论文检索

Zhan Wang, Leiquan Wang, Chunlei Wu, Yu Meng

Image denoising is a fundamental task in computer vision aimed at recovering clean images from noise-corrupted observations. While supervised deep learning methods achieve remarkable performance when trained on paired data with known noise levels, their real-world applicability is limited as noise characteristics are often unknown. Existing unsupervised techniques, such as blind-spot networks or methods based on statistical estimation, either compromise performance due to information loss or suffer from inaccuracies in noise level estimation. To address these challenges, we propose a novel two-stage self-supervised denoising framework that first accurately estimates the noise level directly from noisy images, without requiring clean references or prior noise knowledge. Building upon theoretical insights from Noisier2Noise, we rigorously derive a relationship between the noise level and the variance of the denoised image, enabling robust estimation via a deep learning model and a ternary search strategy. The estimated noise level is then used to synthesize training pairs for supervised denoising. Experiments demonstrate that our method outperforms existing unsupervised approaches and traditional noise estimation techniques, achieving performance competitive with--and in some cases surpassing--supervised methods trained with known noise levels. The proposed framework effectively overcomes the training data pair limitations of supervised approaches for unknown additive white Gaussian noise. Our code is available at https://github.com/zhanzhanblue/CANC.

Danxu Liu, Di Wang, Hebaixu Wang, Haoyang Chen, Wentao Jiang, Yilin Cheng, Haonan Guo, Wei Cui, Jing Zhang

Synthetic Aperture Radar (SAR) imagery plays a critical role in all-weather, day-and-night remote sensing applications. However, existing SAR-oriented deep learning is constrained by data scarcity, while the physically grounded speckle noise in SAR imagery further hampers fine-grained semantic representation learning. To address these challenges, we propose SARMAE, a Noise-Aware Masked Autoencoder for self-supervised SAR representation learning. Specifically, we construct SAR-1M, the first million-scale SAR dataset, with additional paired optical images, to enable large-scale pre-training. Building upon this, we design Speckle-Aware Representation Enhancement (SARE), which injects SAR-specific speckle noise into Masked Autoencoder to facilitate noise-aware and robust representation learning. Furthermore, we introduce Semantic Anchor Representation Constraint (SARC), which leverages paired optical priors to align SAR features and ensure semantic consistency. Extensive experiments across multiple SAR datasets demonstrate that SARMAE achieves state-of-the-art performance on classification, detection, and segmentation tasks. Code, data, and models are available at https://github.com/MiliLab/SARMAE.

Matthew Strong, Wei-Jer Chang, Quentin Herau, Jiezhi Yang, Yihan Hu, Chensheng Peng, Wei Zhan

Ego-centric driving videos available online provide an abundant source of visual data for autonomous driving, yet their lack of annotations makes it difficult to learn representations that capture both semantic structure and 3D geometry. Recent advances in large feedforward spatial models demonstrate that point maps and ego-motion can be inferred in a single forward pass, suggesting a promising direction for scalable driving perception. We therefore propose a label-free, teacher-guided framework for learning autonomous driving representations directly from unposed videos. Unlike prior self-supervised approaches that focus primarily on frame-to-frame consistency, we posit that safe and reactive driving depends critically on temporal context. To this end, we leverage a feedforward architecture equipped with a lightweight autoregressive module, trained using multi-modal supervisory signals that guide the model to jointly predict current and future point maps, camera poses, semantic layouts, and motion masks. Multi-modal teachers provide sequence-level pseudo-supervision, enabling LFG to learn a unified pseudo-4D representation from raw YouTube videos without poses, labels, or LiDAR. The resulting encoder not only transfers effectively to downstream autonomous driving planning on the NAVSIM benchmark, surpassing multi-camera and LiDAR baselines with only a single monocular camera, but also yields strong performance when evaluated on a range of semantic, geometric, and motion prediction tasks. These geometry- and motion-aware features position LFG as a compelling video-centric foundation model for autonomous driving.

Qitao Zhao, Hao Tan, Qianqian Wang, Sai Bi, Kai Zhang, Kalyan Sunkavalli, Shubham Tulsiani, Hanwen Jiang

Self-supervised pre-training has driven rapid progress in foundation models for language, 2D images, and video, yet remains largely unexplored for learning 3D-aware representations from multi-view images. In this paper, we present E-RayZer, a self-supervised 3D vision model that learns geometrically grounded representations directly from unlabeled images. Unlike prior self-supervised methods such as RayZer, which infer 3D indirectly through latent-space view synthesis, E-RayZer operates directly in 3D space, performing self-supervised 3D reconstruction with Explicit geometry. This formulation eliminates shortcut solutions and yields representations that are 3D-aware. To ensure convergence and scalability, we introduce a fine-grained learning curriculum that organizes training from easy to hard samples and harmonizes heterogeneous data sources without any supervision. Experiments show that E-RayZer significantly outperforms RayZer on pose estimation and matches or sometimes surpasses fully supervised reconstruction models such as VGGT. Furthermore, its learned representations outperform leading visual pre-training models (e.g., DINOv3, CroCo v2, VideoMAE V2, and RayZer) on 3D downstream tasks, establishing E-RayZer as a promising paradigm for spatial visual pre-training.

Gustavo Chau Loo Kung, Mohammad Abbasi, Camila Blank, Juze Zhang, Alan Q. Wang, Sophie Ostmeier, Akshay Chaudhari, Kilian Pohl, Ehsan Adeli

Diffusion Magnetic Resonance Imaging (dMRI) plays a critical role in studying microstructural changes in the brain. It is, therefore, widely used in clinical practice; yet progress in learning general-purpose representations from dMRI has been limited. A key challenge is that existing deep learning approaches are not well-suited to capture the unique properties of diffusion signals. Brain dMRI is normally composed of several brain volumes, each with different attenuation characteristics dependent on the direction and strength of the diffusion-sensitized gradients. Thus, there is a need to jointly model spatial, diffusion-weighting, and directional dependencies in dMRI. Furthermore, varying acquisition protocols (e.g., differing numbers of directions) further limit traditional models. To address these gaps, we introduce a diffusion space rotatory positional embedding (D-RoPE) plugged into our dMRI transformer to capture both the spatial structure and directional characteristics of diffusion data, enabling robust and transferable representations across diverse acquisition settings and an arbitrary number of diffusion directions. After self-supervised masked autoencoding pretraining, tests on several downstream tasks show that the learned representations and the pretrained model can provide competitive or superior performance compared to several baselines in these downstream tasks (even compared to a fully trained baseline); the finetuned features from our pretrained encoder resulted in a 6% higher accuracy in classifying mild cognitive impairment and a 0.05 increase in the correlation coefficient when predicting cognitive scores. Code is available at: github.com/gustavochau/D-RoPE

Yingyan Xu, Pramod Rao, Sebastian Weiss, Gaspard Zoss, Markus Gross, Christian Theobalt, Marc Habermann, Derek Bradley

3D Gaussian Splatting (3DGS) has become a standard approach to reconstruct and render photorealistic 3D head avatars. A major challenge is to relight the avatars to match any scene illumination. For high quality relighting, existing methods require subjects to be captured under complex time-multiplexed illumination, such as one-light-at-a-time (OLAT). We propose a new generalized relightable 3D Gaussian head model that can relight any subject observed in a single- or multi-view images without requiring OLAT data for that subject. Our core idea is to learn a mapping from flat-lit 3DGS avatars to corresponding relightable Gaussian parameters for that avatar. Our model consists of two stages: a first stage that models flat-lit 3DGS avatars without OLAT lighting, and a second stage that learns the mapping to physically-based reflectance parameters for high-quality relighting. This two-stage design allows us to train the first stage across diverse existing multi-view datasets without OLAT lighting ensuring cross-subject generalization, where we learn a dataset-specific lighting code for self-supervised lighting alignment. Subsequently, the second stage can be trained on a significantly smaller dataset of subjects captured under OLAT illumination. Together, this allows our method to generalize well and relight any subject from the first stage as if we had captured them under OLAT lighting. Furthermore, we can fit our model to unseen subjects from as little as a single image, allowing several applications in novel view synthesis and relighting for digital avatars.

Junyuan Zhang, Bin Wang, Qintong Zhang, Fan Wu, Zichen Wen, Jialin Lu, Junjie Shan, Ziqi Zhao, Shuya Yang, Ziling Wang 等

Table recognition (TR) aims to transform table images into semi-structured representations such as HTML or Markdown.As a core component of document parsing, TR has long relied on supervised learning, with recent efforts dominated by fine-tuning vision-language models (VLMs) using labeled data.While VLMs have brought TR to the next level, pushing performance further demands large-scale labeled data that is costly to obtain.Consequently, although proprietary models have continuously pushed the performance boundary, open-source models, often trained with limited resources and, in practice, the only viable option for many due to privacy regulations, still lag far behind.To bridge this gap, we introduce TRivia, a self-supervised fine-tuning method that enables pretrained VLMs to learn TR directly from unlabeled table images in the wild. Built upon Group Relative Policy Optimization, TRivia automatically identifies unlabeled samples that most effectively facilitate learning and eliminates the need for human annotations through a question-answering-based reward mechanism. An attention-guided module generates diverse questions for each table image, and the ability to interpret the recognition results and answer them correctly provides feedback to optimize the TR model.This closed-loop process allows the TR model to autonomously learn to recognize, structure, and reason over tables without labeled data. Leveraging this pipeline, we present TRivia-3B, an open-sourced, compact, and state-of-the-art TR model that surpasses existing systems (e.g., Gemini 2.5 Pro, MinerU2.5) on three popular benchmarks.

Guillaume Letellier, Siddharth Srivastava, Frederic Jurie, Gaurav Sharma

Foundation models pre-trained with self-supervised learning (SSL) on large-scale datasets have become powerful general-purpose feature extractors. However, their immense size and computational cost make them prohibitive for deployment on edge devices such as robots and AR/VR headsets. Existing compression techniques like standard knowledge distillation create efficient `specialist' models but sacrifice the crucial, downstream-agnostic generality that makes foundation models so valuable.In this paper, we introduce Foundation Model Distillation (FMD), a new paradigm for compressing large SSL models into compact, efficient, and faithful proxies that retain their general-purpose representational power. We present Foundry, the first implementation of FMD for 3D point clouds. Our approach, Foundry, trains a student to learn a compressed set of SuperTokens that reconstruct the teacher's token-level representations, capturing a compact basis of its latent space. A single distilled model maintains strong transferability across diverse downstream tasks--classification, part segmentation, and few-shot scenarios--approaching full foundation-model performance while using significantly fewer tokens and FLOPs, making such models more practical for deployment on resource-constrained hardware.

Chengan Che, Chao Wang, Xinyue Chen, Sophia Tsoka, Luis C. Garcia-Peraza-Herrera

Procedural activities, ranging from routine cooking to complex surgical operations, are highly structured sequences of actions performed in a specific temporal order. Despite the success of current self-supervised learning (SSL) methods on static images and short clips, these models often overlook the underlying sequential structure of such activities. We expose this lack of procedural awareness with a motivating experiment: models pretrained on forward and time-reversed sequences produce highly similar features, confirming that their representations are blind to the underlying procedural order. To address this shortcoming, we propose PL-Stitch, a self-supervised framework that harnesses the inherent temporal order of video frames as a powerful supervisory signal. Our approach integrates two novel probabilistic objectives based on the Plackett-Luce (PL) model. The primary PL objective trains the model to sort sampled frames chronologically, compelling it to learn the global workflow progression. The secondary objective, a spatio-temporal jigsaw loss, complements the learning by capturing fine-grained, cross-frame object correspondences. Our approach consistently achieves superior performance across five surgical and cooking benchmarks. Specifically, PL-Stitch yields significant gains in surgical phase recognition (e.g., +11.4 pp in k-NN accuracy on Cholec80) and cooking action segmentation (e.g., +5.7 pp in linear probing accuracy on Breakfast), demonstrating its effectiveness for procedural video representation learning. Code and models are available at https://github.com/visurg-ai/PL-Stitch.

David Nordström, Johan Edstedt, Fredrik Kahl, Georg Bökman

Self-supervised learning on images seeks to extract meaningful visual representations from unlabeled data. When scaled to large datasets, this paradigm has achieved state-of-the-art performance and the resulting trained models such as DINOv3 have seen widespread adoption. However, most prior efforts are optimized for semantic understanding rather than geometric reasoning. One important exception is Cross-View Completion, CroCo, which is a form of masked autoencoding (MAE) tailored for 3D understanding. In this work, we continue on the path proposed by CroCo and focus on learning features tailored for 3D vision. In a nutshell, we extend MAE to arbitrarily many views of the same scene. By uniformly masking all views and employing a lightweight decoder with inter-frame attention, our approach is inherently simpler and more scalable than CroCo. We evaluate the resulting model, MuM, extensively on downstream tasks including feedforward reconstruction, dense image matching and relative pose estimation, finding that it outperforms the state-of-the-art visual encoders DINOv3 and CroCo v2.

Jiadong Pan, Liang Li, Yuxin Peng, Yu-Ming Tang, Shuohuan Wang, Yu Sun, Hua Wu, Qingming Huang, Haifeng Wang

Recently, unified multimodal models (UMMs) have made remarkable progress in integrating visual understanding and generation, demonstrating strong potential for complex text-to-image (T2I) tasks. Despite their theoretical promise, a persistent capability gap exists: UMMs typically exhibit superior visual understanding but comparatively weaker generative capabilities. This discrepancy arises largely from the intrinsic decoupling between the understanding and generation processes. While a UMM can accurately interpret fine-grained visual details, it often struggles to produce semantically coherent images from complex textual prompts. To address this challenge, we explore UMMs' internal understanding capability to enhance generation quality. We propose a token-level intrinsic text-image alignment reward mechanism, GvU, enabling the UMM to act simultaneously as teacher and student: it evaluates its own outputs using the understanding branch to guide the generations accordingly. Building upon this, we design a self-supervised reinforcement learning framework, allowing UMMs to iteratively improve their generation quality through understanding-based intrinsic reward signals--without reliance on external supervision. Experimental results show that our method substantially boosts UMMs' generation, which in turn strengthens their fine-grained visual understanding, narrowing the capability gap between UMMs' visual understanding and generation. The project page is https://matrix0721.github.io/gvu.github.io/.

Kangjian Zhu, Haobo Jiang, Jianjun Qian, Jin Xie

In this paper, we propose a cross-view fusion framework that enhances the robustness of 6-DoF grasp pose estimation in corner views.Our framework alleviates occlusion by incorporating an auxiliary view and avoids the time-consuming, task-agnostic multi-view reconstruction through a post-fusion strategy.To enable cross-view fusion, we propose a self-supervised contrastive learning strategy that leverages cross-view associations to regularize point cloud features.In brief, a cross-view point pair is considered a match if the two points correspond to the same 3D location, and a non-match if they represent distinct grasp directions.The learning strategy significantly enhances the spatial consistency and direction distinctiveness of point features, thereby facilitating cross-view fusion and improving estimation robustness.Furthermore, we propose a cross-view-aligned cylinder integration module to fuse grasp-relevant geometry into a comprehensive representation.Specifically, the module first aligns the cross-view points and features according to their similarity to enhance the robustness against noise.Subsequently, these points are registered into the cylindrical coordinate frame, emphasizing the rotation-symmetric geometry which is important for grasping.Finally, local self-attention and seed cross-attention layers are alternately employed, respectively enabling interactions within single views and across views, which supports fine-grained representation of grasp-relevant geometry.Our framework achieves strong performance on the GraspNet-1Billion benchmark and in real-world applications. Code is available at GitHub.

Liang Zeng, Valerio Marsocci, Wufan Zhao, Andrea Nascetti, Maarten Vergauwen

Masked Image Modeling has been one of the most popular self-supervised learning paradigms to learn representations from large-scale, unlabeled Earth Observation images. While incorporating multi-modal and multi-temporal Earth Observation data into Masked Image Modeling has been widely explored, the spatial dependencies between images captured from neighboring areas remains largely overlooked. Since the Earth's surface is continuous, neighboring images are highly related and offer rich contextual information for self-supervised learning. To close this gap, we propose NeighborMAE, which learns spatial dependencies by joint reconstruction of neighboring Earth Observation images. To ensure that the reconstruction remains challenging, we leverage a heuristic strategy to dynamically adjust the mask ratio and the pixel-level loss weight. Experimental results across various pretraining datasets and downstream tasks show that NeighborMAE significantly outperforms existing baselines, underscoring the value of neighboring images in Masked Image Modeling for Earth Observation and the efficacy of our designs. Our code is available on https://github.com/LeungTsang/NeighborMAE.

Yu Luo, Xiaogang Zhu, Shan Zeng, Wei Xiang, Thomas Francis Bishop, Zhiyong Wang, Kun Hu

Accurate crop yield prediction is crucial for sustainable agriculture and global food security. While existing methods are predominantly developed for single-crop prediction, they often struggle to generalize across diverse crop types, without addressing the unique crop phenological responses that are dynamically modulated by complex weather patterns. In this paper, we propose PhenoYieldNet, a multi-crop yield prediction framework that learns crop-specific phenology by explicitly modeling their responses with temporal drivers. Specifically, we develop a crop-aware temporal decoder consisting of a Crop Phenology Bank (CPB) and a Crop Phenology Attention (CPA) module. The CPB integrates a set of learnable embeddings, which leverage a query to guide the CPA module to learn the most relevant phenology patterns for the specific crop. And the CPA module explicitly captures multi-scale trend and variation components to construct temporal contexts, enabling the model to dynamically adjust the attention across different phenological stages. To learn robust and generalizable features for multi-crop prediction, the encoder is initialized with a pre-trained foundation model, and further adapted via a self-supervised Temporal Contrastive Adaptation strategy to align with agricultural temporal dynamics. Extensive experiments conducted on multi-crop datasets indicate that our proposed method significantly outperforms state-of-the-art methods, exhibiting strong generalization capabilities across different regions and crops.

Yuhong Liu, Beichen Zhang, Yuhang Zang, Yuhang Cao, Long Xing, Xiaoyi Dong, Haodong Duan, Dahua Lin, Jiaqi Wang

Spatial understanding remains a weakness of Large Vision-Language Models (LVLMs). Existing supervised fine-tuning (SFT) and recent reinforcement learning with verifiable rewards (RLVR) pipelines depend on costly supervision, specialized tools, or constrained environments that limit scale. We introduce Spatial-SSRL, a self-supervised RL paradigm that derives verifiable signals directly from ordinary RGB or RGB-D images. Spatial-SSRL automatically formulates five pretext tasks that capture 2D and 3D spatial structure: shuffled patch reordering, flipped patch recognition, cropped patch inpainting, regional depth ordering, and relative 3D position prediction. These tasks provide ground-truth answers that are easy to verify and require no human or LVLM annotation. Training on our tasks substantially improves spatial reasoning while preserving general visual capabilities. On seven spatial understanding benchmarks in both image and video settings, Spatial-SSRL delivers average accuracy gains of 4.63% (3B) and 3.89% (7B) over the Qwen2.5-VL baselines. Our results show that simple, intrinsic supervision enables RLVR at scale and provides a practical route to stronger spatial intelligence in LVLMs.

Shan Zhang, Aotian Chen, Kai Zou, Jindong Gu, Yuan Xue, Anton van den Hengel

Symbolic computer vision represents diagrams through explicit logical rules and structured representations, enabling interpretable understanding in machine vision. This requires fundamentally different learning paradigms from pixel-based visual models. Symbolic visual learners parse diagrams into geometric primitives--points, lines, and shapes--whereas pixel-based learners operate on textures and colors. We propose a novel self-supervised symbolic auto-encoder, where the encoder parses diagrams into structured primitives and their relations, and the decoder compiles them back into executable code that reproduces the original diagrams. Central to this architecture is SymHPR (Symbolic Hierarchical Process Reward Modeling), which applies hierarchical step-level rewards to enforce point-on-line, line-on-shape, and shape-on-relation consistency. Since reinforcement learning collapses under vanilla GRPO, our proposed stabilized strategies balance exploration and exploitation. Built upon the symbolic learner, we design two pipelines: one replaces the MLLM's visual encoder with the symbolic encoder; the other feeds interpretable symbolic parses directly to the LLM to guide reasoning. Evaluations across reconstruction, perception, and reasoning tasks demonstrate the effectiveness of our approach: achieving a 98.2% reduction in MSE for geometric diagram reconstruction, surpassing GPT-4o by 0.6% with a 7B model on chart reconstruction, and improving by +13% on the MathGlance perception benchmark, and by +3% on MathVerse and GeoQA reasoning benchmarks.

Bo Zhou, Qiuxia Lai, Zeren Sun, Xiangbo Shu, Yazhou Yao, Wenguan Wang

Robust 3D representation learning forms the perceptual foundation of spatial intelligence, enabling downstream tasks in scene understanding and embodied AI. However, learning such representations directly from unposed multi-view images remains challenging. Recent self-supervised methods attempt to unify geometry, appearance, and semantics in a feed-forward manner, but they often suffer from weak geometry induction, limited appearance detail, and inconsistencies between geometry and semantics.We introduce UniSplat, a feed-forward framework designed to address these limitations through three complementary components. First, we propose a dual-masking strategy that strengthens geometry induction in the encoder. By masking both encoder and decoder tokens, and targeting decoder masks toward geometry-rich regions, the model is forced to infer structural information from incomplete visual cues, yielding geometry-aware representations even under unposed inputs.Second, we develop a coarse-to-fine Gaussian splatting strategy that reduces appearance-semantics inconsistencies by progressively refining the radiance field.Finally, to enforce geometric-semantic consistency, we introduce a pose-conditioned recalibration mechanism that interrelates the outputs of multiple heads by reprojecting predicted 3D point and semantic maps into the image plane using estimated camera parameters, and aligning them with corresponding RGB and semantic predictions to ensure cross-task consistency and resolving geometry-semantic mismatches. Together, these components yield unified 3D representations that are robust to unposed, sparse-view inputs and generalize across diverse tasks, laying a perceptual foundation for spatial intelligence.

Junbo Zhang, Hang Su, Zhaofan Li, Hang Dong, Chao Sun

Audio-Visual Segmentation (AVS) aims to accurately segment sounding objects in video frames by leveraging audio-visual correspondence cues. However, it remains challenging due to the intrinsic semantic incompleteness within a single modality and the semantic gap between audio and visual representations. Traditional feature-fusion decoding approaches struggle to suppress fusion noise effectively, while recent methods that incorporate data-dependent priors often increase the complexity of modeling audio-visual correlations, leading to poor cross-domain generalization. To address these issues, we propose a novel adaptive contrastive and prototype learning framework, BYOAVP, for AVS. Specifically, we design a Self-Supervised Audio Enhancement (SSAE) module that introduces contrastive learning to adaptively align audio representations with gradient-blocked visual embeddings, thus narrowing the semantic gap between modalities. Furthermore, a Dynamic Prototype Constraint (DPC) module is developed to refine pixel-wise category perception via momentum-based prototype updating, while enhancing localization of sounding regions through cross-modal interaction. Extensive experiments demonstrate that our method achieves SOTA performance across two AVS benchmarks and six sub-tasks, exhibiting strong robustness and generalization ability.

Yik San Cheng, Runkai Zhao, Weidong Cai

2D visual foundation models, such as DINOv3, a self-supervised model trained on large-scale natural images, have demonstrated strong zero-shot generalization, capturing rich global context and fine-grained structural cues. However, an analogous 3D foundation model for downstream volumetric neuroimaging remains lacking, largely due to the challenges of 3D image acquisition and the scarcity of high-quality annotations. To address this gap, we propose to adapt the 2D visual representations learned by DINOv3 to a 3D biomedical segmentation model, enabling more data-efficient and morphologically faithful neuronal reconstruction. Specifically, we design an inflation-based adaptation strategy that inflates 2D filters into 3D operators, preserving semantic priors from DINOv3 while adapting to 3D neuronal volume patches. In addition, we introduce a topology-aware skeleton loss to explicitly enforce structural fidelity of graph-based neuronal arbor reconstruction. Extensive experiments on four neuronal imaging datasets, including two from BigNeuron and two public datasets, NeuroFly and CWMBS, demonstrate consistent improvements in reconstruction accuracy over SoTA methods, with average gains of 2.9% in Entire Structure Average, 2.8% in Different Structure Average, and 3.8% in Percentage of Different Structure.

Jiayuan Chen, Ruoqi Liu, Zishan Gu, Ping Zhang

Microscopy-based phenotypic profiling is scalable for drug discovery but lacks the mechanistic depth of transcriptomics, which remains costly and scarce. Existing multimodal approaches either use images to support other modalities or naively align representations by sample identity, ignoring cell-type and dose variations in weakly paired data-limiting generalization to unseen interventions. In this paper, we introduce an intervention-aware distillation framework that leverages perturbational transcriptomics to guide image representation learning. A transcriptome-conditioned teacher integrates gene expression and intervention metadata to produce soft distributions over a chemistry-aware codebook organized by drug similarity. The teacher employs a fine-tuned single-cell foundation model to encode cell-type context and disentangle dose effects. An image-only student learns to predict these distributions from microscopy alone, distilling mechanistic knowledge while operating independently at test time. This design emphasizes intervention semantics rather than identity alignment and explicitly handles dose and cell-type mismatches. We provide theoretical guarantees showing that transcriptomic guidance tightens the risk bound for image-based prediction. On Cell Painting and RxRx datasets paired with L1000, our method significantly improves one-shot transfer to unseen interventions and drug-target gene discovery compared to self-supervised and alignment baselines