representation before training the flow, substantially improving convergence and generation fidelity across multiple resolutions.

Predict-Project-Renoise: Sampling Diffusion Models under Hard Constraints

Addressing the failure of pretrained diffusion models to satisfy strict physical laws and exact observations, this paper derives a corrector transition kernel whose unique stationary distribution is the constrained marginal at each noise level, introducing Predict-Project-Renoise (PPR)—a training-free algorithm that projects through the denoiser and renoises to achieve over 80% feasibility while preserving high data-manifold fidelity.

Primal-Dual Flow Matching for Sample-Wise Constrained Generation

Addressing the challenge of enforcing sample-wise constraints in continuous generative models when constraints are non-differentiable black boxes, Primal-Dual Flow Matching (PDFM) provides provable Total Variation convergence guarantees and direct probability control using only binary membership oracle feedback via a spatiotemporal surrogate discriminator.

Probing Visual Planning in Image Editing Models

This paper introduces the Editing-as-Reasoning (Ear) paradigm and the procedural Amaze benchmark, systematically probing how image editing models compress visual planning into atomic single-step transformations across continuous maze navigation and discrete combinatorial Queen puzzles.

ProEdit: Inversion-based Editing From Prompts Done Right

To tackle editing failures in training-free inversion-based visual editing caused by excessive source prior injection, ProEdit introduces Latents-Shift to eliminate source bias in inverted latents and region-adaptive KV-mix to decouple edited areas from backgrounds without manual selection of layers or heads.

RA-ClipScore: Making Generative Model Evaluation More Interpretable

Addressing the lack of diagnostic interpretability in traditional generative metrics and the competition/spatial blindness inherent in CLIP's contrastive pretraining, RA-CLIPScore introduces training-free dual prompts to decouple attributes alongside penultimate layer patch tokens to capture localized semantics, establishing the first fine-grained spatial and attribute divergence benchmark in a single forward pass.

RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution

RankE addresses the latent covariate shift in discrete autoregressive T2I post-training by co-evolving the autoregressive policy and the VQ decoder via alternating ranking optimization under a Generalized EM framework, simultaneously breaking the fidelity-alignment trade-off.

RaPD: Resolution-Agnostic Pixel Diffusion via Semantics-Enriched Implicit Representations

RaPD establishes text-to-image pixel diffusion directly within a continuous Neural Image Field (NIF) latent space, utilizing Semantic Representation Guidance and a Coordinate-Queried Attention Renderer to synthesize arbitrary-resolution images with fixed diffusion cost.

Recursively Trained Diffusion Models: Limiting Collapse Distribution and Spectral Characterization

This paper develops an operator-theoretic framework proving that even with perfect score learning and exact sampling, early stopping of the reverse SDE drives recursive diffusion models to geometrically converge to a unique infinite Gaussian-smoothed mixture limit, and proves that annealed truncation schedules asymptotically eliminate this recursive collapse.

Representation Fréchet Loss for Visual Generation

By decoupling the population size required for statistical estimation of Fréchet Distance from the batch size used for back-propagation, this work introduces representation-space FD-loss as a scalable post-training objective that substantially improves one-step visual generators and repurposes multi-step denoisers into competitive one-step models without teacher distillation or adversarial training.

Residual Calibration via Local Feature-Space Refinement

Introduces Local Adaptive Probability Estimation (LAPE), a post-hoc calibration framework that treats global calibrator outputs as stable anchors and applies non-parametric local residual corrections in the feature space, boosted by \(V\)-fold cross-fitting and Cascade Feature Augmentation Fusion (CFAF) for multi-view scenarios.

Rethinking Cross-Layer Information Routing in Diffusion Transformers

DAR replaces incremental residual addition in diffusion Transformers with timestep-aware attention over historical sublayer outputs, achieving an unguided ODE FID of 7.56 after 600K steps on ImageNet 256×256, a 2.11 improvement over the SiT baseline trained for 1.75M steps, while remaining compatible with REPA's representation-alignment loss.

Revisiting Diffusion Fine-Tuning for Unsupervised Domain Adaptation

MUSE separates diffusion-generator adaptation into a shared semantic branch trained only with source-label supervision and a target-conditioned style branch, allowing one source-involved fine-tuning process to generate bridge data for multiple targets; it improves average downstream UDA accuracy on three closed-set classification benchmarks while reducing generator fine-tuning time to roughly one-half to two-fifths of Terra's cost.

Revisiting Diffusion Model Predictions Through Dimensionality

By analyzing the decoupled training dynamics of a minimal linear diffusion model across tangent and normal subspaces, this paper derives the analytical optimal prediction target \(k^* = D / (D + d)\) for flow matching, and proposes \(k\)-Diff—an end-to-end framework that learns the optimal prediction target via a single trainable scalar.

Reward Score Matching: Unifying Reward-based Fine-tuning for Flow and Diffusion Models

The paper proposes Reward Score Matching (RSM), a unifying score-matching framework that reframes fragmented reward-based fine-tuning paradigms into regression against an optimal value-guided score target, explaining existing heuristics via bias-variance-compute tradeoffs and delivering redesigns that cut training cost by up to 5x.

Robust Diffusion Models via Divergence-Induced Weighted Denoising

By proving that the per-step likelihood ratio under Gaussian reverse kernels follows a lognormal distribution parameterized by a scalar denoising mismatch, this paper reformulates diffusion training into divergence-induced weighted denoising that achieves superior robustness against corrupted data with negligible computational overhead.

Root-Selecting Fixed-Point Inversion for Rectified Flows via Trajectory Straightness

SelFix resolves the multi-root ambiguity in rectified flow fixed-point inversion by adopting trajectory straightness as a principled root-selection criterion, combining straightness-anchored Halpern iteration with decoupled momentum to achieve state-of-the-art exact reconstruction and source-preserving editing.

RoPEMover: Depth-Aware Object Relocation via Positional Embeddings

RoPEMover enables single-image object relocation by warping 2D rotary position embeddings (RoPE) and injecting metric depth into the unused temporal axis of diffusion transformers, achieving state-of-the-art geometric consistency, accurate occlusion handling, and physical shadow updates with minimal supervision.

SANEval: Open-Vocabulary Compositional Benchmarks with Failure-mode Diagnosis

SANEval addresses the closed-vocabulary bottleneck and lack of interpretability in text-to-image compositional evaluation by decoupling assessment into spatial, attribute, and numeracy axes via LLM synonym expansion, YOLO-E open-world detection, and VLM crop-level judging, producing both continuous adherence scores and structured diagnostic feedback.

Scalable Minimal-Change Learning for Controllable Image Editing

ARRO translates the causal principle of minimal change into an outcome-level reinforcement learning objective for instruction-based image editing, employing vision-language model chain-of-thought auditing and group rubric refinement to construct decoupled efficacy and locality rewards that prevent unintended visual drift.

SCALE: Improving Guidance-Free Visual Generation via Self-Contrastive Alignment for Likelihood Estimation

Introduces SCALE, a guidance-free visual generation framework that uses a self-contrastive alignment signal between real and self-generated samples to push generation toward the true data manifold via reparameterized likelihood estimation, outperforming CFG-mimicking baselines and guided sampling in a single forward pass.

Search-Augmented Masked Diffusion Models for Constrained Generation

SearchDiff is a training-free inference framework that embeds candidate search sampling and local neighborhood search directly into the reverse denoising steps of masked discrete diffusion models, steering sampling toward feasible discrete manifolds under non-differentiable black-box constraints.

Secure Seed-Based Multi-bit Watermarking for Diffusion Models from First Principles

This paper exposes the pitfalls of evaluating diffusion seed watermarks via empirical generation pipelines, develops an architecture-agnostic theoretical framework decoupling inverse diffusion from decision mechanisms via Shannon capacity, fidelity, and random-matrix security, and introduces SSB, a nested-lattice scheme achieving provably perfect security under non-zero capacity.

SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers

Addressing attention dilution and frequency imbalance when extrapolating Diffusion Transformers beyond their training resolutions, SEGA introduces a training-free framework that dynamically scales individual RoPE dimensions according to the latent's spectral energy, dramatically improving both global structural coherence and fine-detail fidelity.

Sparsely Supervised Diffusion

This paper introduces Sparsely Supervised Diffusion (SSD), a minimal masking strategy in the denoising regression loss that supervises only unmasked pixels (safe up to 98% masking), fundamentally reshaping the covariance spectrum to suppress spurious correlations, stabilize small-dataset training, and eliminate spatial inconsistency artifacts.

SpatialFlow-GRPO: Where Spatial Credit Drives Image Editing

Addressing the spatial uniformity limitation of whole-image rewards in Flow-GRPO post-training, this paper proposes SpatialFlow-GRPO, which introduces region-aware rewards, semantic-group advantage estimation, and latent-region-aligned policy optimization to achieve precise spatial credit assignment and significantly enhance multi-target image editing.

Specificity-Aware Diffusion Steering via Variance-Reduced Sequential Monte Carlo

This paper formulates specificity-aware diffusion guidance as sampling from an overlap-minimizing target distribution with a decoupled powered likelihood-ratio gate, and introduces a Sequential Monte Carlo sampler with a matched backward kernel and fixed-displacement JVP optimization that minimizes proposal weight variance without retraining the underlying model.

Statistical Convergence of Spherical First Hitting Diffusion Models

This paper provides the first theoretical convergence analysis for first hitting diffusion models (FHDM) with random adaptive generation time, proving that sparse ReLU neural networks achieve the minimax optimal total variation convergence rate \(\mathcal{O}(n^{-\alpha/(2\alpha+d-1)})\) (up to logarithmic factors) for Sobolev smooth distributions on the sphere \(\partial B \subset \mathbb{R}^d\).

Statistical Matching via Schrödinger Bridge beyond Conditional Independence

Addressing the fundamental limitation of classical statistical matching that enforces the conditional independence assumption, this paper introduces a dependency-aware Schrödinger bridge framework that exponentially tilts the conservative baseline via a compatibility cost, recovering informative joint distributions with exact Gaussian recovery guarantees and bidirectional probabilistic posterior completion.

Stay Fair! Ensuring Group Fairness in Diffusion Models Across Guidance Scales

This paper exposes guidance bias as an overlooked, scale-dependent driver of demographic disparity in diffusion models, establishes Strong Demographic Parity (SDP) for guidance functions, and introduces StayFair to decouple group fairness from guidance scale tuning across both classifier and classifier-free guidance.

Steering Away from Memorization: Reachability-Constrained Reinforcement Learning for Text-to-Image Diffusion

To prevent text-to-image diffusion models from replicating copyrighted or private training images without sacrificing image fidelity or prompt adherence, RADS models the denoising trajectory as a controlled dynamical system, computes the backward reachable tube of memorization basins, and deploys reachability-constrained reinforcement learning to steer generations safely via minimal caption embedding perturbations.

Steering Fields: Adaptive Vector Fields for Safe Image Generation and Beyond

Addressing the structural distortion and manifold mismatch of static activation steering, this paper introduces Steering Fields—a trajectory-adaptive vector field that re-estimates directional velocities in closed form at each flow matching step to simultaneously induce target concepts and repel harmful ones while enabling inversion-free image editing.

SteerVTE: Seamless Video Text Editing with Style and Glyph Control

SteerVTE steers a frozen pre-trained video diffusion transformer via a lightweight Text Context Adapter combining native-resolution VLM style encoding and dual-granularity glyph priors, optimized via a glyph-aware spatial-focal loss and a progressive three-stage curriculum to achieve seamless, temporally coherent video text editing.

Stochastic Interpolants via Conditional Dependent Coupling

This paper introduces CDC-FM, a unified multi-stage flow matching framework operating in pixel space via Conditional Dependent Coupling (CDC), parameterizing a coarse-to-fine trajectory with a single shared Diffusion Transformer to achieve end-to-end optimization, significantly lower transport costs, and an FID of 1.81 on ImageNet-256.

Structured State-Space Regularization for Generation-Friendly Image Tokenization

This paper introduces structured state-space regularization by conceptualizing state-space models as a basis projection coupled with continuous input transformations; by modeling 2D orthogonal basis coefficient dynamics under continuous Gaussian blur via the heat equation, it endows tokenizer latent channels with clear spectral organization, substantially improving downstream diffusion generation metrics with negligible loss in reconstruction fidelity.

StyleStream 2.0: Fast and Controllable Streaming Voice Style Conversion

StyleStream 2.0 cuts streaming voice style conversion latency by 2x down to 520 ms with 4-step pixel MeanFlow and autoregressive rollout fine-tuning, while removing mel-context conditioning to enable natural language voice design and single-attribute editing via a unified flow-matching style controller.

Temporal Alignment Guidance: On-manifold Sampling in Diffusion Models

To tackle the off-manifold drift and severe fidelity degradation in diffusion models caused by external guidance and accelerated sampling, this paper proposes Temporal Alignment Guidance (TAG), which leverages a lightweight time predictor to evaluate Time-Linked Scores and actively pulls perturbed samples back to the target data manifold at every reverse step.

TerraVis: Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows

TerraVis addresses the critical blind spot where visually photorealistic text-to-image outputs violate real-world physical and anatomical laws by proposing a multi-stage MLLM evaluation workflow over an 18-type hierarchical taxonomy, achieving state-of-the-art alignment with human plausibility judgments.

Tessellations of Semi-Discrete Flow Matching

This paper investigates the intrinsic geometry and topology of terminal assignment cells in semi-discrete Flow Matching (transporting a Gaussian source to a finitely supported target), proving they are open, simply connected, and homeomorphic to open balls, while constructing a planar counterexample that demonstrates non-convexity, curved boundaries, non-monotonicity, and distinct adjacency structures from optimal transport Laguerre cells.

Text-Vision Co-Instructed Image Editing

Addressing the spatial under-specification of textual editing and the semantic blindness of sparse visual dragging, TV-Edit introduces a unified text-vision co-instructed editing framework with a lightweight Content-Aware Spatial Controller, trained on a newly curated 23K paired video dataset to achieve precise geometry control alongside faithful semantic execution.

The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning

This paper resolves the fundamental paradox of noise-agnostic diffusion models by showing they implicitly execute Riemannian gradient flow on a marginal energy landscape, proving that velocity-based parameterization guarantees sampling stability via bounded gain while noise-prediction models suffer catastrophic divergence through an amplified Jensen Gap.

The Interplay of Data Structure and Imbalance in the Learning Dynamics of Diffusion Models

By analyzing random feature score matching on Gaussian mixtures via high-dimensional random matrix theory, this paper proves that class variance primarily dictates generalization and memorization schedules (consistently favoring higher-variance classes), with centroid geometry playing a secondary role and sampling imbalance acting as a modulator that can invert learning order and delay backward speciation.

The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL

Reveals that flow matching fails to capture perceptual data realism due to train-test distribution shifts and \(\ell_2\) velocity geometry mismatch, and proves that discriminator logit in a frozen self-supervised feature space provides the optimal density-ratio reward for KL-regularized RL via adjoint matching.

The Unreasonable Effectiveness of Text Embedding Interpolation for Continuous Image Steering

This paper introduces a training-free framework for continuous and controllable visual generation by extracting debiased difference-of-means steering vectors from LLM-constructed contrastive prompt pairs and adaptively determining smooth slider bounds via an elastic band search algorithm.

The Value of Covariance Matching in Gaussian DDPMs and the Lanczos Sampler

The paper theoretically proves that matching the full posterior covariance breaks the \(\Omega(1/T)\) path-space KL divergence barrier of isotropic/diagonal Gaussian DDPMs to \(\mathcal{O}(1/T^2)\), and introduces the Lanczos Gaussian Sampler (LGS)—a training-free, matrix-free sampler that samples from the full covariance via Jacobian-vector products with exponential convergence in iteration steps.

Think, Plan, Paint: Layout-Aware Reasoning for Controllable Image Generation in Unified Models

ATLAS establishes a unified multimodal framework around a human-like "Think, Plan, Paint" paradigm, employing layout as a shared cognitive bridge and leveraging Flow-GRPO reinforcement learning layout alignment to achieve state-of-the-art controllable image synthesis and interactive editing.

Thinking in Pictures: A Systematic Benchmark for Reasoning-driven Image Generation

Addressing the failure of modern generative models to reason beyond surface-level text rendering, this paper establishes RIG-Bench—the first systematic benchmark requiring models to induce latent rules purely from visual context across 4 task families and 11 subtasks (2,000 items) and synthesize solutions directly as images—uncovering a severe reasoning-generation gap where even top models produce locally plausible but globally illogical outputs.

TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment

TMPO identifies that scalar reward maximization in diffusion reinforcement learning fundamentally minimizes reverse KL divergence and inherently induces mode collapse; to overcome this, it introduces Softmax Trajectory Balance (Softmax-TB) to match full-trajectory policy distributions to a Boltzmann reward target—inheriting the mode-covering guarantees of forward KL—alongside Dynamic Stochastic Tree Sampling that amortizes denoising compute, improving generative diversity by 9.1% while reducing iteration training time by up to 27%.

Token-to-Token Alignment of Text Embeddings for Semantic Blending

This paper introduces Token-to-Token (T2T) alignment, a training-free framework establishing explicit token-wise semantic correspondences across prompts via two stages—concept-level structural reshaping and embedding-level soft projection—enabling smooth and coherent continuous image editing, blending, and concept synthesis through simple linear interpolation.

Total Variation Rates for Riemannian Flow Matching

This paper establishes the first non-asymptotic Total Variation (TV) convergence bounds for Riemannian Flow Matching (RFM) with Euler discretization and learned velocity fields, proving polynomial iteration complexities on both compact spheres and non-compact Hadamard manifolds.

Towards More General Control of Diffusion Models Using Jeffrey Guidance

This paper introduces Jeffrey guidance, a principled and training-free diffusion control framework founded on Jeffrey's rule of conditioning that updates feature marginal distributions to prescribed targets while preserving conditional structure, enabling substantial FID reductions in Inception space and robust statistical attribute decorrelation for fair generation.

Training-Free Generative Sampling via Moment-Matched Score Smoothing

Proposes Moment-Matched Score-Smoothed Overdamped Langevin Dynamics (MM-SOLD), a training-free interacting particle sampler that enforces empirical mean and covariance constraints on a centered Stiefel manifold, eliminating barycentric collapse in score smoothing and enabling high-fidelity, diverse generative sampling on CPUs.

Training Data Attribution in Diffusion Models via Mirrored Unlearning and Noise-Consistent Skew

Addressing the fragility and computational bottlenecks of training data attribution in diffusion models, MUCS formulates mirrored unlearning as bounded gradient ascent regularized by training set fine-tuning against a null-loss ceiling, and scores sample influence via noise-consistent normalized loss skew, outperforming existing baselines across leave-\(k\)-out counterfactual benchmarks while running 1.5–2.9\(\times\) faster.

Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation

By identifying blind stochastic jumps at RK-4 midpoints as the core bottleneck in discrete flow distillation, TS-DFM introduces a training-only, time-unconditioned energy compass to navigate midpoint trajectories via a sequence-to-token policy, achieving 8-step generation that beats the 1024-step teacher with zero inference overhead.

UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction

Addressing the architectural imbalance of isotropic Diffusion Transformers and the self-attention disruption in existing U-Net DiTs caused by fixed local downsampling, UDT introduces a parameter-free, data-adaptive token merging and unmerging U-Net framework that preserves hidden token dimensions while slashing computational cost and accelerating training convergence.

UltraDiff: Transferring High-Fidelity Priors to Compressed Latent Spaces for High-Resolution Image Generation

UltraDiff addresses the detail degradation and semantic ambiguity of aggressive spatial downsampling (\(f32\)) by training a dual-branch MixTeacher over paired \(f8\)–\(f32\) latents and distilling its marginal vector field directly into native \(f32\) space, achieving \(f8\)-level fidelity while generating 2K images in 6.2s (\(40.3\times\) faster).

UltraFlash: Accelerating Megapixel Visual Synthesis

UltraFlash shifts the acceleration bottleneck of high-resolution cascade visual synthesis from diffusion denoising to VAE pixel-space transitions, bypassing intermediate RGB round-trips with a learned Latent Hyperloop shortcut and enabling artifact-free sparse patch VAE decoding via shared GroupNorm statistics, cutting 4K generation latency by 2x while boosting fidelity.

UniCustom: Unified Visual Conditioning for Multi-reference Image Generation

UniCustom establishes a unified visual conditioning paradigm for multi-reference image generation by early-fusing semantic ViT features and appearance-rich VAE features via an identity-preserving linear projection before VLM encoding, bridged with slot-wise binding regularization and a two-stage training scheme to achieve state-of-the-art open-source performance on OmniContext and MICo-Bench.

Uniform Diffusion Models revisited: Leave-One-Out Denoiser and Absorbing State Reformulation

This paper demonstrates that the standard bridge plug-in parameterization of Uniform Diffusion Models (UDMs) minimizes the ELBO towards a leave-one-out posterior rather than the conventional denoising posterior, establishes exact closed-form bidirectional conversions to decouple parameterization from training losses, and introduces an absorbing state reformulation (AUDM/ReAUDM) that achieves MDM-level generative quality while preserving UDM joint laws.

UniVL: Unified Vision-Language Embedding for Spatially Grounded Contextual Image Generation

UniVL unifies spatial layout and per-region semantic instructions into a single visual condition by rendering text labels directly onto image masks as pixels, completely eliminating the standalone text encoder while reducing inference TFLOPs by up to 52% and boosting image fidelity.

Unlocking the Duality between Flow and Field Matching

This paper proves a rigorous constructive bijective duality between probability-driven Conditional Flow Matching (CFM) and physics-geometry-driven forward-only Interaction Field Matching (IFM) at the level of generative dynamics, reveals that generalized IFM possesses higher expressive power beyond CFM, and derives a multi-sample velocity estimation objective for CFM that significantly reduces gradient variance.

Unrestrained Simplex Denoising for Discrete Data. A Non-Markovian Approach Applied to Graph Generation

Addressing the abrupt structural disruptions of discrete diffusion and compounding errors in flow matching, Unrestrained Simplex Denoising (UNSID) introduces an explicit Dirichlet probability path and a non-Markovian resampling mechanism directly on the probability simplex, achieving theoretical error contraction and state-of-the-art graph generation quality without handcrafted features.

Velocity Ambiguity Profiles: Time-Resolved Bayes-Risk Diagnostics for Flow Matching

This paper introduces the Velocity Ambiguity Profile (VAP) as a time-local Bayes-risk floor for flow matching, decomposes time-resolved loss into intrinsic ambiguity and reducible excess error, and leverages frozen excess-risk proxies to guide training time resampling.

VTV-FM: Flow Matching through Variational Terminal-Velocity Closure

VTV-FM establishes a second-order phase-space Flow Matching framework by variationally eliminating missing terminal velocities via minimum-acceleration action, thereby unifying closed-form terminal velocity closure, optimal transport pairing, and acceleration field supervision.

What Time Is It? How Data Geometry Makes Time Conditioning Optional for Flow Matching

By decomposing the time-blind flow matching loss into learnable error, coupling variance, and time-blindness gap, this paper proves that high-dimensional data geometry allows time to be identified from the noise subspace at an \(O_p(1/\sqrt{d-k})\) rate, making the time-blindness gap asymptotically negligible relative to the coupling variance and showing that sample coupling is the dominant performance lever.

When Riemann flows with Wasserstein: Generative Modeling of Probability Distributions on Manifolds

Riemannian Wasserstein Entropic Flow Matching (RWEFM) lifts conditional flow matching to the quadratic Wasserstein space \(\mathcal{P}_2(\mathcal{M})\) of a Riemannian manifold, using GPU-scalable Riemannian Entropic Maps to generate high-fidelity probability distributions on curved geometries.

Where to Look Is Not How to Fix: Pre-Denoising Diagnostics and Modality-Dependent Control in Diffusion Composition

Through a rigorously controlled anchor-stress evaluation protocol, this work uncovers a diagnosis-control dissociation in diffusion models: compositional stress is detectable before denoising via a non-factorization residual index (CSI) and peaks at the deep encoder (D2), yet causal semantic repair concentrates downstream at decoder cross-attention blocks.

Where, What, Why, and Importance: Structured Defect Grounding for Text-to-Image Feedback

Structured Defect Grounding (SDG) reformulates dense text-to-image failure diagnosis from continuous heatmap regression into variable-cardinality set prediction over (location, type, reason, importance) tuples, utilizing BoxFlow-GRPO to enable spatially localized, importance-weighted diffusion policy alignment without reward hacking.

Which Tokens to Merge? Diffusion Dynamics for Efficient Image Generation

Introduces Dynamics-aware Token Merging (DaTMe), a training-free framework that leverages Tweedie clean-space inter-step residuals and spatial high-frequency residuals to adaptively partition tokens into protected structural anchors and safe merge sources, resolving CFG's prompt-centric bias and maintaining image-wide fidelity on PixArt-\(\alpha\) and FLUX.

X2HDR: HDR Image Generation in a Perceptually Uniform Space

X2HDR reformulates high-dynamic-range (HDR) image generation by mapping data into a perceptually uniform space (PU21), proving that off-the-shelf, frozen LDR VAEs reconstruct HDR content with near-LDR fidelity, thereby enabling text-to-HDR synthesis and single-RAW HDR reconstruction via simple, parameter-efficient LoRA finetuning.

You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning

Formalizing concepts as overlapping geometric regions in cross-attention activation space, this paper proves a theoretical lower bound on collateral damage during diffusion unlearning (\(\gamma \ge \kappa(1-\varepsilon)/L\)), demonstrating that concept leakage and collateral forgetting are unavoidable dual symptoms of representational entanglement rather than algorithm-specific flaws.