Selected work

Representative Papers

FlowDIS: Language-Guided Dichotomous Image Segmentation with Flow Matching

May 06, 2026

Accurate image segmentation is essential for modern computer vision applications such as image editing, autonomous driving, and medical image analysis. In recent years, Dichotomous Image Segmentation (DIS) has become a standard task for training and evaluating highly accurate segmentation models. Existing DIS approaches often fail to preserve fine-grained details or fully capture the semantic structure of the foreground. To address these challenges, we present FlowDIS, a novel dichotomous image segmentation method built on the flow matching framework, which learns a time-dependent vector field to transport the image distribution to the corresponding mask distribution, optionally conditioned on a text prompt. Moreover, with our Position-Aware Instance Pairing (PAIP) training strategy, FlowDIS offers strong controllability through text prompts, enabling precise, pixel-level object segmentation. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art approaches both with and without language guidance. Compared with the best prior DIS method, FlowDIS achieves a 5.5% higher $F_β^ω$ measure and 43% lower MAE ($\mathcal{M}$) on the DIS-TE test set. The code is available at: https://github.com/Picsart-AI-Research/FlowDIS

0 citationsRead paper

Beyond Realism: Learning the Art of Expressive Composition with StickerNet

Nov 25, 2025

Traditional image synthesis prioritizes photorealism and semantic plausibility, yet users in creative platforms emphasize artistic expression, playfulness, and social appeal—motivating a shift toward expressive synthesis, which values stylistic diversity and flexible layout logic to model real-world sticker editing behavior. Method: We introduce StickerNet, a two-stage framework (classification followed by regression) that jointly predicts sticker transparency, position, scale, and mask, trained on the first large-scale dataset derived from real user editing logs. Crucially, our approach abandons pixel-level realism in favor of modeling human creative intent. Contribution/Results: Extensive user studies and quantitative evaluations demonstrate that StickerNet significantly outperforms existing baselines in style compatibility, layout合理性, and social acceptability—closely aligning with authentic creative practices. This validates the effectiveness of learning expressive priors directly from behavioral data.

0 citationsRead paper

Distilling Normalizing Flows

Jun 26, 2025

This work addresses the performance bottlenecks of compact normalizing flow (NF) models in density estimation and sample quality. We propose a novel knowledge distillation framework specifically designed for NF architectures, moving beyond conventional output-layer distillation to enable asymmetric, structure-aware knowledge transfer at intermediate latent layers—particularly suited to the modular design of compositional NFs. By explicitly modeling probabilistic flow mappings between corresponding teacher and student layers, our method significantly improves parameter efficiency and inference speed of student models. Experiments demonstrate that distilled compact NFs achieve 23–37% lower density estimation error, 18–41% improvement in sampling Fréchet Inception Distance (FID), 2.1× higher throughput, and 58% reduction in computational overhead on standard benchmarks. The approach establishes a scalable paradigm for lightweight generative modeling.

0 citationsRead paper

RAGME: Retrieval Augmented Video Generation for Enhanced Motion Realism

Apr 09, 2025

Existing text-to-video generation methods exhibit notable deficiencies in modeling motion complexity and physical plausibility, frequently yielding static artifacts, jitter, or physically implausible motion. To address this, we propose RAG-Video, the first retrieval-augmented diffusion video generation framework that integrates Retrieval-Augmented Generation (RAG) into the video synthesis pipeline. During diffusion sampling, RAG-Video dynamically retrieves semantically and kinematically similar reference videos and extracts multi-granularity dynamic priors as conditional guidance—enabling enhanced motion modeling without large-scale retraining. The method is compatible with mainstream text-to-video diffusion architectures and synergistically combines cross-modal retrieval, conditional fine-tuning, and motion-aligned feature integration. Evaluated on novel benchmarks—including MotionScore and PhysEval—RAG-Video significantly outperforms state-of-the-art approaches, markedly reducing motion distortion while improving temporal coherence and physical realism of generated videos.

0 citationsRead paper

Safe Vision-Language Models via Unsafe Weights Manipulation

Mar 14, 2025

Vision-language models (VLMs) inherit dataset biases, and existing training-based safety alignment methods often degrade performance on safe inputs. Method: We propose a training-free Unsafe Weights Manipulation (UWM) paradigm. Leveraging fine-grained evaluation from SafeGround, UWM identifies critical parameters by contrasting intra-layer activations between safe and unsafe samples, then applies targeted weight sign flipping. Contribution/Results: UWM is the first method to uncover and mitigate the counterintuitive performance drop on safe inputs induced by safety alignment. Without altering model architecture or updating parameters, it simultaneously enhances safety and preserves knowledge capabilities. Experiments show that UWM significantly improves safety on unsafe queries while outperforming all state-of-the-art training-based methods on safe queries—achieving near-lossless retention of knowledge capacity.

0 citationsRead paper

Recent publications

Latest Papers

FlowDIS: Language-Guided Dichotomous Image Segmentation with Flow Matching

May 06, 2026

Accurate image segmentation is essential for modern computer vision applications such as image editing, autonomous driving, and medical image analysis. In recent years, Dichotomous Image Segmentation (DIS) has become a standard task for training and evaluating highly accurate segmentation models. Existing DIS approaches often fail to preserve fine-grained details or fully capture the semantic structure of the foreground. To address these challenges, we present FlowDIS, a novel dichotomous image segmentation method built on the flow matching framework, which learns a time-dependent vector field to transport the image distribution to the corresponding mask distribution, optionally conditioned on a text prompt. Moreover, with our Position-Aware Instance Pairing (PAIP) training strategy, FlowDIS offers strong controllability through text prompts, enabling precise, pixel-level object segmentation. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art approaches both with and without language guidance. Compared with the best prior DIS method, FlowDIS achieves a 5.5% higher $F_β^ω$ measure and 43% lower MAE ($\mathcal{M}$) on the DIS-TE test set. The code is available at: https://github.com/Picsart-AI-Research/FlowDIS

0 citationsRead paper

Beyond Realism: Learning the Art of Expressive Composition with StickerNet

Nov 25, 2025

Traditional image synthesis prioritizes photorealism and semantic plausibility, yet users in creative platforms emphasize artistic expression, playfulness, and social appeal—motivating a shift toward expressive synthesis, which values stylistic diversity and flexible layout logic to model real-world sticker editing behavior. Method: We introduce StickerNet, a two-stage framework (classification followed by regression) that jointly predicts sticker transparency, position, scale, and mask, trained on the first large-scale dataset derived from real user editing logs. Crucially, our approach abandons pixel-level realism in favor of modeling human creative intent. Contribution/Results: Extensive user studies and quantitative evaluations demonstrate that StickerNet significantly outperforms existing baselines in style compatibility, layout合理性, and social acceptability—closely aligning with authentic creative practices. This validates the effectiveness of learning expressive priors directly from behavioral data.

0 citationsRead paper

Distilling Normalizing Flows

Jun 26, 2025

This work addresses the performance bottlenecks of compact normalizing flow (NF) models in density estimation and sample quality. We propose a novel knowledge distillation framework specifically designed for NF architectures, moving beyond conventional output-layer distillation to enable asymmetric, structure-aware knowledge transfer at intermediate latent layers—particularly suited to the modular design of compositional NFs. By explicitly modeling probabilistic flow mappings between corresponding teacher and student layers, our method significantly improves parameter efficiency and inference speed of student models. Experiments demonstrate that distilled compact NFs achieve 23–37% lower density estimation error, 18–41% improvement in sampling Fréchet Inception Distance (FID), 2.1× higher throughput, and 58% reduction in computational overhead on standard benchmarks. The approach establishes a scalable paradigm for lightweight generative modeling.

0 citationsRead paper

RAGME: Retrieval Augmented Video Generation for Enhanced Motion Realism

Apr 09, 2025

Existing text-to-video generation methods exhibit notable deficiencies in modeling motion complexity and physical plausibility, frequently yielding static artifacts, jitter, or physically implausible motion. To address this, we propose RAG-Video, the first retrieval-augmented diffusion video generation framework that integrates Retrieval-Augmented Generation (RAG) into the video synthesis pipeline. During diffusion sampling, RAG-Video dynamically retrieves semantically and kinematically similar reference videos and extracts multi-granularity dynamic priors as conditional guidance—enabling enhanced motion modeling without large-scale retraining. The method is compatible with mainstream text-to-video diffusion architectures and synergistically combines cross-modal retrieval, conditional fine-tuning, and motion-aligned feature integration. Evaluated on novel benchmarks—including MotionScore and PhysEval—RAG-Video significantly outperforms state-of-the-art approaches, markedly reducing motion distortion while improving temporal coherence and physical realism of generated videos.

0 citationsRead paper

Safe Vision-Language Models via Unsafe Weights Manipulation

Mar 14, 2025

Vision-language models (VLMs) inherit dataset biases, and existing training-based safety alignment methods often degrade performance on safe inputs. Method: We propose a training-free Unsafe Weights Manipulation (UWM) paradigm. Leveraging fine-grained evaluation from SafeGround, UWM identifies critical parameters by contrasting intra-layer activations between safe and unsafe samples, then applies targeted weight sign flipping. Contribution/Results: UWM is the first method to uncover and mitigate the counterintuitive performance drop on safe inputs induced by safety alignment. Without altering model architecture or updating parameters, it simultaneously enhances safety and preserves knowledge capabilities. Experiments show that UWM significantly improves safety on unsafe queries while outperforming all state-of-the-art training-based methods on safe queries—achieving near-lossless retention of knowledge capacity.

0 citationsRead paper