CLIP models learn transferable multi-modal features via image-text contrastive learning on internet-scale data. They are widely used in zero-shot classification, multi-modal retrieval, text-to-image diffusion, and as image encoders in large vision-language models. However, CLIP's pretraining is dominated by images paired with short captions, biasing the model toward encoding simple descriptions of salient objects and leading to coarse alignment on complex scenes and dense descriptions. While recent work mitigates this by fine-tuning on small-scale long-caption datasets, we identify an important common bias: both human- and LLM-generated long captions typically begin with a one-sentence summary followed by a detailed description. We show that this acts as a shortcut during training, concentrating attention on the opening sentence and early tokens and weakening alignment over the rest of the caption. To resolve this, we introduce DeBias-CLIP, which removes the summary sentence during training and applies sentence sub-sampling and text token padding to distribute supervision across all token positions. DeBias-CLIP achieves state-of-the-art long-text retrieval, improves short-text retrieval, and is less sensitive to sentence order permutations. It is a drop-in replacement for Long-CLIP with no additional trainable parameters.
论文检索
输入标题、作者或关键词，从 9,256 篇学术成果中精准定位
Recent subject-driven image customization excels in fidelity, yet fine-grained instance-level spatial control remains an elusive challenge, hindering real-world applications. This limitation stems from two factors: a scarcity of scalable, position-annotated datasets, and the entanglement of identity and layout by global attention mechanisms. To this end, we introduce PositionIC, a unified framework for high-fidelity, spatially controllable multi-subject customization. First, we present BMPDS, the first automatic data-synthesis pipeline for position-annotated multi-subject datasets, effectively providing crucial spatial supervision. Second, we design a lightweight, layout-aware diffusion framework that integrates a novel visibility-aware attention mechanism. This mechanism explicitly models spatial relationships via an NeRF-inspired volumetric weight regulation to effectively decouple instance-level spatial embeddings from semantic identity features, enabling precise, occlusion-aware placement of multiple subjects. Extensive experiments demonstrate PositionIC achieves state-of-the-art performance on public benchmarks, setting new records for spatial precision and identity consistency. Our work represents a significant step towards truly controllable, high-fidelity image customization in multi-entity scenarios.Code and data: https://github.com/MeiGen-AI/PositionIC.
BuildingGPT: Auto-Regressive Building Wireframe Reconstruction Model with Reinforcement Learning
PDF ↗In this paper, we propose BuildingGPT, a novel auto-regressive model for building wireframe reconstruction from point clouds with reinforcement learning.Unlike prior works based on detection or diffusion models, BuildingGPT reformulates the building wireframe reconstruction task into a sequence prediction problem.Based on the hierarchical building wireframe tokenization, the wireframe sequences are organized in a structurally- and semantically-aware order for the next-token prediction.The point cloud encoder first transforms the input point cloud into a fixed-length latent code that serves as the starting of the sequence.Then, BuildingGPT auto-regressively predicts tokens conditioned on the latent code and previously generated tokens.With token sequence predicted, the building wireframe is obtained through detokenization.To enhance the model performance, we adopt a two-stage training paradigm including the pre-training and post-training.After the auto-regressive pre-training, Direct Preference Optimization (DPO) is employed as a post-training strategy to align reconstruction results with human preferences.Extensive experiments on the large-scale MunichWF dataset show that BuildingGPT outperforms existing state-of-the-art methods.We commit to release the code and dataset.
Recent visual generative models often struggle with consistency during image editing due to the entangled nature of raster images, where all visual content is fused into a single canvas. In contrast, professional design tools employ layered representations, allowing isolated edits while preserving consistency. Motivated by this, we propose Qwen-Image-Layered, an end-to-end diffusion model that decomposes a single RGB image into multiple semantically disentangled RGBA layers, enabling inherent editability, where each RGBA layer can be independently manipulated without affecting other content. To support variable-length decomposition, we introduce three key components: (1) an RGBA-VAE to unify the latent representations of RGB and RGBA images; (2) a VLD-MMDiT (Variable Layers Decomposition MMDiT) architecture capable of decomposing a variable number of image layers; and (3) a Multi-stage Training strategy to adapt a pretrained image generation model into a multilayer image decomposer. Furthermore, to address the scarcity of high-quality multilayer training images, we build a pipeline to extract and annotate multilayer images from Photoshop documents (PSD). Experiments demonstrate that our method significantly surpasses existing approaches in decomposition quality and establishes a new paradigm for consistent image editing.
PropFly: Learning to Propagate via On-the-Fly Supervision from Pre-trained Video Diffusion Models
PDF ↗Propagation-based video editing enables precise user control by propagating a single edited frame into following frames while maintaining the original context such as motion and structures.However, training such models requires large-scale, paired (source and edited) video datasets, which are costly and complex to acquire.Hence, we propose the PropFly, a training pipeline for Propagation-based video editing, relying on on-the-Fly supervision from pre-trained video diffusion models (VDMs) instead of requiring off-the-shelf or precomputed paired video editing datasets.Specifically, our PropFly leverages one-step clean latent estimations from intermediate noised latents with varying Classifier-Free Guidance (CFG) scales to synthesize diverse pairs of 'source' (low-CFG) and 'edited' (high-CFG) latents on-the-fly. The source latent serves as structural information of the video, while the edited latent provides the target transformation for learning propagation.Our pipeline enables an additional adapter attached to the pre-trained VDM to learn to propagate edits via Guidance-Modulated Flow Matching (GMFM) loss, which guides the model to replicate the target transformation.Our on-the-fly supervision ensures the model to learn temporally consistent and dynamic transformations.Extensive experiments demonstrate that our PropFly significantly outperforms the state-of-the-art methods on various video editing tasks, producing high-quality editing results.
Multimodal autoregressive (AR) models, based on next-token prediction and transformer architecture, have demonstrated remarkable capabilities in various multimodal tasks including text-to-image (T2I) generation. Despite their strong performance in general T2I tasks, our research reveals that these models initially struggle with subject-driven image generation compared to dominant diffusion models. To address this limitation, we introduce Proxy-Tuning, leveraging diffusion models to enhance AR models' capabilities in subject-specific image generation. Our method reveals a striking weak-to-strong phenomenon: fine-tuned AR models consistently outperform their diffusion model supervisors in both subject fidelity and prompt adherence. We analyze this performance shift and identify scenarios where AR models excel, particularly in multi-subject compositions and contextual understanding. This work not only demonstrates impressive results in subject-driven AR image generation, but also unveils the potential of weak-to-strong generalization in the image generation domain, contributing to a deeper understanding of different architectures' strengths and limitations.
Diffusion models (DMs) can generate anatomically realistic medical images, offering a compelling route to improving generalization through synthetic augmentation. Yet high visual realism does not necessarily translate into improved downstream utility. This work addresses two key questions in diffusion-driven augmentation. First, what should be synthesized? We show that synthetic adversariality, namely the expected empirical loss induced by synthetic samples, is a key driver of generalization. More importantly, only native adversariality, arising from hard examples supported by the diffusion model distribution, yields consistent gains, whereas artificial adversariality induced by attack-based perturbations is detrimental. Second, how should such samples be synthesized? We propose the Adversariality Miner, a lightweight module that optimizes the initial noise to mine natively adversarial samples without modifying or retraining the diffusion model. Extensive experiments across diverse diffusion backbones and medical benchmarks confirm the effectiveness of our approach, establishing a principled path toward diffusion-driven generalization.
A fundamental challenge in embodied intelligence is developing expressive and compact state representations for efficient world modeling and decision making. However, existing methods often fail to achieve this balance, yielding representations that are either overly redundant or lacking in task-critical information. We propose an unsupervised approach that learns a highly compressed two-token state representation using a lightweight encoder and a pre-trained Diffusion Transformer (DiT) decoder, capitalizing on its strong generative prior. Our representation is efficient, interpretable, and integrates seamlessly into existing VLA-based models, improving performance by 11.6% on LIBERO and 31% in real-world task success with minimal inference overhead. More importantly, we find that the difference between these tokens, obtained via latent interpolation, naturally serves as a highly effective latent action, which can be further decoded into executable robot actions. This emergent capability reveals that our representation captures structured dynamics without explicit supervision. We name our method StaMo for its ability to learn generalizable robotic Motion from compact State representation, which is encoded from static images, challenging the prevalent dependence to learning latent action on complex architectures and video data. The resulting latent actions also enhance policy co-training, outperforming prior methods by 10.4% with improved interpretability. Moreover, our approach scales effectively across diverse data sources, including real-world robot data, simulation, and human egocentric video.
Low-Light Image Enhancement (LLIE) is a challenging task, as severe information loss means a single input can correspond to multiple plausible restorations. This inherent ambiguity causes conventional regression-based models to produce overly-smooth results that lack detail. While recent generative models can create richer details, their common unidirectional design often compromises content fidelity by distorting original structures. We introduce Bi-Bridge, a unified framework that models both enhancement and its inverse degradation within a single symmetric diffusion bridge. By compelling the network to preserve essential content structures across both transformations, this bidirectional learning acts as a powerful constraint, leading to significantly more faithful and realistic restorations. Extensive experiments show that Bi-Bridge outperforms state-of-the-art (SOTA) methods across multiple benchmarks, establishing a new standard for fidelity and perceptual quality.
Face relighting aims to synthesize realistic portraits under novel illumination while preserving identity and geometry. However, progress remains constrained by the limited availability of large-scale, physically consistent illumination data. To address this, we introduce POLAR, a large-scale and physically calibrated One-Light-at-a-Time (OLAT) dataset containing over 200 subjects captured under 156 lighting directions, multiple views, and diverse expressions. Building upon POLAR, we develop a flow-based generative model POLARNet that predicts per-light OLAT responses from a single portrait, capturing fine-grained and direction-aware illumination effects while preserving facial identity. Unlike diffusion or background-conditioned methods that rely on statistical or contextual cues, our formulation models illumination as a continuous, physically interpretable transformation between lighting states, enabling scalable and controllable relighting. Together, POLAR and POLARNet form a unified illumination learning framework that links real data, generative synthesis, and physically grounded relighting, establishing a self-sustaining "chicken-and-egg" cycle for scalable and reproducible portrait illumination.
Egocentric walking tour videos provide a rich source of image data to develop rich and diverse visual models of environments around the world. However, the significant presence of humans in frames of these videos due to crowds and eye-level camera perspectives mitigates their usefulness in environment modeling applications. We focus on addressing this challenge by developing a generative algorithm that can realistically remove (i.e., inpaint) humans and their associated shadow effects from walking tour videos. Key to our approach is the construction of a rich semi-synthetic dataset of video clip pairs to train this generative model. Each pair in the dataset consists of an environment-only background clip, and a composite clip of walking humans with simulated shadows overlaid on the background. We randomly sourced both foreground and background components from real egocentric walking tour videos around the world to maintain visual diversity. We then used this dataset to fine-tune the state-of-the-art Casper video diffusion model for object and effects inpainting, and demonstrate that the resulting model performs far better than Casper both qualitatively and quantitatively at removing humans from walking tour clips with significant human presence and complex backgrounds. Finally, we show that the resulting generated clips can be used to build successful 3D/4D models of urban locations.
Despite strong results on recognition and segmentation, current 3D visual pre-training methods often underperform on robotic manipulation. We attribute this gap to two factors: the lack of state-action-state dynamics modeling and the unnecessary redundancy of explicit geometric reconstruction. We introduce AFRO, a self-supervised framework that learns dynamics-aware 3D representations without action or reconstruction supervision. AFRO casts state prediction as a generative diffusion process and jointly models forward and inverse dynamics in a shared latent space to capture causal transition structure. To prevent feature leakage in action learning, we employ feature differencing and inverse-consistency supervision, improving the quality and stability of visual features. When combined with Diffusion Policy, AFRO substantially increases manipulation success rates across 16 simulated and 4 real-world tasks, outperforming existing pre-training approaches. The framework also scales favorably with data volume and task complexity. Qualitative visualizations indicate that AFRO learns semantically rich, discriminative features, offering an effective pre-training solution for 3D representation learning in robotics. Project page: https://kolakivy.github.io/AFRO/
Unified image generation and editing models suffer from severe task interference in dense diffusion transformers architectures, where a shared parameter space must compromise between conflicting objectives (e.g., local editing v.s. subject-driven generation). While the sparse Mixture-of-Experts (MoE) paradigm is a promising solution, its gating networks remain task-agnostic, operating based on local features, unaware of global task intent. This task-agnostic nature prevents meaningful specialization and fails to resolve the underlying task interference.In this paper, we propose a novel framework to inject semantic intent into MoE routing. We introduce a Hierarchical Task Semantic Annotation scheme to create structured task descriptors (e.g., scope, type, preservation). We then design Predictive Alignment Regularization to align internal routing decisions with the task's high-level semantics. This regularization evolves the gating network from a task-agnostic executor to a dispatch center. Our model effectively mitigates task interference, outperforming dense baselines in fidelity and quality, and our analysis shows that experts naturally develop clear and semantically correlated specializations.
Autoregressive video models are promising for world modeling via next-frame prediction, but they suffer from exposure bias: a mismatch between training on clean contexts and inference on self-generated frames, causing errors to compound and quality to drift over time. We introduce Backwards Aggregation (BAgger), a self-supervised scheme that constructs corrective trajectories from the model's own rollouts, teaching it to recover from its mistakes. Unlike prior approaches that rely on few-step distillation and distribution-matching losses --which can hurt quality and diversity -- BAgger trains with standard score or flow matching objectives, avoiding large teachers and long-chain backpropagation through time. We instantiate BAgger on causal diffusion transformers and evaluate on text-to-video, video extension, and multi-prompt generation, observing more stable long-horizon motion and better visual consistency with reduced drift.
We propose Tea-Adapter, a plug-and-play adapter designed to efficiently integrate conditional knowledge from a smaller teacher model into a larger student video diffusion model. Existing controllable video DiT methods face critical challenges: full fine-tuning of billion-parameter models is extremely expensive. At the same time, ControlNets introduce substantial parameter overhead and exhibit limited flexibility for novel multi-condition compositions. To overcome these issues, Tea-Adapter introduces a novel reverse distillation method that enables large video diffusion models to inherit precise control capabilities from smaller, efficiently tuned teacher diffusion models, eliminating the need for full fine-tuning. Moreover, recognizing the intrinsic relationships between different conditions, we replace the cascaded ControlNet design with a Mixture of Condition Experts (MCE) layer. This structure dynamically routes diverse conditional inputs within a unified architecture, supporting both single-condition control and multiple condition combinations without additional training cost. To achieve cross-scale knowledge transfer, we further develop a Feature Propagation Module to ensure efficient and temporally consistent feature propagation across video frames. Experiments demonstrate that Tea-Adapter enables high-fidelity, multi-condition video synthesis, making advanced, controllable video generation feasible on low-resource hardware.
Flow-GRPO successfully applies reinforcement learning to flow models, but uses uniform credit assignment across all steps. This ignores the temporal structure of diffusion generation: early steps determine composition and content (low-frequency structure), while late steps resolve details and textures (high-frequency details). Moreover, assigning uniform credit based solely on the final image can inadvertently reward suboptimal intermediate steps, especially when errors are corrected later in the diffusion trajectory. We propose Stepwise-Flow-GRPO, which assigns credit based on each step's reward improvement. By leveraging Tweedie's formula to obtain intermediate reward estimates and introducing gain-based advantages, our method achieves superior sample efficiency and faster convergence. We also introduce a DDIM-inspired SDE that improves reward quality while preserving stochasticity for policy gradients.
Forecast the Principal, Stabilize the Residual: Subspace-Aware Feature Caching for Diffusion Transformers
PDF ↗Diffusion Transformer (DiT) models have achieved unprecedented quality in image and video generation, yet their iterative sampling process remains computationally prohibitive. To accelerate inference, feature caching methods have emerged by reusing or forecasting intermediate representations across timesteps. However, existing caching approaches treat all feature components uniformly. We reveal that DiT feature spaces contain distinct principal and residual subspaces with divergent temporal behavior: the principal subspace evolves smoothly and predictably, while the residual subspace exhibits volatile, low-energy oscillations that resist accurate prediction. Building on this insight, we propose SVD-Cache, a subspace-aware caching framework that decomposes diffusion features via Singular Value Decomposition (SVD), applies exponential moving average (EMA) prediction to the dominant low-rank components, and directly reuses the residual subspace. Extensive experiments demonstrate that SVD-Cache achieves near-lossless across diverse models and methods, including 5.55xspeedup on FLUX and HunyuanVideo, and compatibility with model acceleration techniques including distillation, quantization and sparse attention. Our code is available at \href https://github.com/BlackMaple1203/SVDCache https://github.com/BlackMaple1203/SVDCache .
Copy-Transform-Paste: Zero-Shot Object-Object Alignment Guided by Vision-Language and Geometric Constraints
PDF ↗We study zero-shot 3D alignment of two given meshes from a short text prompt describing their spatial relation---an essential capability for content creation and scene assembly. Earlier approaches primarily rely on geometric alignment procedures, while recent work leverages pretrained 2D diffusion models to model language-conditioned object-object spatial relationships. In contrast, we directly optimize the relative pose at test time---updating translation, rotation, and isotropic scale with CLIP-driven gradients via a differentiable renderer---without training a new model.Our framework augments language supervision with geometry-aware objectives: a variant of soft-Iterative Closest Point (ICP) term to encourage controlled surface attachment and a penetration loss to discourage interpenetration. A phased schedule strengthens contact constraints over time, camera control concentrates views on the interaction region, and randomized restarts improve robustness.To enable evaluation, we curate a benchmark of 50 \ mesh pair, prompt\ cases spanning diverse categories and relations, and compare against baselines. Across the benchmark, our method yields semantically faithful and physically plausible alignments, improving CLIP similarity while reducing intersection volume.
cryoSENSE: Compressive Sensing Enables High-throughput Microscopy with Sparse and Generative Priors on the Protein Cryo-EM Image Manifold
PDF ↗Cryo-electron microscopy (cryo-EM) enables the atomic-resolution visualization of biomolecules; however, modern direct detectors generate data volumes that far exceed the available storage and transfer bandwidth, thereby constraining practical throughput. We introduce cryoSENSE, the computational realization of a hardware-software co-designed framework for compressive cryo-EM sensing and acquisition. We show that cryo-EM images of proteins lie on low-dimensional manifolds that can be independently represented using sparse priors in predefined bases and generative priors captured by a denoising diffusion model. cryoSENSE leverages these low-dimensional manifolds to enable faithful image reconstruction from spatial and Fourier-domain undersampled measurements while preserving downstream structural resolution. In experiments, cryoSENSE increases acquisition throughput by up to 2.5xwhile retaining the original 3D resolution, offering controllable trade-offs between the number of masked measurements and the level of downsampling. Sparse priors favor faithful reconstruction from Fourier-domain measurements and moderate compression, whereas generative diffusion priors achieve accurate recovery from pixel-domain measurements and more severe undersampling.
Recent approaches have demonstrated the promise of using diffusion models to generate interactive and explorable worlds. However, most of these methods face critical challenges such as excessively large parameter sizes, reliance on lengthy inference steps, and rapidly growing historical context, which severely limit real-time performance and lack text-controlled generation capabilities.To address these challenges, we propose Yume1.5, a novel framework designed to generate realistic, interactive, and continuous worlds from a single image or text prompt. Yume1.5 achieves this through a carefully designed framework that supports keyboard-based exploration of the generated worlds. The framework comprises three core components: (1) a long-video generation method combining unified context compression and linear attention; (2) a context compression-based bidirectional attention distillation approach with an enhanced text embedding scheme for real-time streaming video generation. Yume1.5 achieves an average generation speed of 12 fps at 540p resolution using only a single A100 GPU; (3) a text-controlled method for generating world events. Github: https://github.com/stdstu12/YUME.