arXiv subjects

Guandong Li

Publications and source records attributed to Guandong Li.

At least 19 recordsLinked to original sources

arXiv2026

AttnRouter: Per-Category Attention Routing for Training-Free Image Editing on MMDiT

We study training-free image editing on Qwen-Image-Edit-2511, a 60-block multi-modal diffusion transformer (MMDiT) that concatenates noise and source-image tokens within a single attention stream. We make three contributions. (i) We introduce KVInject, a single-forward attention manipulation that alpha-blends source-half key/value projections into the noise-half within a localized layer/step band. KVInject is simpler than the classical two-pass MasaCtrl recipe and avoids the prompt-mismatch failure mode that disables MasaCtrl on MMDiT (composite score drops 31% versus baseline). (ii) We show that no single attention operation dominates across edit types, motivating AttnRouter, a per-category routing table that dispatches edits to the operation that best preserves source structure for that type. With ground-truth categories the router improves the CLIP-T+DINO-I composite by 6.4% over the editing baseline; an automatic CLIP zero-shot classifier closes 98% of this gap despite only 55% category accuracy. (iii) Through layer-, step-, and alpha-band ablations we localize the editing-effective attention sub-circuit: K/V injection in early denoising steps (S0-7) recovers nearly all of the gain of full-step injection, while injection in early (L0-15) or late (L45-60) layer bands fails to drive editing entirely; alpha in [0.3, 0.5] is a stable sweet spot. We also report negative results that highlight what does not transfer from the UNet folklore: simple K/V rescaling never beats baseline and aggressive variants collapse generation entirely (composite 0.084). We release code, pre-computed routing tables, and a 100-sample stratified subset of ImgEdit-Bench used in all ablations.

cs.CV↗

arXiv2026

PhysEdit: Physically-Consistent Region-Aware Image Editing via Adaptive Spatio-Temporal Reasoning

Image editing instructions are heterogeneous: a color swap, an object insertion, and a physical-action edit all demand different spatial coverage and different reasoning depth, yet existing reasoning-based editors apply a single fixed inference recipe to every instruction. We argue that adaptivity along both the spatial and temporal axes is the missing degree of freedom, and we present PhysEdit, an editing framework built around this principle. PhysEdit introduces two inference-time modules that compose without retraining the backbone. At its core, (1) Complexity-Adaptive Reasoning Depth (CARD) predicts edit complexity directly from the instruction and reference image and allocates the reasoning step count N_r and reasoning-token length r per sample -- turning a previously fixed inference schedule into a conditional-computation problem. CARD is supported by (2) a Spatial Reasoning Mask (SRM) that extracts an instruction-conditioned spatial prior from cross-attention to confine reasoning to regions that semantically require it. On the full 737-case ImgEdit Basic-Edit Suite, PhysEdit delivers a 1.18x wall-clock speedup (64.3s vs. 76.1s per sample) over a strong reasoning baseline while slightly improving instruction adherence (CLIP-T 0.2283 vs. 0.2266, +0.7%) and matching identity preservation within noise (CLIP-I 0.8246 vs. 0.8280). The speedup is category-dependent and reaches 1.52x on appearance-level edits, validating CARD's adaptive allocation as the principal source of efficiency gain. A 30-sample pilot with full ablations isolates the contribution of each module.

cs.CV↗

arXiv2026

Edit Fidelity Field: Semantics-Aware Region Isolation for Training-Free Scene Text Editing

Scene text editing (STE) has achieved remarkable progress in accurately rendering target text through diffusion-based methods. However, we identify a critical yet overlooked problem: edit spillover -- when editing a target text region, existing methods inadvertently modify non-target regions, particularly neighboring text. Through systematic evaluation on 50 real-world scenes across four categories, we reveal that state-of-the-art diffusion editing models exhibit a spillover rate of 94%, meaning nearly all non-target text regions are altered during editing. To address this, we propose the Edit Fidelity Field (EFF), a semantics-aware continuous field that controls per-pixel editing fidelity. Unlike binary masks, EFF leverages OCR-detected text regions to construct a four-zone field: Edit Core (fully editable), Transition Zone (smooth decay), Protected Zone (non-target text, explicitly locked), and Background (strictly preserved). EFF operates as a training-free, model-agnostic post-processing module applicable to any diffusion-based STE method. We further propose per-region spillover quantification, a novel evaluation protocol that measures edit leakage at each non-target text region individually. Experiments demonstrate that EFF reduces spillover rate from 94% to 25% while improving non-target region preservation by +91.4 dB PSNR.

cs.CV↗

arXiv2026

LayerCache: Exploiting Layer-wise Velocity Heterogeneity for Efficient Flow Matching Inference

Flow Matching models achieve state-of-the-art image generation quality but incur substantial inference cost due to iterative denoising through large Transformer networks. We observe that different layer groups within a Transformer exhibit markedly heterogeneous velocity dynamics: shallow layers are highly stable and amenable to aggressive caching, while deep layers undergo large velocity changes that demand full computation. Existing caching methods, however, treat the entire Transformer as a monolithic unit, applying a single caching decision per timestep and thus failing to exploit this heterogeneity. Based on this finding, we propose LayerCache, a layer-aware caching framework that partitions the Transformer into layer groups and makes independent, per-group caching decisions at each denoising step. LayerCache introduces an adaptive JVP span K selection mechanism that leverages per-group stability measurements to balance estimation accuracy and computational savings. We formulate a three-dimensional scheduling problem over timesteps, layer groups, and JVP span, and solve it with a greedy budget allocation algorithm. On Qwen-Image (1024x1024, 50 steps), LayerCache achieves PSNR 37.46 dB (+5.38 dB over MeanCache), SSIM 0.9834, and LPIPS 0.0178 (a 70% reduction over MeanCache) at 1.37x speedup, dominating all prior caching methods on the quality-speed Pareto frontier.

cs.CV↗

arXiv2026

AdaEdit: Adaptive Temporal and Channel Modulation for Flow-Based Image Editing

Inversion-based image editing in flow matching models has emerged as a powerful paradigm for training-free, text-guided image manipulation. A central challenge in this paradigm is the injection dilemma: injecting source features during denoising preserves the background of the original image but simultaneously suppresses the model's ability to synthesize edited content. Existing methods address this with fixed injection strategies -- binary on/off temporal schedules, uniform spatial mixing ratios, and channel-agnostic latent perturbation -- that ignore the inherently heterogeneous nature of injection demand across both the temporal and channel dimensions. In this paper, we present AdaEdit, a training-free adaptive editing framework that resolves this dilemma through two complementary innovations. First, we propose a Progressive Injection Schedule that replaces hard binary cutoffs with continuous decay functions (sigmoid, cosine, or linear), enabling a smooth transition from source-feature preservation to target-feature generation and eliminating feature discontinuity artifacts. Second, we introduce Channel-Selective Latent Perturbation, which estimates per-channel importance based on the distributional gap between the inverted and random latents and applies differentiated perturbation strengths accordingly -- strongly perturbing edit-relevant channels while preserving structure-encoding channels. Extensive experiments on the PIE-Bench benchmark (700 images, 10 editing types) demonstrate that AdaEdit achieves an 8.7% reduction in LPIPS, a 2.6% improvement in SSIM, and a 2.3% improvement in PSNR over strong baselines, while maintaining competitive CLIP similarity. AdaEdit is fully plug-and-play and compatible with multiple ODE solvers including Euler, RF-Solver, and FireFlow. Code is available at https://github.com/leeguandong/AdaEdit

cs.CV↗

arXiv2026

Edit Spillover as a Probe: Do Image Editing Models Implicitly Understand World Relations?

Instruction-following image editing models are expected to modify only the specified region while keeping the rest of the image unchanged. However, in practice, we observe a pervasive phenomenon -- edit spillover: models alter semantically related but unspecified content outside the edit region. This raises a fundamental question -- does spillover reflect genuine implicit world understanding, or is it merely attention leakage? We propose EditSpilloverProbe, a systematic framework that repurposes edit spillover as a natural probe for world knowledge in image editing models. We introduce a spillover taxonomy (spatial, semantic, mixed, random), an automated detection-and-classification pipeline, and a benchmark dataset constructed from real-world Chinese text editing tasks, EditSpilloverBench. Systematic evaluation of 5 representative editing models reveals three core findings: (1) spillover rates vary dramatically across architectures, from 3.49% to 11.46%, with a 3.3x ratio; (2) absolute semantic spillover quantity reveals models' world understanding capability -- nano_banana produces the most semantic spillover (27.8 per image), while qwen_2511 has the most precise editing control but lower semantic spillover (16.3 per image), revealing a trade-off between editing control and world understanding; (3) spatial decay analysis shows spillover area density decays exponentially with distance, but the proportion of semantically relevant spillover remains constant (40%-58%), providing direct evidence that semantic spillover reflects genuine world understanding rather than spatial diffusion.

cs.CV↗

arXiv2026

Frequency-Aware Error-Bounded Caching for Accelerating Diffusion Transformers

Diffusion Transformers (DiTs) have emerged as the dominant architecture for high-quality image and video generation, yet their iterative denoising process incurs substantial computational cost during inference. Existing caching methods accelerate DiTs by reusing intermediate computations across timesteps, but they share a common limitation: treating the denoising process as uniform across time,depth, and feature dimensions. In this work, we identify three orthogonal axes of non-uniformity in DiT denoising: (1) temporal -- sensitivity to caching errors varies dramatically across the denoising trajectory; (2) depth -- consecutive caching decisions lead to cascading approximation errors; and (3) feature -- different components of the hidden state exhibit heterogeneous temporal dynamics. Based on these observations, we propose SpectralCache, a unified caching framework comprising Timestep-Aware Dynamic Scheduling (TADS), Cumulative Error Budgets (CEB), and Frequency-Decomposed Caching (FDC). On FLUX.1-schnell at 512x512 resolution, SpectralCache achieves 2.46x speedup with LPIPS 0.217 and SSIM 0.727, outperforming TeaCache (2.12x, LPIPS 0.215, SSIM 0.734) by 16% in speed while maintaining comparable quality (LPIPS difference < 1%). Our approach is training-free, plug-and-play, and compatible with existing DiT architectures.

cs.CV↗

arXiv2026

Dual-Channel Attention Guidance for Training-Free Image Editing Control in Diffusion Transformers

Training-free control over editing intensity is a critical requirement for diffusion-based image editing models built on the Diffusion Transformer (DiT) architecture. Existing attention manipulation methods focus exclusively on the Key space to modulate attention routing, leaving the Value space -- which governs feature aggregation -- entirely unexploited. In this paper, we first reveal that both Key and Value projections in DiT's multi-modal attention layers exhibit a pronounced bias-delta structure, where token embeddings cluster tightly around a layer-specific bias vector. Building on this observation, we propose Dual-Channel Attention Guidance (DCAG), a training-free framework that simultaneously manipulates both the Key channel (controlling where to attend) and the Value channel (controlling what to aggregate). We provide a theoretical analysis showing that the Key channel operates through the nonlinear softmax function, acting as a coarse control knob, while the Value channel operates through linear weighted summation, serving as a fine-grained complement. Together, the two-dimensional parameter space $(\delta_k, \delta_v)$ enables more precise editing-fidelity trade-offs than any single-channel method. Extensive experiments on the PIE-Bench benchmark (700 images, 10 editing categories) demonstrate that DCAG consistently outperforms Key-only guidance across all fidelity metrics, with the most significant improvements observed in localized editing tasks such as object deletion (4.9% LPIPS reduction) and object addition (3.2% LPIPS reduction).

cs.CV↗

arXiv2026

Inject Where It Matters: Training-Free Spatially-Adaptive Identity Preservation for Text-to-Image Personalization

Personalized text-to-image generation aims to integrate specific identities into arbitrary contexts. However, existing tuning-free methods typically employ Spatially Uniform Visual Injection, causing identity features to contaminate non-facial regions (e.g., backgrounds and lighting) and degrading text adherence. To address this without expensive fine-tuning, we propose SpatialID, a training-free spatially-adaptive identity modulation framework. SpatialID fundamentally decouples identity injection into face-relevant and context-free regions using a Spatial Mask Extractor derived from cross-attention responses. Furthermore, we introduce a Temporal-Spatial Scheduling strategy that dynamically adjusts spatial constraints - transitioning from Gaussian priors to attention-based masks and adaptive relaxation - to align with the diffusion generation dynamics. Extensive experiments on IBench demonstrate that SpatialID achieves state-of-the-art performance in text adherence (CLIP-T: 0.281), visual consistency (CLIP-I: 0.827), and image quality (IQ: 0.523), significantly eliminating background contamination while maintaining robust identity preservation.

cs.CV↗

arXiv2026

FastUSP: A Multi-Level Collaborative Acceleration Framework for Distributed Diffusion Model Inference

Large-scale diffusion models such as FLUX (12B parameters) and Stable Diffusion 3 (8B parameters) require multi-GPU parallelism for efficient inference. Unified Sequence Parallelism (USP), which combines Ulysses and Ring attention mechanisms, has emerged as the state-of-the-art approach for distributed attention computation. However, existing USP implementations suffer from significant inefficiencies including excessive kernel launch overhead and suboptimal computation-communication scheduling. In this paper, we propose \textbf{FastUSP}, a multi-level optimization framework that integrates compile-level optimization (graph compilation with CUDA Graphs and computation-communication reordering), communication-level optimization (FP8 quantized collective communication), and operator-level optimization (pipelined Ring attention with double buffering). We evaluate FastUSP on FLUX (12B) and Qwen-Image models across 2, 4, and 8 NVIDIA RTX 5090 GPUs. On FLUX, FastUSP achieves consistent \textbf{1.12$\times$--1.16$\times$} end-to-end speedup over baseline USP, with compile-level optimization contributing the dominant improvement. On Qwen-Image, FastUSP achieves \textbf{1.09$\times$} speedup on 2 GPUs; on 4--8 GPUs, we identify a PyTorch Inductor compatibility limitation with Ring attention that prevents compile optimization, while baseline USP scales to 1.30$\times$--1.46$\times$ of 2-GPU performance. We further provide a detailed analysis of the performance characteristics of distributed diffusion inference, revealing that kernel launch overhead -- rather than communication latency -- is the primary bottleneck on modern high-bandwidth GPU interconnects.

cs.CV↗

arXiv2026

FlexID: Training-Free Flexible Identity Injection via Intent-Aware Modulation for Text-to-Image Generation

Personalized text-to-image generation aims to seamlessly integrate specific identities into textual descriptions. However, existing training-free methods often rely on rigid visual feature injection, creating a conflict between identity fidelity and textual adaptability. To address this, we propose FlexID, a novel training-free framework utilizing intent-aware modulation. FlexID orthogonally decouples identity into two dimensions: a Semantic Identity Projector (SIP) that injects high-level priors into the language space, and a Visual Feature Anchor (VFA) that ensures structural fidelity within the latent space. Crucially, we introduce a Context-Aware Adaptive Gating (CAG) mechanism that dynamically modulates the weights of these streams based on editing intent and diffusion timesteps. By automatically relaxing rigid visual constraints when strong editing intent is detected, CAG achieves synergy between identity preservation and semantic variation. Extensive experiments on IBench demonstrate that FlexID achieves a state-of-the-art balance between identity consistency and text adherence, offering an efficient solution for complex narrative generation.

cs.CV↗

arXiv2025

DVI: Disentangling Semantic and Visual Identity for Training-Free Personalized Generation

Recent tuning-free identity customization methods achieve high facial fidelity but often overlook visual context, such as lighting, skin texture, and environmental tone. This limitation leads to ``Semantic-Visual Dissonance,'' where accurate facial geometry clashes with the input's unique atmosphere, causing an unnatural ``sticker-like'' effect. We propose **DVI (Disentangled Visual-Identity)**, a zero-shot framework that orthogonally disentangles identity into fine-grained semantic and coarse-grained visual streams. Unlike methods relying solely on semantic vectors, DVI exploits the inherent statistical properties of the VAE latent space, utilizing mean and variance as lightweight descriptors for global visual atmosphere. We introduce a **Parameter-Free Feature Modulation** mechanism that adaptively modulates semantic embeddings with these visual statistics, effectively injecting the reference's ``visual soul'' without training. Furthermore, a **Dynamic Temporal Granularity Scheduler** aligns with the diffusion process, prioritizing visual atmosphere in early denoising stages while refining semantic details later. Extensive experiments demonstrate that DVI significantly enhances visual consistency and atmospheric fidelity without parameter fine-tuning, maintaining robust identity preservation and outperforming state-of-the-art methods in IBench evaluations.

cs.CV↗

arXiv2025

EditIDv2: Editable ID Customization with Data-Lubricated ID Feature Integration for Text-to-Image Generation

We propose EditIDv2, a tuning-free solution specifically designed for high-complexity narrative scenes and long text inputs. Existing character editing methods perform well under simple prompts, but often suffer from degraded editing capabilities, semantic understanding biases, and identity consistency breakdowns when faced with long text narratives containing multiple semantic layers, temporal logic, and complex contextual relationships. In EditID, we analyzed the impact of the ID integration module on editability. In EditIDv2, we further explore and address the influence of the ID feature integration module. The core of EditIDv2 is to discuss the issue of editability injection under minimal data lubrication. Through a sophisticated decomposition of PerceiverAttention, the introduction of ID loss and joint dynamic training with the diffusion model, as well as an offline fusion strategy for the integration module, we achieve deep, multi-level semantic editing while maintaining identity consistency in complex narrative environments using only a small amount of data lubrication. This meets the demands of long prompts and high-quality image generation, and achieves excellent results in the IBench evaluation.

cs.CV↗

arXiv2025

MVNet: Hyperspectral Remote Sensing Image Classification Based on Hybrid Mamba-Transformer Vision Backbone Architecture

Hyperspectral image (HSI) classification faces challenges such as high-dimensional data, limited training samples, and spectral redundancy, which often lead to overfitting and insufficient generalization capability. This paper proposes a novel MVNet network architecture that integrates 3D-CNN's local feature extraction, Transformer's global modeling, and Mamba's linear complexity sequence modeling capabilities, achieving efficient spatial-spectral feature extraction and fusion. MVNet features a redesigned dual-branch Mamba module, including a State Space Model (SSM) branch and a non-SSM branch employing 1D convolution with SiLU activation, enhancing modeling of both short-range and long-range dependencies while reducing computational latency in traditional Mamba. The optimized HSI-MambaVision Mixer module overcomes the unidirectional limitation of causal convolution, capturing bidirectional spatial-spectral dependencies in a single forward pass through decoupled attention that focuses on high-value features, alleviating parameter redundancy and the curse of dimensionality. On IN, UP, and KSC datasets, MVNet outperforms mainstream hyperspectral image classification methods in both classification accuracy and computational efficiency, demonstrating robust capability in processing complex HSI data.

cs.CV↗

arXiv2025

Hyperspectral Image Classification via Transformer-based Spectral-Spatial Attention Decoupling and Adaptive Gating

Deep neural networks face several challenges in hyperspectral image classification, including high-dimensional data, sparse distribution of ground objects, and spectral redundancy, which often lead to classification overfitting and limited generalization capability. To more effectively extract and fuse spatial context with fine spectral information in hyperspectral image (HSI) classification, this paper proposes a novel network architecture called STNet. The core advantage of STNet stems from the dual innovative design of its Spatial-Spectral Transformer module: first, the fundamental explicit decoupling of spatial and spectral attention ensures targeted capture of key information in HSI; second, two functionally distinct gating mechanisms perform intelligent regulation at both the fusion level of attention flows (adaptive attention fusion gating) and the internal level of feature transformation (GFFN). This characteristic demonstrates superior feature extraction and fusion capabilities compared to traditional convolutional neural networks, while reducing overfitting risks in small-sample and high-noise scenarios. STNet enhances model representation capability without increasing network depth or width. The proposed method demonstrates superior performance on IN, UP, and KSC datasets, outperforming mainstream hyperspectral image classification approaches.

cs.CV↗

arXiv2025

Dynamic 3D KAN Convolution with Adaptive Grid Optimization for Hyperspectral Image Classification

Deep neural networks face several challenges in hyperspectral image classification, including high-dimensional data, sparse distribution of ground objects, and spectral redundancy, which often lead to classification overfitting and limited generalization capability. To more efficiently adapt to ground object distributions while extracting image features without introducing excessive parameters and skipping redundant information, this paper proposes KANet based on an improved 3D-DenseNet model, consisting of 3D KAN Conv and an adaptive grid update mechanism. By introducing learnable univariate B-spline functions on network edges, specifically by flattening three-dimensional neighborhoods into vectors and applying B-spline-parameterized nonlinear activation functions to replace the fixed linear weights of traditional 3D convolutional kernels, we precisely capture complex spectral-spatial nonlinear relationships in hyperspectral data. Simultaneously, through a dynamic grid adjustment mechanism, we adaptively update the grid point positions of B-splines based on the statistical characteristics of input data, optimizing the resolution of spline functions to match the non-uniform distribution of spectral features, significantly improving the model's accuracy in high-dimensional data modeling and parameter efficiency, effectively alleviating the curse of dimensionality. This characteristic demonstrates superior neural scaling laws compared to traditional convolutional neural networks and reduces overfitting risks in small-sample and high-noise scenarios. KANet enhances model representation capability through a 3D dynamic expert convolution system without increasing network depth or width. The proposed method demonstrates superior performance on IN, UP, and KSC datasets, outperforming mainstream hyperspectral image classification approaches.

cs.CV↗

arXiv2025

Expert Kernel Generation Network Driven by Contextual Mapping for Hyperspectral Image Classification

Deep neural networks face several challenges in hyperspectral image classification, including high-dimensional data, sparse distribution of ground objects, and spectral redundancy, which often lead to classification overfitting and limited generalization capability. To more efficiently adapt to ground object distributions while extracting image features without introducing excessive parameters and skipping redundant information, this paper proposes EKGNet based on an improved 3D-DenseNet model, consisting of a context-aware mapping network and a dynamic kernel generation module. The context-aware mapping module translates global contextual information of hyperspectral inputs into instructions for combining base convolutional kernels, while the dynamic kernels are composed of K groups of base convolutions, analogous to K different types of experts specializing in fundamental patterns across various dimensions. The mapping module and dynamic kernel generation mechanism form a tightly coupled system - the former generates meaningful combination weights based on inputs, while the latter constructs an adaptive expert convolution system using these weights. This dynamic approach enables the model to focus more flexibly on key spatial structures when processing different regions, rather than relying on the fixed receptive field of a single static convolutional kernel. EKGNet enhances model representation capability through a 3D dynamic expert convolution system without increasing network depth or width. The proposed method demonstrates superior performance on IN, UP, and KSC datasets, outperforming mainstream hyperspectral image classification approaches.

cs.CV↗

arXiv2025

3D Wavelet Convolutions with Extended Receptive Fields for Hyperspectral Image Classification

Deep neural networks face numerous challenges in hyperspectral image classification, including high-dimensional data, sparse ground object distributions, and spectral redundancy, which often lead to classification overfitting and limited generalization capability. To better adapt to ground object distributions while expanding receptive fields without introducing excessive parameters and skipping redundant information, this paper proposes WCNet, an improved 3D-DenseNet model integrated with wavelet transforms. We introduce wavelet transforms to effectively extend convolutional receptive fields and guide CNNs to better respond to low frequencies through cascading, termed wavelet convolution. Each convolution focuses on different frequency bands of the input signal with gradually increasing effective ranges. This process enables greater emphasis on low-frequency components while adding only a small number of trainable parameters. This dynamic approach allows the model to flexibly focus on critical spatial structures when processing different regions, rather than relying on fixed receptive fields of single static kernels. The Wavelet Conv module enhances model representation capability by expanding receptive fields through 3D wavelet transforms without increasing network depth or width. Experimental results demonstrate superior performance on the IN, UP, and KSC datasets, outperforming mainstream hyperspectral image classification methods.

cs.CV↗