Daily Papers

Conditional human motion generation remains a fundamental challenge in computer vision and robotics. Despite significant progress, current methods are often constrained by fixed modality configurations and task-specific architectures, leaving cross-modal interactions and the scaling laws of multimodal-conditioned synthesis largely underexplored. A key bottleneck is the scarcity of large-scale modality-aligned motion data, limiting generalization across diverse control signals. In this work, we introduce OmniHuMo, a large-scale, high-quality dataset comprising over 5,000 hours of motion and 3.2 million sequences with precisely aligned multimodal annotations (e.g., text, speech, music, and trajectory). Leveraging OmniHuMo, we propose AnyMo, a unified multimodal framework combining a Residual FSQ-based motion tokenizer with a scalable masked modeling transformer, enabling high-quality motion synthesis under arbitrary modality combinations. Extensive experiments show that AnyMo achieves high-fidelity synthesis while offering flexible control over both spatial and stylistic attributes.

3D indoor scene generation is an important problem for the design of digital and real-world environments. To automate this process, a scene generation model should be able to not only generate plausible scene layouts, but also take into consideration visual features and style preferences. Existing methods for this task exhibit very limited control over these attributes, only allowing text inputs in the form of simple object-level descriptions or pairwise spatial relationships. Our proposed method Decorum enables users to control the scene generation process with natural language by adopting language-based representations at each stage. This enables us to harness recent advancements in Large Language Models (LLMs) to model language-to-language mappings. In addition, we show that using a text-based representation allows us to select furniture for our scenes using a novel object retrieval method based on multimodal LLMs. Evaluations on the benchmark 3D-FRONT dataset show that our methods achieve improvements over existing work in text-conditioned scene synthesis and object retrieval.

Synthesizing controllable 6-DOF object manipulation trajectories in 3D environments is essential for enabling robots to interact with complex scenes, yet remains challenging due to the need for accurate spatial reasoning, physical feasibility, and multimodal scene understanding. Existing approaches often rely on 2D or partial 3D representations, limiting their ability to capture full scene geometry and constraining trajectory precision. We present GMT, a multimodal transformer framework that generates realistic and goal-directed object trajectories by jointly leveraging 3D bounding box geometry, point cloud context, semantic object categories, and target end poses. The model represents trajectories as continuous 6-DOF pose sequences and employs a tailored conditioning strategy that fuses geometric, semantic, contextual, and goaloriented information. Extensive experiments on synthetic and real-world benchmarks demonstrate that GMT outperforms state-of-the-art human motion and human-object interaction baselines, such as CHOIS and GIMO, achieving substantial gains in spatial accuracy and orientation control. Our method establishes a new benchmark for learningbased manipulation planning and shows strong generalization to diverse objects and cluttered 3D environments. Project page: https://huajian- zeng.github. io/projects/gmt/.

Recent advances in diffusion-based video generation have substantially improved visual fidelity and temporal coherence. However, most existing approaches remain task-specific and rely primarily on textual instructions, limiting their ability to handle multimodal inputs, contextual references, and diverse video generation and editing scenarios within a unified framework. Moreover, many video editing methods depend on carefully engineered pipelines tailored to individual operations, which hinders scalability and composability. In this paper, we propose Tele-Omni, a unified multimodal framework for video generation and editing that follows multimodal instructions, including text, images, and reference videos, within a single model. Tele-Omni leverages pretrained multimodal large language models to parse heterogeneous instructions and infer structured generation or editing intents, while diffusion-based generators perform high-quality video synthesis conditioned on these structured signals. To enable joint training across heterogeneous video tasks, we introduce a task-aware data processing pipeline that unifies multimodal inputs into a structured instruction format while preserving task-specific constraints. Tele-Omni supports a wide range of video-centric tasks, including text-to-video generation, image-to-video generation, first-last-frame video generation, in-context video generation, and in-context video editing. By decoupling instruction parsing from video synthesis and combining it with task-aware data design, Tele-Omni achieves flexible multimodal control while maintaining strong temporal coherence and visual consistency. Experimental results demonstrate that Tele-Omni achieves competitive performance across multiple tasks.

Image generation models have evolved from text-conditioned pixel synthesis toward multimodal agents endowed with visual comprehension and tool invocation capabilities. Yet, existing agents remain at the mercy of underlying black-box image models. Their workflow is trapped in a repetitive cycle of prompt rewriting for generation refinement, leaving them with no mechanism to directly manipulate the canvas. In essence, the potential of LLMs to serve as a genuine "brush" for precise visual construction remains largely untapped. In this paper, we propose GenClaw, a code-driven agentic image generation paradigm that empowers the agent to create like a human artist: first conceptualizing, then sketching, and finally coloring. Specifically, the agent first constructs the conceptual knowledge and context through search and reasoning. It then utilizes code (e.g., SVG, HTML, Three.js) to render executable visual sketches. Finally, it employs an image generation model to supplement textures, materials, and photorealism. In this workflow, code serves as a controllable intermediate canvas bridging linguistic reasoning and pixel synthesis, seamlessly integrating programmatic logic with the visual expressiveness of generative models. By transforming image generation from a black-box paradigm into a staged process akin to authentic human creation, GenClaw offers a step toward for highly controllable and interpretable visual generation systems.

Diffusion probabilistic models (DPMs) have become a popular approach to conditional generation, due to their promising results and support for cross-modal synthesis. A key desideratum in conditional synthesis is to achieve high correspondence between the conditioning input and generated output. Most existing methods learn such relationships implicitly, by incorporating the prior into the variational lower bound. In this work, we take a different route -- we explicitly enhance input-output connections by maximizing their mutual information. To this end, we introduce a Conditional Discrete Contrastive Diffusion (CDCD) loss and design two contrastive diffusion mechanisms to effectively incorporate it into the denoising process, combining the diffusion training and contrastive learning for the first time by connecting it with the conventional variational objectives. We demonstrate the efficacy of our approach in evaluations with diverse multimodal conditional synthesis tasks: dance-to-music generation, text-to-image synthesis, as well as class-conditioned image synthesis. On each, we enhance the input-output correspondence and achieve higher or competitive general synthesis quality. Furthermore, the proposed approach improves the convergence of diffusion models, reducing the number of required diffusion steps by more than 35% on two benchmarks, significantly increasing the inference speed.

Along with the explosion of large language models, improvements in speech synthesis, advancements in hardware, and the evolution of computer graphics, the current bottleneck in creating digital humans lies in generating character movements that correspond naturally to text or speech inputs. In this work, we present DeepGesture, a diffusion-based gesture synthesis framework for generating expressive co-speech gestures conditioned on multimodal signals - text, speech, emotion, and seed motion. Built upon the DiffuseStyleGesture model, DeepGesture introduces novel architectural enhancements that improve semantic alignment and emotional expressiveness in generated gestures. Specifically, we integrate fast text transcriptions as semantic conditioning and implement emotion-guided classifier-free diffusion to support controllable gesture generation across affective states. To visualize results, we implement a full rendering pipeline in Unity based on BVH output from the model. Evaluation on the ZeroEGGS dataset shows that DeepGesture produces gestures with improved human-likeness and contextual appropriateness. Our system supports interpolation between emotional states and demonstrates generalization to out-of-distribution speech, including synthetic voices - marking a step forward toward fully multimodal, emotionally aware digital humans. Project page: https://deepgesture.github.io

Existing multimodal conditional image synthesis (MCIS) methods generate images conditioned on any combinations of various modalities that require all of them must be exactly conformed, hindering the synthesis controllability and leaving the potential of cross-modality under-exploited. To this end, we propose to generate images conditioned on the compositions of multimodal control signals, where modalities are imperfectly complementary, i.e., composed multimodal conditional image synthesis (CMCIS). Specifically, we observe two challenging issues of the proposed CMCIS task, i.e., the modality coordination problem and the modality imbalance problem. To tackle these issues, we introduce a Mixture-of-Modality-Tokens Transformer (MMoT) that adaptively fuses fine-grained multimodal control signals, a multimodal balanced training loss to stabilize the optimization of each modality, and a multimodal sampling guidance to balance the strength of each modality control signal. Comprehensive experimental results demonstrate that MMoT achieves superior performance on both unimodal conditional image synthesis (UCIS) and MCIS tasks with high-quality and faithful image synthesis on complex multimodal conditions. The project website is available at https://jabir-zheng.github.io/MMoT.

We propose to synthesize high-quality and synchronized audio, given video and optional text conditions, using a novel multimodal joint training framework MMAudio. In contrast to single-modality training conditioned on (limited) video data only, MMAudio is jointly trained with larger-scale, readily available text-audio data to learn to generate semantically aligned high-quality audio samples. Additionally, we improve audio-visual synchrony with a conditional synchronization module that aligns video conditions with audio latents at the frame level. Trained with a flow matching objective, MMAudio achieves new video-to-audio state-of-the-art among public models in terms of audio quality, semantic alignment, and audio-visual synchronization, while having a low inference time (1.23s to generate an 8s clip) and just 157M parameters. MMAudio also achieves surprisingly competitive performance in text-to-audio generation, showing that joint training does not hinder single-modality performance. Code and demo are available at: https://hkchengrex.github.io/MMAudio

Generating audio that is acoustically consistent with a scene is essential for immersive virtual environments. Recent neural acoustic field methods enable spatially continuous sound rendering but remain scene-specific, requiring dense audio measurements and costly training for each environment. Few-shot approaches improve scalability across rooms but still rely on multiple recordings and, being deterministic, fail to capture the inherent uncertainty of scene acoustics under sparse context. We introduce flow-matching acoustic generation (FLAC), a probabilistic method for few-shot acoustic synthesis that models the distribution of plausible room impulse responses (RIRs) given minimal scene context. FLAC leverages a diffusion transformer trained with a flow-matching objective to generate RIRs at arbitrary positions in novel scenes, conditioned on spatial, geometric, and acoustic cues. FLAC outperforms state-of-the-art eight-shot baselines with one-shot on both the AcousticRooms and Hearing Anything Anywhere datasets. To complement standard perceptual metrics, we further introduce AGREE, a joint acoustic-geometry embedding, enabling geometry-consistent evaluation of generated RIRs through retrieval and distributional metrics. This work is the first to apply generative flow matching to explicit RIR synthesis, establishing a new direction for robust and data-efficient acoustic synthesis.

Recent advances in audio-driven avatar video generation have significantly enhanced audio-visual realism. However, existing methods treat instruction conditioning merely as low-level tracking driven by acoustic or visual cues, without modeling the communicative purpose conveyed by the instructions. This limitation compromises their narrative coherence and character expressiveness. To bridge this gap, we introduce Kling-Avatar, a novel cascaded framework that unifies multimodal instruction understanding with photorealistic portrait generation. Our approach adopts a two-stage pipeline. In the first stage, we design a multimodal large language model (MLLM) director that produces a blueprint video conditioned on diverse instruction signals, thereby governing high-level semantics such as character motion and emotions. In the second stage, guided by blueprint keyframes, we generate multiple sub-clips in parallel using a first-last frame strategy. This global-to-local framework preserves fine-grained details while faithfully encoding the high-level intent behind multimodal instructions. Our parallel architecture also enables fast and stable generation of long-duration videos, making it suitable for real-world applications such as digital human livestreaming and vlogging. To comprehensively evaluate our method, we construct a benchmark of 375 curated samples covering diverse instructions and challenging scenarios. Extensive experiments demonstrate that Kling-Avatar is capable of generating vivid, fluent, long-duration videos at up to 1080p and 48 fps, achieving superior performance in lip synchronization accuracy, emotion and dynamic expressiveness, instruction controllability, identity preservation, and cross-domain generalization. These results establish Kling-Avatar as a new benchmark for semantically grounded, high-fidelity audio-driven avatar synthesis.

Recent progress in video-to-video (V2V) translation has enabled realistic resimulation of embodied AI demonstrations, a capability that allows pretrained robot policies to be transferable to new environments without additional data collection. However, prior works can only operate on a single view at a time, while embodied AI tasks are commonly captured from multiple synchronized cameras to support policy learning. Naively applying single-view models independently to each camera leads to inconsistent appearance across views, and standard transformer architectures do not scale to multi-view settings due to the quadratic cost of cross-view attention. We present VideoWeaver, the first multimodal multi-view V2V translation framework. VideoWeaver is initially trained as a single-view flow-based V2V model. To achieve an extension to the multi-view regime, we propose to ground all views in a shared 4D latent space derived from a feed-forward spatial foundation model, namely, Pi3. This encourages view-consistent appearance even under wide baselines and dynamic camera motion. To scale beyond a fixed number of cameras, we train views at distinct diffusion timesteps, enabling the model to learn both joint and conditional view distributions. This in turn allows autoregressive synthesis of new viewpoints conditioned on existing ones. Experiments show superior or similar performance to the state-of-the-art on the single-view translation benchmarks and, for the first time, physically and stylistically consistent multi-view translations, including challenging egocentric and heterogeneous-camera setups central to world randomization for robot learning.

Recent breakthroughs of transformer-based diffusion models, particularly with Multimodal Diffusion Transformers (MMDiT) driven models like FLUX and Qwen Image, have facilitated thrilling experiences in text-to-image generation and editing. To understand the internal mechanism of MMDiT-based models, existing methods tried to analyze the effect of specific components like positional encoding and attention layers. Yet, a comprehensive understanding of how different blocks and their interactions with textual conditions contribute to the synthesis process remains elusive. In this paper, we first develop a systematic pipeline to comprehensively investigate each block's functionality by removing, disabling and enhancing textual hidden-states at corresponding blocks. Our analysis reveals that 1) semantic information appears in earlier blocks and finer details are rendered in later blocks, 2) removing specific blocks is usually less disruptive than disabling text conditions, and 3) enhancing textual conditions in selective blocks improves semantic attributes. Building on these observations, we further propose novel training-free strategies for improved text alignment, precise editing, and acceleration. Extensive experiments demonstrated that our method outperforms various baselines and remains flexible across text-to-image generation, image editing, and inference acceleration. Our method improves T2I-Combench++ from 56.92% to 63.00% and GenEval from 66.42% to 71.63% on SD3.5, without sacrificing synthesis quality. These results advance understanding of MMDiT models and provide valuable insights to unlock new possibilities for further improvements.

AI-assisted graphic design has emerged as a powerful tool for automating the creation and editing of design elements such as posters, banners, and advertisements. While diffusion-based text-to-image models have demonstrated strong capabilities in visual content generation, their text rendering performance, particularly for small-scale typography and non-Latin scripts, remains limited. In this paper, we propose UTDesign, a unified framework for high-precision stylized text editing and conditional text generation in design images, supporting both English and Chinese scripts. Our framework introduces a novel DiT-based text style transfer model trained from scratch on a synthetic dataset, capable of generating transparent RGBA text foregrounds that preserve the style of reference glyphs. We further extend this model into a conditional text generation framework by training a multi-modal condition encoder on a curated dataset with detailed text annotations, enabling accurate, style-consistent text synthesis conditioned on background images, prompts, and layout specifications. Finally, we integrate our approach into a fully automated text-to-design (T2D) pipeline by incorporating pre-trained text-to-image (T2I) models and an MLLM-based layout planner. Extensive experiments demonstrate that UTDesign achieves state-of-the-art performance among open-source methods in terms of stylistic consistency and text accuracy, and also exhibits unique advantages compared to proprietary commercial approaches. Code and data for this paper are available at https://github.com/ZYM-PKU/UTDesign.

Biomedical research increasingly relies on integrating diverse data modalities, including gene expression profiles, medical images, and clinical metadata. While medical images and clinical metadata are routinely collected in clinical practice, gene expression data presents unique challenges for widespread research use, mainly due to stringent privacy regulations and costly laboratory experiments. To address these limitations, we present GeMM-GAN, a novel Generative Adversarial Network conditioned on histopathology tissue slides and clinical metadata, designed to synthesize realistic gene expression profiles. GeMM-GAN combines a Transformer Encoder for image patches with a final Cross Attention mechanism between patches and text tokens, producing a conditioning vector to guide a generative model in generating biologically coherent gene expression profiles. We evaluate our approach on the TCGA dataset and demonstrate that our framework outperforms standard generative models and generates more realistic and functionally meaningful gene expression profiles, improving by more than 11\% the accuracy on downstream disease type prediction compared to current state-of-the-art generative models. Code will be available at: https://github.com/francescapia/GeMM-GAN

In this paper, we approach an overlooked yet critical task Graph2Image: generating images from multimodal attributed graphs (MMAGs). This task poses significant challenges due to the explosion in graph size, dependencies among graph entities, and the need for controllability in graph conditions. To address these challenges, we propose a graph context-conditioned diffusion model called InstructG2I. InstructG2I first exploits the graph structure and multimodal information to conduct informative neighbor sampling by combining personalized page rank and re-ranking based on vision-language features. Then, a Graph-QFormer encoder adaptively encodes the graph nodes into an auxiliary set of graph prompts to guide the denoising process of diffusion. Finally, we propose graph classifier-free guidance, enabling controllable generation by varying the strength of graph guidance and multiple connected edges to a node. Extensive experiments conducted on three datasets from different domains demonstrate the effectiveness and controllability of our approach. The code is available at https://github.com/PeterGriffinJin/InstructG2I.

We present Compound Conditioned ControlNet, C3Net, a novel generative neural architecture taking conditions from multiple modalities and synthesizing multimodal contents simultaneously (e.g., image, text, audio). C3Net adapts the ControlNet architecture to jointly train and make inferences on a production-ready diffusion model and its trainable copies. Specifically, C3Net first aligns the conditions from multi-modalities to the same semantic latent space using modality-specific encoders based on contrastive training. Then, it generates multimodal outputs based on the aligned latent space, whose semantic information is combined using a ControlNet-like architecture called Control C3-UNet. Correspondingly, with this system design, our model offers an improved solution for joint-modality generation through learning and explaining multimodal conditions instead of simply taking linear interpolations on the latent space. Meanwhile, as we align conditions to a unified latent space, C3Net only requires one trainable Control C3-UNet to work on multimodal semantic information. Furthermore, our model employs unimodal pretraining on the condition alignment stage, outperforming the non-pretrained alignment even on relatively scarce training data and thus demonstrating high-quality compound condition generation. We contribute the first high-quality tri-modal validation set to validate quantitatively that C3Net outperforms or is on par with first and contemporary state-of-the-art multimodal generation. Our codes and tri-modal dataset will be released.

Achieving realistic, vivid, and human-like synthesized conversational gestures conditioned on multi-modal data is still an unsolved problem due to the lack of available datasets, models and standard evaluation metrics. To address this, we build Body-Expression-Audio-Text dataset, BEAT, which has i) 76 hours, high-quality, multi-modal data captured from 30 speakers talking with eight different emotions and in four different languages, ii) 32 millions frame-level emotion and semantic relevance annotations. Our statistical analysis on BEAT demonstrates the correlation of conversational gestures with facial expressions, emotions, and semantics, in addition to the known correlation with audio, text, and speaker identity. Based on this observation, we propose a baseline model, Cascaded Motion Network (CaMN), which consists of above six modalities modeled in a cascaded architecture for gesture synthesis. To evaluate the semantic relevancy, we introduce a metric, Semantic Relevance Gesture Recall (SRGR). Qualitative and quantitative experiments demonstrate metrics' validness, ground truth data quality, and baseline's state-of-the-art performance. To the best of our knowledge, BEAT is the largest motion capture dataset for investigating human gestures, which may contribute to a number of different research fields, including controllable gesture synthesis, cross-modality analysis, and emotional gesture recognition. The data, code and model are available on https://pantomatrix.github.io/BEAT/.

Acquiring aligned visuo-tactile datasets is slow and costly, requiring specialised hardware and large-scale data collection. Synthetic generation is promising, but prior methods are typically single-modality, limiting cross-modal learning. We present MultiDiffSense, a unified diffusion model that synthesises images for multiple vision-based tactile sensors (ViTac, TacTip, ViTacTip) within a single architecture. Our approach uses dual conditioning on CAD-derived, pose-aligned depth maps and structured prompts that encode sensor type and 4-DoF contact pose, enabling controllable, physically consistent multi-modal synthesis. Evaluating on 8 objects (5 seen, 3 novel) and unseen poses, MultiDiffSense outperforms a Pix2Pix cGAN baseline in SSIM by +36.3% (ViTac), +134.6% (ViTacTip), and +64.7% (TacTip). For downstream 3-DoF pose estimation, mixing 50% synthetic with 50% real halves the required real data while maintaining competitive performance. MultiDiffSense alleviates the data-collection bottleneck in tactile sensing and enables scalable, controllable multi-modal dataset generation for robotic applications.

In this work, we study Human-Object Interaction Video Generation (HOIVG), which aims to synthesize high-quality human-object interaction videos conditioned on text, reference images, audio, and pose. This task holds significant practical value for automating content creation in real-world applications, such as e-commerce demonstrations, short video production, and interactive entertainment. However, existing approaches fail to accommodate all these requisite conditions. We present OmniShow, an end-to-end framework tailored for this practical yet challenging task, capable of harmonizing multimodal conditions and delivering industry-grade performance. To overcome the trade-off between controllability and quality, we introduce Unified Channel-wise Conditioning for efficient image and pose injection, and Gated Local-Context Attention to ensure precise audio-visual synchronization. To effectively address data scarcity, we develop a Decoupled-Then-Joint Training strategy that leverages a multi-stage training process with model merging to efficiently harness heterogeneous sub-task datasets. Furthermore, to fill the evaluation gap in this field, we establish HOIVG-Bench, a dedicated and comprehensive benchmark for HOIVG. Extensive experiments demonstrate that OmniShow achieves overall state-of-the-art performance across various multimodal conditioning settings, setting a solid standard for the emerging HOIVG task.

Text-driven human motion generation is a multimodal task that synthesizes human motion sequences conditioned on natural language. It requires the model to satisfy textual descriptions under varying conditional inputs, while generating plausible and realistic human actions with high diversity. Existing diffusion model-based approaches have outstanding performance in the diversity and multimodality of generation. However, compared to autoregressive methods that train motion encoders before inference, diffusion methods lack in fitting the distribution of human motion features which leads to an unsatisfactory FID score. One insight is that the diffusion model lack the ability to learn the motion relations among spatio-temporal semantics through contextual reasoning. To solve this issue, in this paper, we proposed Motion Masked Diffusion Model (MMDM), a novel human motion masked mechanism for diffusion model to explicitly enhance its ability to learn the spatio-temporal relationships from contextual joints among motion sequences. Besides, considering the complexity of human motion data with dynamic temporal characteristics and spatial structure, we designed two mask modeling strategies: time frames mask and body parts mask. During training, MMDM masks certain tokens in the motion embedding space. Then, the diffusion decoder is designed to learn the whole motion sequence from masked embedding in each sampling step, this allows the model to recover a complete sequence from incomplete representations. Experiments on HumanML3D and KIT-ML dataset demonstrate that our mask strategy is effective by balancing motion quality and text-motion consistency.

Customized video generation aims to produce videos featuring specific subjects under flexible user-defined conditions, yet existing methods often struggle with identity consistency and limited input modalities. In this paper, we propose HunyuanCustom, a multi-modal customized video generation framework that emphasizes subject consistency while supporting image, audio, video, and text conditions. Built upon HunyuanVideo, our model first addresses the image-text conditioned generation task by introducing a text-image fusion module based on LLaVA for enhanced multi-modal understanding, along with an image ID enhancement module that leverages temporal concatenation to reinforce identity features across frames. To enable audio- and video-conditioned generation, we further propose modality-specific condition injection mechanisms: an AudioNet module that achieves hierarchical alignment via spatial cross-attention, and a video-driven injection module that integrates latent-compressed conditional video through a patchify-based feature-alignment network. Extensive experiments on single- and multi-subject scenarios demonstrate that HunyuanCustom significantly outperforms state-of-the-art open- and closed-source methods in terms of ID consistency, realism, and text-video alignment. Moreover, we validate its robustness across downstream tasks, including audio and video-driven customized video generation. Our results highlight the effectiveness of multi-modal conditioning and identity-preserving strategies in advancing controllable video generation. All the code and models are available at https://hunyuancustom.github.io.

Different modalities hold considerable gaps in optimization trajectories, including speeds and paths, which lead to modality laziness and modality clash when jointly training multimodal models, resulting in insufficient and imbalanced multimodal learning. Existing methods focus on enforcing the weak modality by adding modality-specific optimization objectives, aligning their optimization speeds, or decomposing multimodal learning to enhance unimodal learning. These methods fail to achieve both unimodal sufficiency and multimodal balance. In this paper, we, for the first time, address both concerns by proposing multimodal Data Remixing, including decoupling multimodal data and filtering hard samples for each modality to mitigate modality imbalance; and then batch-level reassembling to align the gradient directions and avoid cross-modal interference, thus enhancing unimodal learning sufficiency. Experimental results demonstrate that our method can be seamlessly integrated with existing approaches, improving accuracy by approximately 6.50%uparrow on CREMAD and 3.41%uparrow on Kinetic-Sounds, without training set expansion or additional computational overhead during inference. The source code is available at https://github.com/MatthewMaxy/Remix_ICML2025.

This work introduces a new task, text-conditioned selective video-to-audio (V2A) generation, which produces only the user-intended sound from a multi-object video. This capability is especially crucial in multimedia production, where audio tracks are handled individually for each sound source for precise editing, mixing, and creative control. However, current approaches generate single source-mixed sounds at once, largely because visual features are entangled, and region cues or prompts often fail to specify the source. We propose SelVA, a novel text-conditioned V2A model that treats the text prompt as an explicit selector of target source and modulates video encoder to distinctly extract prompt-relevant video features. The proposed supplementary tokens promote cross-attention by suppressing text-irrelevant activations with efficient parameter tuning, yielding robust semantic and temporal grounding. SelVA further employs a self-augmentation scheme to overcome the lack of mono audio track supervision. We evaluate SelVA on VGG-MONOAUDIO, a curated benchmark of clean single-source videos for such a task. Extensive experiments and ablations consistently verify its effectiveness across audio quality, semantic alignment, and temporal synchronization. Code and demo are available at https://jnwnlee.github.io/selva-demo/.

Multimodal generative models have shown remarkable progress in single-modality video and audio synthesis, yet truly joint audio-video generation remains an open challenge. In this paper, I explore four key contributions to advance this field. First, I release two high-quality, paired audio-video datasets. The datasets consisting on 13 hours of video-game clips and 64 hours of concert performances, each segmented into consistent 34-second samples to facilitate reproducible research. Second, I train the MM-Diffusion architecture from scratch on our datasets, demonstrating its ability to produce semantically coherent audio-video pairs and quantitatively evaluating alignment on rapid actions and musical cues. Third, I investigate joint latent diffusion by leveraging pretrained video and audio encoder-decoders, uncovering challenges and inconsistencies in the multimodal decoding stage. Finally, I propose a sequential two-step text-to-audio-video generation pipeline: first generating video, then conditioning on both the video output and the original prompt to synthesize temporally synchronized audio. My experiments show that this modular approach yields high-fidelity generations of audio video generation.

Unified Multimodal Models (UMMs) have demonstrated remarkable performance in text-to-image generation (T2I) and editing (TI2I), whether instantiated as assembled unified frameworks which couple powerful vision-language model (VLM) with diffusion-based generator, or as naive Unified Multimodal Models with an early fusion of understanding and generation modalities. We contend that in current unified frameworks, the crucial capability of multimodal generative reasoning which encompasses instruction understanding, grounding, and image referring for identity preservation and faithful reconstruction, is intrinsically entangled with high-fidelity synthesis. In this work, we introduce Query-Kontext, a novel approach that bridges the VLM and diffusion model via a multimodal ``kontext'' composed of semantic cues and coarse-grained image conditions encoded from multimodal inputs. This design delegates the complex ability of multimodal generative reasoning to powerful VLM while reserving diffusion model's role for high-quality visual synthesis. To achieve this, we propose a three-stage progressive training strategy. First, we connect the VLM to a lightweight diffusion head via multimodal kontext tokens to unleash the VLM's generative reasoning ability. Second, we scale this head to a large, pre-trained diffusion model to enhance visual detail and realism. Finally, we introduce a low-level image encoder to improve image fidelity and perform instruction tuning on downstream tasks. Furthermore, we build a comprehensive data pipeline integrating real, synthetic, and open-source datasets, covering diverse multimodal reference-to-image scenarios, including image generation, instruction-driven editing, customized generation, and multi-subject composition. Experiments show that our approach matches strong unified baselines and even outperforms task-specific state-of-the-art methods in several cases.

Recent work has studied text-to-audio synthesis using large amounts of paired text-audio data. However, audio recordings with high-quality text annotations can be difficult to acquire. In this work, we approach text-to-audio synthesis using unlabeled videos and pretrained language-vision models. We propose to learn the desired text-audio correspondence by leveraging the visual modality as a bridge. We train a conditional diffusion model to generate the audio track of a video, given a video frame encoded by a pretrained contrastive language-image pretraining (CLIP) model. At test time, we first explore performing a zero-shot modality transfer and condition the diffusion model with a CLIP-encoded text query. However, we observe a noticeable performance drop with respect to image queries. To close this gap, we further adopt a pretrained diffusion prior model to generate a CLIP image embedding given a CLIP text embedding. Our results show the effectiveness of the proposed method, and that the pretrained diffusion prior can reduce the modality transfer gap. While we focus on text-to-audio synthesis, the proposed model can also generate audio from image queries, and it shows competitive performance against a state-of-the-art image-to-audio synthesis model in a subjective listening test. This study offers a new direction of approaching text-to-audio synthesis that leverages the naturally-occurring audio-visual correspondence in videos and the power of pretrained language-vision models.

Recent progress in image generation has sparked research into controlling these models through condition signals, with various methods addressing specific challenges in conditional generation. Instead of proposing another specialized technique, we introduce a simple, unified framework to handle diverse conditional generation tasks involving a specific image-condition correlation. By learning a joint distribution over a correlated image pair (e.g. image and depth) with a diffusion model, our approach enables versatile capabilities via different inference-time sampling schemes, including controllable image generation (e.g. depth to image), estimation (e.g. image to depth), signal guidance, joint generation (image & depth), and coarse control. Previous attempts at unification often introduce significant complexity through multi-stage training, architectural modification, or increased parameter counts. In contrast, our simple formulation requires a single, computationally efficient training stage, maintains the standard model input, and adds minimal learned parameters (15% of the base model). Moreover, our model supports additional capabilities like non-spatially aligned and coarse conditioning. Extensive results show that our single model can produce comparable results with specialized methods and better results than prior unified methods. We also demonstrate that multiple models can be effectively combined for multi-signal conditional generation.

The recently developed discrete diffusion models perform extraordinarily well in the text-to-image task, showing significant promise for handling the multi-modality signals. In this work, we harness these traits and present a unified multimodal generation model that can conduct both the "modality translation" and "multi-modality generation" tasks using a single model, performing text-based, image-based, and even vision-language simultaneous generation. Specifically, we unify the discrete diffusion process for multimodal signals by proposing a unified transition matrix. Moreover, we design a mutual attention module with fused embedding layer and a unified objective function to emphasise the inter-modal linkages, which are vital for multi-modality generation. Extensive experiments indicate that our proposed method can perform comparably to the state-of-the-art solutions in various generation tasks.

When trained at sufficient scale, auto-regressive language models exhibit the notable ability to learn a new language task after being prompted with just a few examples. Here, we present a simple, yet effective, approach for transferring this few-shot learning ability to a multimodal setting (vision and language). Using aligned image and caption data, we train a vision encoder to represent each image as a sequence of continuous embeddings, such that a pre-trained, frozen language model prompted with this prefix generates the appropriate caption. The resulting system is a multimodal few-shot learner, with the surprising ability to learn a variety of new tasks when conditioned on examples, represented as a sequence of multiple interleaved image and text embeddings. We demonstrate that it can rapidly learn words for new objects and novel visual categories, do visual question-answering with only a handful of examples, and make use of outside knowledge, by measuring a single model on a variety of established and new benchmarks.

Music exists in various modalities, such as score images, symbolic scores, MIDI, and audio. Translations between each modality are established as core tasks of music information retrieval, such as automatic music transcription (audio-to-MIDI) and optical music recognition (score image to symbolic score). However, most past work on multimodal translation trains specialized models on individual translation tasks. In this paper, we propose a unified approach, where we train a general-purpose model on many translation tasks simultaneously. Two key factors make this unified approach viable: a new large-scale dataset and the tokenization of each modality. Firstly, we propose a new dataset that consists of more than 1,300 hours of paired audio-score image data collected from YouTube videos, which is an order of magnitude larger than any existing music modal translation datasets. Secondly, our unified tokenization framework discretizes score images, audio, MIDI, and MusicXML into a sequence of tokens, enabling a single encoder-decoder Transformer to tackle multiple cross-modal translation as one coherent sequence-to-sequence task. Experimental results confirm that our unified multitask model improves upon single-task baselines in several key areas, notably reducing the symbol error rate for optical music recognition from 24.58% to a state-of-the-art 13.67%, while similarly substantial improvements are observed across the other translation tasks. Notably, our approach achieves the first successful score-image-conditioned audio generation, marking a significant breakthrough in cross-modal music generation.

Generating high quality music that complements the visual content of a video is a challenging task. Most existing visual conditioned music generation systems generate symbolic music data, such as MIDI files, instead of raw audio waveform. Given the limited availability of symbolic music data, such methods can only generate music for a few instruments or for specific types of visual input. In this paper, we propose a novel approach called V2Meow that can generate high-quality music audio that aligns well with the visual semantics of a diverse range of video input types. Specifically, the proposed music generation system is a multi-stage autoregressive model which is trained with a number of O(100K) music audio clips paired with video frames, which are mined from in-the-wild music videos, and no parallel symbolic music data is involved. V2Meow is able to synthesize high-fidelity music audio waveform solely conditioned on pre-trained visual features extracted from an arbitrary silent video clip, and it also allows high-level control over the music style of generation examples via supporting text prompts in addition to the video frames conditioning. Through both qualitative and quantitative evaluations, we demonstrate that our model outperforms several existing music generation systems in terms of both visual-audio correspondence and audio quality.

The bifurcation of generative modeling into autoregressive approaches for discrete data (text) and diffusion approaches for continuous data (images) hinders the development of truly unified multimodal systems. While Masked Language Models (MLMs) offer efficient bidirectional context, they traditionally lack the generative fidelity of autoregressive models and the semantic continuity of diffusion models. Furthermore, extending masked generation to multimodal settings introduces severe alignment challenges and training instability. In this work, we propose CoM-DAD (Coupled Manifold Discrete Absorbing Diffusion), a novel probabilistic framework that reformulates multimodal generation as a hierarchical dual-process. CoM-DAD decouples high-level semantic planning from low-level token synthesis. First, we model the semantic manifold via a continuous latent diffusion process; second, we treat token generation as a discrete absorbing diffusion process, regulated by a Variable-Rate Noise Schedule, conditioned on these evolving semantic priors. Crucially, we introduce a Stochastic Mixed-Modal Transport strategy that aligns disparate modalities without requiring heavy contrastive dual-encoders. Our method demonstrates superior stability over standard masked modeling, establishing a new paradigm for scalable, unified text-image generation.

In this paper, we introduce MIO, a novel foundation model built on multimodal tokens, capable of understanding and generating speech, text, images, and videos in an end-to-end, autoregressive manner. While the emergence of large language models (LLMs) and multimodal large language models (MM-LLMs) propels advancements in artificial general intelligence through their versatile capabilities, they still lack true any-to-any understanding and generation. Recently, the release of GPT-4o has showcased the remarkable potential of any-to-any LLMs for complex real-world tasks, enabling omnidirectional input and output across images, speech, and text. However, it is closed-source and does not support the generation of multimodal interleaved sequences. To address this gap, we present MIO, which is trained on a mixture of discrete tokens across four modalities using causal multimodal modeling. MIO undergoes a four-stage training process: (1) alignment pre-training, (2) interleaved pre-training, (3) speech-enhanced pre-training, and (4) comprehensive supervised fine-tuning on diverse textual, visual, and speech tasks. Our experimental results indicate that MIO exhibits competitive, and in some cases superior, performance compared to previous dual-modal baselines, any-to-any model baselines, and even modality-specific baselines. Moreover, MIO demonstrates advanced capabilities inherent to its any-to-any feature, such as interleaved video-text generation, chain-of-visual-thought reasoning, visual guideline generation, instructional image editing, etc.

Unified vision-language models have made significant progress in multimodal understanding and generation, yet they largely fall short in producing multimodal interleaved outputs, which is a crucial capability for tasks like visual storytelling and step-by-step visual reasoning. In this work, we propose a reinforcement learning-based post-training strategy to unlock this capability in existing unified models, without relying on large-scale multimodal interleaved datasets. We begin with a warm-up stage using a hybrid dataset comprising curated interleaved sequences and limited data for multimodal understanding and text-to-image generation, which exposes the model to interleaved generation patterns while preserving its pretrained capabilities. To further refine interleaved generation, we propose a unified policy optimization framework that extends Group Relative Policy Optimization (GRPO) to the multimodal setting. Our approach jointly models text and image generation within a single decoding trajectory and optimizes it with our novel hybrid rewards covering textual relevance, visual-text alignment, and structural fidelity. Additionally, we incorporate process-level rewards to provide step-wise guidance, enhancing training efficiency in complex multimodal tasks. Experiments on MMIE and InterleavedBench demonstrate that our approach significantly enhances the quality and coherence of multimodal interleaved generation.

Current machine learning models for vision are often highly specialized and limited to a single modality and task. In contrast, recent large language models exhibit a wide range of capabilities, hinting at a possibility for similarly versatile models in computer vision. In this paper, we take a step in this direction and propose a multimodal training scheme called 4M. It consists of training a single unified Transformer encoder-decoder using a masked modeling objective across a wide range of input/output modalities - including text, images, geometric, and semantic modalities, as well as neural network feature maps. 4M achieves scalability by unifying the representation space of all modalities through mapping them into discrete tokens and performing multimodal masked modeling on a small randomized subset of tokens. 4M leads to models that exhibit several key capabilities: (1) they can perform a diverse set of vision tasks out of the box, (2) they excel when fine-tuned for unseen downstream tasks or new input modalities, and (3) they can function as a generative model that can be conditioned on arbitrary modalities, enabling a wide variety of expressive multimodal editing capabilities with remarkable flexibility. Through experimental analyses, we demonstrate the potential of 4M for training versatile and scalable foundation models for vision tasks, setting the stage for further exploration in multimodal learning for vision and other domains.

Visually-conditioned language models (VLMs) have seen growing adoption in applications such as visual dialogue, scene understanding, and robotic task planning; adoption that has fueled a wealth of new models such as LLaVa, InstructBLIP, and PaLI-3. Despite the volume of new releases, key design decisions around image preprocessing, architecture, and optimization are under-explored, making it challenging to understand what factors account for model performance - a challenge further complicated by the lack of objective, consistent evaluations. To address these gaps, we first compile a suite of standardized evaluations spanning visual question answering, object localization from language, and targeted challenge sets that probe properties such as hallucination; evaluations that provide calibrated, fine-grained insight into a VLM's capabilities. Second, we rigorously investigate VLMs along key design axes, including pretrained visual representations and quantifying the tradeoffs of using base vs. instruct-tuned language models, amongst others. We couple our analysis with three resource contributions: (1) a unified framework for evaluating VLMs, (2) optimized, flexible code for VLM training, and (3) checkpoints for all models, including a family of VLMs at the 7-13B scale that strictly outperform InstructBLIP and LLaVa v1.5, the state-of-the-art in open-source VLMs.

We propose a novel step-by-step video-to-audio generation method that sequentially produces individual audio tracks, each corresponding to a specific sound event in the video. Our approach mirrors traditional Foley workflows, aiming to capture all sound events induced by a given video comprehensively. Each generation step is formulated as a guided video-to-audio synthesis task, conditioned on a target text prompt and previously generated audio tracks. This design is inspired by the idea of concept negation from prior compositional generation frameworks. To enable this guided generation, we introduce a training framework that leverages pre-trained video-to-audio models and eliminates the need for specialized paired datasets, allowing training on more accessible data. Experimental results demonstrate that our method generates multiple semantically distinct audio tracks for a single input video, leading to higher-quality composite audio synthesis than existing baselines.

Multimodal learning helps to comprehensively understand the world, by integrating different senses. Accordingly, multiple input modalities are expected to boost model performance, but we actually find that they are not fully exploited even when the multimodal model outperforms its uni-modal counterpart. Specifically, in this paper we point out that existing multimodal discriminative models, in which uniform objective is designed for all modalities, could remain under-optimized uni-modal representations, caused by another dominated modality in some scenarios, e.g., sound in blowing wind event, vision in drawing picture event, etc. To alleviate this optimization imbalance, we propose on-the-fly gradient modulation to adaptively control the optimization of each modality, via monitoring the discrepancy of their contribution towards the learning objective. Further, an extra Gaussian noise that changes dynamically is introduced to avoid possible generalization drop caused by gradient modulation. As a result, we achieve considerable improvement over common fusion methods on different multimodal tasks, and this simple strategy can also boost existing multimodal methods, which illustrates its efficacy and versatility. The source code is available at https://github.com/GeWu-Lab/OGM-GE_CVPR2022.

Recent work has shown great progress in integrating spatial conditioning to control large, pre-trained text-to-image diffusion models. Despite these advances, existing methods describe the spatial image content using hand-crafted conditioning inputs, which are either semantically ambiguous (e.g., edges) or require expensive manual annotations (e.g., semantic segmentation). To address these limitations, we propose a new label-free way of conditioning diffusion models to enable fine-grained spatial control. We introduce the concept of neural semantic image synthesis, which uses neural layouts extracted from pre-trained foundation models as conditioning. Neural layouts are advantageous as they provide rich descriptions of the desired image, containing both semantics and detailed geometry of the scene. We experimentally show that images synthesized via neural semantic image synthesis achieve similar or superior pixel-level alignment of semantic classes compared to those created using expensive semantic label maps. At the same time, they capture better semantics, instance separation, and object orientation than other label-free conditioning options, such as edges or depth. Moreover, we show that images generated by neural layout conditioning can effectively augment real data for training various perception tasks.

Motion-centric video editing remains difficult for large generative video models, which often respond well to appearance changes but struggle to produce specific, localized actions or state transitions in an existing clip. We introduce Sound Sparks Motion, a training-free framework that enables motion editing in an audio-visual video generation model by tuning its internal multimodal conditioning signals at test time. Rather than modifying model weights, our method tunes only two lightweight variables: an audio latent derived from the source video and a residual perturbation in the text-conditioning. We find that this combination can encourage motion edits that the underlying model often struggles to realize under prompt-only control. Since there is no direct way to evaluate temporal alignment between text and motion, we guide the tuning process using a vision-language model that provides feedback indicating whether the intended motion appears in the generated video. This simple supervision yields an effective semantic objective for motion editing, while regularization and perceptual-temporal constraints help preserve content and visual quality. Beyond per-video tuning, we show that the learned latent controls are transferable across videos, suggesting that they capture reusable motion-edit directions rather than overfitting to a single example. Our results highlight multimodal conditioning tuning, particularly through the audio pathway, as a promising direction for motion-aware video editing, and suggest that test-time tuning can serve as a lightweight probing mechanism that helps reveal latent motion controls embedded in the model's multimodal conditioning. Code and data are available via our project page: https://amirhossein-razlighi.github.io/Sound_Sparks_Motion/

The goal of this work is to simultaneously generate natural talking faces and speech outputs from text. We achieve this by integrating Talking Face Generation (TFG) and Text-to-Speech (TTS) systems into a unified framework. We address the main challenges of each task: (1) generating a range of head poses representative of real-world scenarios, and (2) ensuring voice consistency despite variations in facial motion for the same identity. To tackle these issues, we introduce a motion sampler based on conditional flow matching, which is capable of high-quality motion code generation in an efficient way. Moreover, we introduce a novel conditioning method for the TTS system, which utilises motion-removed features from the TFG model to yield uniform speech outputs. Our extensive experiments demonstrate that our method effectively creates natural-looking talking faces and speech that accurately match the input text. To our knowledge, this is the first effort to build a multimodal synthesis system that can generalise to unseen identities.

In complex embodied long-horizon manipulation tasks, effective task decomposition and execution require synergistic integration of textual logical reasoning and visual-spatial imagination to ensure efficient and accurate operation. Current methods fail to adopt a unified generation framework for multimodal planning, lead to inconsistent in multimodal planning. To address this challenge, we present EVLP (Embodied Vision-Language Planner), an innovative multimodal unified generation framework that jointly models linguistic reasoning and visual generation. Our approach achieves multimodal planning for long-horizon tasks through a novel training pipeline incorporating dynamic pretraining and reinforced alignment. Our core innovations consist of three key components: 1) Unified Multimodal Generation Framework: For understanding, We integrate semantic information with spatial features to provide comprehensive visual perception. For generation, we directly learn the joint distribution of discrete images for one-step visual synthesis, enabling coordinated language-visual modeling through learnable cross-modal attention mechanisms. 2) Dynamic Perception Pretraining: We propose a bidirectional dynamic alignment strategy employing inverse dynamics tasks and forward dynamics tasks, effectively strengthening multimodal correlations within a unified feature space. 3) Reinforced Supervised Fine-Tuning: While conducting instruction-based fine-tuning in the unified generation space, we construct a reinforce loss to align the spatial logic between textual actions and generated images, enabling the model to acquire spatio-awared multimodal planning capabilities.

Multimodal embedding models have gained significant attention for their ability to map data from different modalities, such as text and images, into a unified representation space. However, the limited labeled multimodal data often hinders embedding performance. Recent approaches have leveraged data synthesis to address this problem, yet the quality of synthetic data remains a critical bottleneck. In this work, we identify three criteria for high-quality synthetic multimodal data. First, broad scope ensures that the generated data covers diverse tasks and modalities, making it applicable to various downstream scenarios. Second, robust cross-modal alignment makes different modalities semantically consistent. Third, high fidelity ensures that the synthetic data maintains realistic details to enhance its reliability. Guided by these principles, we synthesize datasets that: (1) cover a wide range of tasks, modality combinations, and languages, (2) are generated via a deep thinking process within a single pass of a multimodal large language model, and (3) incorporate real-world images with accurate and relevant texts, ensuring fidelity through self-evaluation and refinement. Leveraging these high-quality synthetic and labeled datasets, we train a multimodal multilingual E5 model mmE5. Extensive experiments demonstrate that mmE5 achieves state-of-the-art performance on the MMEB Benchmark and superior multilingual performance on the XTD benchmark. Our codes, datasets and models are released in https://github.com/haon-chen/mmE5.

Although humans engaged in face-to-face conversation simultaneously communicate both verbally and non-verbally, methods for joint and unified synthesis of speech audio and co-speech 3D gesture motion from text are a new and emerging field. These technologies hold great promise for more human-like, efficient, expressive, and robust synthetic communication, but are currently held back by the lack of suitably large datasets, as existing methods are trained on parallel data from all constituent modalities. Inspired by student-teacher methods, we propose a straightforward solution to the data shortage, by simply synthesising additional training material. Specifically, we use unimodal synthesis models trained on large datasets to create multimodal (but synthetic) parallel training data, and then pre-train a joint synthesis model on that material. In addition, we propose a new synthesis architecture that adds better and more controllable prosody modelling to the state-of-the-art method in the field. Our results confirm that pre-training on large amounts of synthetic data improves the quality of both the speech and the motion synthesised by the multimodal model, with the proposed architecture yielding further benefits when pre-trained on the synthetic data. See https://shivammehta25.github.io/MAGI/ for example output.

Recent progress in multimodal reasoning has been driven largely by undisclosed datasets and proprietary data synthesis recipes, leaving open questions about how to systematically build large-scale, vision-centric reasoning datasets, particularly for tasks that go beyond visual math. In this work, we introduce a new reasoning data generation framework spanning diverse skills and levels of complexity with over 1M high-quality synthetic vision-centric questions. The dataset also includes preference data and instruction prompts supporting both offline and online RL. Our synthesis framework proceeds in two stages: (1) scale; and (2) complexity. Reasoning traces are then synthesized through a two-stage process that leverages VLMs and reasoning LLMs, producing CoT traces for VLMs that capture the richness and diverse cognitive behaviors found in frontier reasoning models. Remarkably, we show that finetuning Qwen2.5-VL-7B on our data outperforms all open-data baselines across all evaluated vision-centric benchmarks, and even surpasses strong closed-data models such as MiMo-VL-7B-RL on V* Bench, CV-Bench and MMStar-V. Perhaps most surprising, despite being entirely vision-centric, our data transfers positively to text-only reasoning (MMLU-Pro) and audio reasoning (MMAU), demonstrating its effectiveness. Similarly, despite not containing videos or embodied visual data, we observe notable gains when evaluating on a single-evidence embodied QA benchmark (NiEH). Finally, we use our data to analyze the entire VLM post-training pipeline. Our empirical analysis highlights that (i) SFT on high-quality data with non-linear reasoning traces is essential for effective online RL, (ii) staged offline RL matches online RL's performance while reducing compute demands, and (iii) careful SFT on high quality data can substantially improve out-of-domain, cross-modality transfer.

Training vision-language models (VLMs) typically requires large-scale, high-quality image-text pairs, but collecting or synthesizing such data is costly. In contrast, text data is abundant and inexpensive, prompting the question: can high-quality multimodal training data be synthesized purely from text? To tackle this, we propose a cross-integrated three-stage multimodal data synthesis framework, which generates two datasets: Unicorn-1.2M and Unicorn-471K-Instruction. In Stage 1: Diverse Caption Data Synthesis, we construct 1.2M semantically diverse high-quality captions by expanding sparse caption seeds using large language models (LLMs). In Stage 2: Instruction-Tuning Data Generation, we further process 471K captions into multi-turn instruction-tuning tasks to support complex reasoning. Finally, in Stage 3: Modality Representation Transfer, these textual captions representations are transformed into visual representations, resulting in diverse synthetic image representations. This three-stage process enables us to construct Unicorn-1.2M for pretraining and Unicorn-471K-Instruction for instruction-tuning, without relying on real images. By eliminating the dependency on real images while maintaining data quality and diversity, our framework offers a cost-effective and scalable solution for VLMs training. Code is available at https://github.com/Yu-xm/Unicorn.git.

Generating sound effects for videos often requires creating artistic sound effects that diverge significantly from real-life sources and flexible control in the sound design. To address this problem, we introduce MultiFoley, a model designed for video-guided sound generation that supports multimodal conditioning through text, audio, and video. Given a silent video and a text prompt, MultiFoley allows users to create clean sounds (e.g., skateboard wheels spinning without wind noise) or more whimsical sounds (e.g., making a lion's roar sound like a cat's meow). MultiFoley also allows users to choose reference audio from sound effects (SFX) libraries or partial videos for conditioning. A key novelty of our model lies in its joint training on both internet video datasets with low-quality audio and professional SFX recordings, enabling high-quality, full-bandwidth (48kHz) audio generation. Through automated evaluations and human studies, we demonstrate that MultiFoley successfully generates synchronized high-quality sounds across varied conditional inputs and outperforms existing methods. Please see our project page for video results: https://ificl.github.io/MultiFoley/

Large Language Models (LLMs) have shown remarkable success, and their multimodal expansions (MLLMs) further unlock capabilities spanning images, videos, and other modalities beyond text. However, despite this shift, prompt optimization approaches, designed to reduce the burden of manual prompt crafting while maximizing performance, remain confined to text, ultimately limiting the full potential of MLLMs. Motivated by this gap, we introduce the new problem of multimodal prompt optimization, which expands the prior definition of prompt optimization to the multimodal space defined by the pairs of textual and non-textual prompts. To tackle this problem, we then propose the Multimodal Prompt Optimizer (MPO), a unified framework that not only performs the joint optimization of multimodal prompts through alignment-preserving updates but also guides the selection process of candidate prompts by leveraging earlier evaluations as priors in a Bayesian-based selection strategy. Through extensive experiments across diverse modalities that go beyond text, such as images, videos, and even molecules, we demonstrate that MPO outperforms leading text-only optimization methods, establishing multimodal prompt optimization as a crucial step to realizing the potential of MLLMs.

This paper presents a multimodal framework that attempts to unify visual understanding and generation within a shared discrete semantic representation. At its core is the Text-Aligned Tokenizer (TA-Tok), which converts images into discrete tokens using a text-aligned codebook projected from a large language model's (LLM) vocabulary. By integrating vision and text into a unified space with an expanded vocabulary, our multimodal LLM, Tar, enables cross-modal input and output through a shared interface, without the need for modality-specific designs. Additionally, we propose scale-adaptive encoding and decoding to balance efficiency and visual detail, along with a generative de-tokenizer to produce high-fidelity visual outputs. To address diverse decoding needs, we utilize two complementary de-tokenizers: a fast autoregressive model and a diffusion-based model. To enhance modality fusion, we investigate advanced pre-training tasks, demonstrating improvements in both visual understanding and generation. Experiments across benchmarks show that Tar matches or surpasses existing multimodal LLM methods, achieving faster convergence and greater training efficiency. Code, models, and data are available at https://tar.csuhan.com

In recent years, advances in the large-scale pretraining of language and text-to-image models have revolutionized the field of machine learning. Yet, integrating these two modalities into a single, robust model capable of generating seamless multimodal outputs remains a significant challenge. To address this gap, we present the Joint Autoregressive Mixture (JAM) framework, a modular approach that systematically fuses existing text and image generation models. We also introduce a specialized, data-efficient instruction-tuning strategy, tailored for mixed-modal generation tasks. Our final instruct-tuned model demonstrates unparalleled performance in generating high-quality multimodal outputs and represents the first model explicitly designed for this purpose.

Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex reasoning tasks through long Chain-of-Thought (CoT) reasoning. Extending these successes to multimodal reasoning remains challenging due to the increased complexity of integrating diverse input modalities and the scarcity of high-quality long CoT training data. Existing multimodal datasets and CoT synthesis methods still suffer from limited reasoning depth, modality conversion errors, and rigid generation pipelines, hindering model performance and stability. To this end, in this paper, we propose SynSelect, a novel three-stage Synthesis-Selection framework for generating high-quality long CoT data tailored to multimodal reasoning tasks. Specifically, SynSelect first leverages multiple heterogeneous multimodal LRMs to produce diverse candidate CoTs, and then applies both instance and batch level selection to filter high-quality CoTs that can effectively enhance the model's reasoning capabilities. Extensive experiments on multiple multimodal benchmarks demonstrate that models supervised fine-tuned on SynSelect-generated data significantly outperform baselines and achieve further improvements after reinforcement learning post-training. Our results validate SynSelect as an effective approach for advancing multimodal LRMs reasoning capabilities.

Multimodal Large Language Models (MLLMs) are making significant progress in multimodal reasoning. Early approaches focus on pure text-based reasoning. More recent studies have incorporated multimodal information into the reasoning steps; however, they often follow a single task-specific reasoning pattern, which limits their generalizability across various multimodal tasks. In fact, there are numerous multimodal tasks requiring diverse reasoning skills, such as zooming in on a specific region or marking an object within an image. To address this, we propose unified generative multimodal reasoning, which unifies diverse multimodal reasoning skills by generating intermediate images during the reasoning process. We instantiate this paradigm with Omni-R1, a two-stage SFT+RL framework featuring perception alignment loss and perception reward, thereby enabling functional image generation. Additionally, we introduce Omni-R1-Zero, which eliminates the need for multimodal annotations by bootstrapping step-wise visualizations from text-only reasoning data. Empirical results show that Omni-R1 achieves unified generative reasoning across a wide range of multimodal tasks, and Omni-R1-Zero can match or even surpass Omni-R1 on average, suggesting a promising direction for generative multimodal reasoning.

Audio-Visual Intelligence (AVI) has emerged as a central frontier in artificial intelligence, bridging auditory and visual modalities to enable machines that can perceive, generate, and interact in the multimodal real world. In the era of large foundation models, joint modeling of audio and vision has become increasingly crucial, i.e., not only for understanding but also for controllable generation and reasoning across dynamic, temporally grounded signals. Recent advances, such as Meta MovieGen and Google Veo-3, highlight the growing industrial and academic focus on unified audio-vision architectures that learn from massive multimodal data. However, despite rapid progress, the literature remains fragmented, spanning diverse tasks, inconsistent taxonomies, and heterogeneous evaluation practices that impede systematic comparison and knowledge integration. This survey provides the first comprehensive review of AVI through the lens of large foundation models. We establish a unified taxonomy covering the broad landscape of AVI tasks, ranging from understanding (e.g., speech recognition, sound localization) to generation (e.g., audio-driven video synthesis, video-to-audio) and interaction (e.g., dialogue, embodied, or agentic interfaces). We synthesize methodological foundations, including modality tokenization, cross-modal fusion, autoregressive and diffusion-based generation, large-scale pretraining, instruction alignment, and preference optimization. Furthermore, we curate representative datasets, benchmarks, and evaluation metrics, offering a structured comparison across task families and identifying open challenges in synchronization, spatial reasoning, controllability, and safety. By consolidating this rapidly expanding field into a coherent framework, this survey aims to serve as a foundational reference for future research on large-scale AVI.

We present Unified-IO 2, the first autoregressive multimodal model that is capable of understanding and generating image, text, audio, and action. To unify different modalities, we tokenize inputs and outputs -- images, text, audio, action, bounding boxes, etc., into a shared semantic space and then process them with a single encoder-decoder transformer model. Since training with such diverse modalities is challenging, we propose various architectural improvements to stabilize model training. We train our model from scratch on a large multimodal pre-training corpus from diverse sources with a multimodal mixture of denoisers objective. To learn an expansive set of skills, such as following multimodal instructions, we construct and finetune on an ensemble of 120 datasets with prompts and augmentations. With a single unified model, Unified-IO 2 achieves state-of-the-art performance on the GRIT benchmark and strong results in more than 35 benchmarks, including image generation and understanding, natural language understanding, video and audio understanding, and robotic manipulation. We release all our models to the research community.

Expanding multimodal representations to novel modalities is constrained by reliance on large-scale paired datasets (e.g., text-image, text-audio, text-3D, text-molecule), which are costly and often infeasible in domains requiring expert annotation such as medical imaging and molecular analysis. We introduce TextME, the first text-only modality expansion framework, to the best of our knowledge, projecting diverse modalities into LLM embedding space as a unified anchor. Our approach exploits the geometric structure of pretrained contrastive encoders to enable zero-shot cross-modal transfer using only text descriptions, without paired supervision. We empirically validate that such consistent modality gaps exist across image, video, audio, 3D, X-ray, and molecular domains, demonstrating that text-only training can preserve substantial performance of pretrained encoders. We further show that our framework enables emergent cross-modal retrieval between modality pairs not explicitly aligned during training (e.g., audio-to-image, 3D-to-image). These results establish text-only training as a practical alternative to paired supervision for modality expansion.

We explore multi-step reasoning in vision-language models (VLMs). The problem is challenging, as reasoning data consisting of multiple steps of visual and language processing are barely available. To overcome the challenge, we first introduce a least-to-most visual reasoning paradigm, which interleaves steps of decomposing a question into sub-questions and invoking external tools for resolving sub-questions. Based on the paradigm, we further propose a novel data synthesis approach that can automatically create questions and multi-step reasoning paths for an image in a bottom-up manner. Our approach divides the complex synthesis task into a few simple sub-tasks, and (almost entirely) relies on open-sourced models to accomplish the sub-tasks. Therefore, the entire synthesis process is reproducible and cost-efficient, and the synthesized data is quality guaranteed. With the approach, we construct 50k visual reasoning examples. Then, we develop a visual reasoner through supervised fine-tuning, which is capable of generally enhancing the reasoning abilities of a wide range of existing VLMs in a plug-and-play fashion. Extensive experiments indicate that the visual reasoner can consistently and significantly improve four VLMs on four VQA benchmarks. Our code and dataset are available at https://github.com/steven-ccq/VisualReasoner.

Multimodal learning holds promise for richer information extraction by capturing dependencies across data sources. Yet, current training methods often underperform due to modality competition, a phenomenon where modalities contend for training resources leaving some underoptimized. This raises a pivotal question: how can we address training imbalances, ensure adequate optimization across all modalities, and achieve consistent performance improvements as we transition from unimodal to multimodal data? This paper proposes the Multimodal Competition Regularizer (MCR), inspired by a mutual information (MI) decomposition designed to prevent the adverse effects of competition in multimodal training. Our key contributions are: 1) A game-theoretic framework that adaptively balances modality contributions by encouraging each to maximize its informative role in the final prediction 2) Refining lower and upper bounds for each MI term to enhance the extraction of both task-relevant unique and shared information across modalities. 3) Proposing latent space permutations for conditional MI estimation, significantly improving computational efficiency. MCR outperforms all previously suggested training strategies and simple baseline, clearly demonstrating that training modalities jointly leads to important performance gains on both synthetic and large real-world datasets. We release our code and models at https://github.com/kkontras/MCR.

Advancements in language foundation models have primarily fueled the recent surge in artificial intelligence. In contrast, generative learning of non-textual modalities, especially videos, significantly trails behind language modeling. This thesis chronicles our endeavor to build multi-task models for generating videos and other modalities under diverse conditions, as well as for understanding and compression applications. Given the high dimensionality of visual data, we pursue concise and accurate latent representations. Our video-native spatial-temporal tokenizers preserve high fidelity. We unveil a novel approach to mapping bidirectionally between visual observation and interpretable lexical terms. Furthermore, our scalable visual token representation proves beneficial across generation, compression, and understanding tasks. This achievement marks the first instances of language models surpassing diffusion models in visual synthesis and a video tokenizer outperforming industry-standard codecs. Within these multi-modal latent spaces, we study the design of multi-task generative models. Our masked multi-task transformer excels at the quality, efficiency, and flexibility of video generation. We enable a frozen language model, trained solely on text, to generate visual content. Finally, we build a scalable generative multi-modal transformer trained from scratch, enabling the generation of videos containing high-fidelity motion with the corresponding audio given diverse conditions. Throughout the course, we have shown the effectiveness of integrating multiple tasks, crafting high-fidelity latent representation, and generating multiple modalities. This work suggests intriguing potential for future exploration in generating non-textual data and enabling real-time, interactive experiences across various media forms.

The visual world offers a critical axis for advancing foundation models beyond language. Despite growing interest in this direction, the design space for native multimodal models remains opaque. We provide empirical clarity through controlled, from-scratch pretraining experiments, isolating the factors that govern multimodal pretraining without interference from language pretraining. We adopt the Transfusion framework, using next-token prediction for language and diffusion for vision, to train on diverse data including text, video, image-text pairs, and even action-conditioned video. Our experiments yield four key insights: (i) Representation Autoencoder (RAE) provides an optimal unified visual representation by excelling at both visual understanding and generation; (ii) visual and language data are complementary and yield synergy for downstream capabilities; (iii) unified multimodal pretraining leads naturally to world modeling, with capabilities emerging from general training; and (iv) Mixture-of-Experts (MoE) enables efficient and effective multimodal scaling while naturally inducing modality specialization. Through IsoFLOP analysis, we compute scaling laws for both modalities and uncover a scaling asymmetry: vision is significantly more data-hungry than language. We demonstrate that the MoE architecture harmonizes this scaling asymmetry by providing the high model capacity required by language while accommodating the data-intensive nature of vision, paving the way for truly unified multimodal models.

Multimodal learning typically relies on the assumption that all modalities are fully available during both the training and inference phases. However, in real-world scenarios, consistently acquiring complete multimodal data presents significant challenges due to various factors. This often leads to the issue of missing modalities, where data for certain modalities are absent, posing considerable obstacles not only for the availability of multimodal pretrained models but also for their fine-tuning and the preservation of robustness in downstream tasks. To address these challenges, we propose a novel framework integrating parameter-efficient fine-tuning of unimodal pretrained models with a self-supervised joint-embedding learning method. This framework enables the model to predict the embedding of a missing modality in the representation space during inference. Our method effectively predicts the missing embedding through prompt tuning, leveraging information from available modalities. We evaluate our approach on several multimodal benchmark datasets and demonstrate its effectiveness and robustness across various scenarios of missing modalities.

We introduce AnyGPT, an any-to-any multimodal language model that utilizes discrete representations for the unified processing of various modalities, including speech, text, images, and music. AnyGPT can be trained stably without any alterations to the current large language model (LLM) architecture or training paradigms. Instead, it relies exclusively on data-level preprocessing, facilitating the seamless integration of new modalities into LLMs, akin to the incorporation of new languages. We build a multimodal text-centric dataset for multimodal alignment pre-training. Utilizing generative models, we synthesize the first large-scale any-to-any multimodal instruction dataset. It consists of 108k samples of multi-turn conversations that intricately interweave various modalities, thus equipping the model to handle arbitrary combinations of multimodal inputs and outputs. Experimental results demonstrate that AnyGPT is capable of facilitating any-to-any multimodal conversation while achieving performance comparable to specialized models across all modalities, proving that discrete representations can effectively and conveniently unify multiple modalities within a language model. Demos are shown in https://junzhan2000.github.io/AnyGPT.github.io/

Recent developments in Multimodal Large Language Models (MLLMs) have shown rapid progress, moving towards the goal of creating versatile MLLMs that understand inputs from various modalities. However, existing methods typically rely on joint training with paired multimodal instruction data, which is resource-intensive and challenging to extend to new modalities. In this paper, we propose a new paradigm through the model composition of existing MLLMs to create a new model that retains the modal understanding capabilities of each original model. Our basic implementation, NaiveMC, demonstrates the effectiveness of this paradigm by reusing modality encoders and merging LLM parameters. Furthermore, we introduce DAMC to address parameter interference and mismatch issues during the merging process, thereby enhancing the model performance. To facilitate research in this area, we propose MCUB, a benchmark for assessing ability of MLLMs to understand inputs from diverse modalities. Experiments on this benchmark and four other multimodal understanding tasks show significant improvements over baselines, proving that model composition can create a versatile model capable of processing inputs from multiple modalities.

Recent text-to-image models produce high-quality results but still struggle with precise visual control, balancing multimodal inputs, and requiring extensive training for complex multimodal image generation. To address these limitations, we propose MENTOR, a novel autoregressive (AR) framework for efficient Multimodal-conditioned Tuning for Autoregressive multimodal image generation. MENTOR combines an AR image generator with a two-stage training paradigm, enabling fine-grained, token-level alignment between multimodal inputs and image outputs without relying on auxiliary adapters or cross-attention modules. The two-stage training consists of: (1) a multimodal alignment stage that establishes robust pixel- and semantic-level alignment, followed by (2) a multimodal instruction tuning stage that balances the integration of multimodal inputs and enhances generation controllability. Despite modest model size, suboptimal base components, and limited training resources, MENTOR achieves strong performance on the DreamBench++ benchmark, outperforming competitive baselines in concept preservation and prompt following. Additionally, our method delivers superior image reconstruction fidelity, broad task adaptability, and improved training efficiency compared to diffusion-based methods. Dataset, code, and models are available at: https://github.com/HaozheZhao/MENTOR

Multimodal generative models have shown impressive advances with the help of powerful diffusion models. Despite the progress, generating sound solely from text poses challenges in ensuring comprehensive scene depiction and temporal alignment. Meanwhile, video-to-sound generation limits the flexibility to prioritize sound synthesis for specific objects within the scene. To tackle these challenges, we propose a novel video-and-text-to-sound generation method, called ReWaS, where video serves as a conditional control for a text-to-audio generation model. Our method estimates the structural information of audio (namely, energy) from the video while receiving key content cues from a user prompt. We employ a well-performing text-to-sound model to consolidate the video control, which is much more efficient for training multimodal diffusion models with massive triplet-paired (audio-video-text) data. In addition, by separating the generative components of audio, it becomes a more flexible system that allows users to freely adjust the energy, surrounding environment, and primary sound source according to their preferences. Experimental results demonstrate that our method shows superiority in terms of quality, controllability, and training efficiency. Our demo is available at https://naver-ai.github.io/rewas

In this paper, we propose and investigate the use of neural audio codec language models for the automatic generation of sample-based musical instruments based on text or reference audio prompts. Our approach extends a generative audio framework to condition on pitch across an 88-key spectrum, velocity, and a combined text/audio embedding. We identify maintaining timbral consistency within the generated instruments as a major challenge. To tackle this issue, we introduce three distinct conditioning schemes. We analyze our methods through objective metrics and human listening tests, demonstrating that our approach can produce compelling musical instruments. Specifically, we introduce a new objective metric to evaluate the timbral consistency of the generated instruments and adapt the average Contrastive Language-Audio Pretraining (CLAP) score for the text-to-instrument case, noting that its naive application is unsuitable for assessing this task. Our findings reveal a complex interplay between timbral consistency, the quality of generated samples, and their correspondence to the input prompt.

Multimodal music generation aims to produce music from diverse input modalities, including text, videos, and images. Existing methods use a common embedding space for multimodal fusion. Despite their effectiveness in other modalities, their application in multimodal music generation faces challenges of data scarcity, weak cross-modal alignment, and limited controllability. This paper addresses these issues by using explicit bridges of text and music for multimodal alignment. We introduce a novel method named Visuals Music Bridge (VMB). Specifically, a Multimodal Music Description Model converts visual inputs into detailed textual descriptions to provide the text bridge; a Dual-track Music Retrieval module that combines broad and targeted retrieval strategies to provide the music bridge and enable user control. Finally, we design an Explicitly Conditioned Music Generation framework to generate music based on the two bridges. We conduct experiments on video-to-music, image-to-music, text-to-music, and controllable music generation tasks, along with experiments on controllability. The results demonstrate that VMB significantly enhances music quality, modality, and customization alignment compared to previous methods. VMB sets a new standard for interpretable and expressive multimodal music generation with applications in various multimedia fields. Demos and code are available at https://github.com/wbs2788/VMB.

In this work, we build a simple but strong baseline for sounding video generation. Given base diffusion models for audio and video, we integrate them with additional modules into a single model and train it to make the model jointly generate audio and video. To enhance alignment between audio-video pairs, we introduce two novel mechanisms in our model. The first one is timestep adjustment, which provides different timestep information to each base model. It is designed to align how samples are generated along with timesteps across modalities. The second one is a new design of the additional modules, termed Cross-Modal Conditioning as Positional Encoding (CMC-PE). In CMC-PE, cross-modal information is embedded as if it represents temporal position information, and the embeddings are fed into the model like positional encoding. Compared with the popular cross-attention mechanism, CMC-PE provides a better inductive bias for temporal alignment in the generated data. Experimental results validate the effectiveness of the two newly introduced mechanisms and also demonstrate that our method outperforms existing methods.

Multimodal learning seeks to combine data from multiple input sources to enhance the performance of different downstream tasks. In real-world scenarios, performance can degrade substantially if some input modalities are missing. Existing methods that can handle missing modalities involve custom training or adaptation steps for each input modality combination. These approaches are either tied to specific modalities or become computationally expensive as the number of input modalities increases. In this paper, we propose Masked Modality Projection (MMP), a method designed to train a single model that is robust to any missing modality scenario. We achieve this by randomly masking a subset of modalities during training and learning to project available input modalities to estimate the tokens for the masked modalities. This approach enables the model to effectively learn to leverage the information from the available modalities to compensate for the missing ones, enhancing missing modality robustness. We conduct a series of experiments with various baseline models and datasets to assess the effectiveness of this strategy. Experiments demonstrate that our approach improves robustness to different missing modality scenarios, outperforming existing methods designed for missing modalities or specific modality combinations.

Recent large-scale text-driven synthesis models have attracted much attention thanks to their remarkable capabilities of generating highly diverse images that follow given text prompts. Such text-based synthesis methods are particularly appealing to humans who are used to verbally describe their intent. Therefore, it is only natural to extend the text-driven image synthesis to text-driven image editing. Editing is challenging for these generative models, since an innate property of an editing technique is to preserve most of the original image, while in the text-based models, even a small modification of the text prompt often leads to a completely different outcome. State-of-the-art methods mitigate this by requiring the users to provide a spatial mask to localize the edit, hence, ignoring the original structure and content within the masked region. In this paper, we pursue an intuitive prompt-to-prompt editing framework, where the edits are controlled by text only. To this end, we analyze a text-conditioned model in depth and observe that the cross-attention layers are the key to controlling the relation between the spatial layout of the image to each word in the prompt. With this observation, we present several applications which monitor the image synthesis by editing the textual prompt only. This includes localized editing by replacing a word, global editing by adding a specification, and even delicately controlling the extent to which a word is reflected in the image. We present our results over diverse images and prompts, demonstrating high-quality synthesis and fidelity to the edited prompts.

We use (multi)modal deep neural networks (DNNs) to probe for sites of multimodal integration in the human brain by predicting stereoencephalography (SEEG) recordings taken while human subjects watched movies. We operationalize sites of multimodal integration as regions where a multimodal vision-language model predicts recordings better than unimodal language, unimodal vision, or linearly-integrated language-vision models. Our target DNN models span different architectures (e.g., convolutional networks and transformers) and multimodal training techniques (e.g., cross-attention and contrastive learning). As a key enabling step, we first demonstrate that trained vision and language models systematically outperform their randomly initialized counterparts in their ability to predict SEEG signals. We then compare unimodal and multimodal models against one another. Because our target DNN models often have different architectures, number of parameters, and training sets (possibly obscuring those differences attributable to integration), we carry out a controlled comparison of two models (SLIP and SimCLR), which keep all of these attributes the same aside from input modality. Using this approach, we identify a sizable number of neural sites (on average 141 out of 1090 total sites or 12.94%) and brain regions where multimodal integration seems to occur. Additionally, we find that among the variants of multimodal training techniques we assess, CLIP-style training is the best suited for downstream prediction of the neural activity in these sites.

The goal of this work is to enhance balanced multimodal understanding in audio-visual large language models (AV-LLMs) by addressing modality bias without requiring additional training. In current AV-LLMs, audio and video features are typically processed jointly in the decoder. While this strategy facilitates unified multimodal understanding, it may introduce modality bias, where the model tends to over-rely on one modality due to imbalanced training signals. To mitigate this, we propose Fork-Merge Decoding (FMD), a simple yet effective inference-time strategy that requires no additional training or architectural modifications. FMD first performs modality-specific reasoning by processing audio-only and video-only inputs through the early decoder layers (a fork phase), and then merges the resulting hidden states for joint reasoning in the remaining layers (a merge phase). This approach promotes balanced modality contributions and leverages complementary information across modalities. We evaluate our method on two representative AV-LLMs, VideoLLaMA2 and video-SALMONN, using three benchmark datasets. Experimental results demonstrate consistent performance improvements on tasks focused on audio, video, and combined audio-visual reasoning, demonstrating the effectiveness of inference-time interventions for robust multimodal understanding.

Traditional multimodal learners find unified representations for tasks like visual question answering, but rely heavily on paired datasets. However, an overlooked yet potentially powerful question is: can one leverage auxiliary unpaired multimodal data to directly enhance representation learning in a target modality? We introduce UML: Unpaired Multimodal Learner, a modality-agnostic training paradigm in which a single model alternately processes inputs from different modalities while sharing parameters across them. This design exploits the assumption that different modalities are projections of a shared underlying reality, allowing the model to benefit from cross-modal structure without requiring explicit pairs. Theoretically, under linear data-generating assumptions, we show that unpaired auxiliary data can yield representations strictly more informative about the data-generating process than unimodal training. Empirically, we show that using unpaired data from auxiliary modalities -- such as text, audio, or images -- consistently improves downstream performance across diverse unimodal targets such as image and audio. Our project page: https://unpaired-multimodal.github.io/

Audio-video joint generation has progressed rapidly, yet substantial challenges still remain. Non-commercial approaches still suffer audio-visual asynchrony, poor lip-speech alignment, and unimodal degradation, which can be stemmed from weak audio-visual correspondence modeling, limited generalization, and scarce high-quality dense-caption data. To address these issues, we introduce Klear and delve into three axes--model architecture, training strategy, and data curation. Architecturally, we adopt a single-tower design with unified DiT blocks and an Omni-Full Attention mechanism, achieving tight audio-visual alignment and strong scalability. Training-wise, we adopt a progressive multitask regime--random modality masking to joint optimization across tasks, and a multistage curriculum, yielding robust representations, strengthening A-V aligned world knowledge, and preventing unimodal collapse. For datasets, we present the first large-scale audio-video dataset with dense captions, and introduce a novel automated data-construction pipeline which annotates and filters millions of diverse, high-quality, strictly aligned audio-video-caption triplets. Building on this, Klear scales to large datasets, delivering high-fidelity, semantically and temporally aligned, instruction-following generation in both joint and unimodal settings while generalizing robustly to out-of-distribution scenarios. Across tasks, it substantially outperforms prior methods by a large margin and achieves performance comparable to Veo 3, offering a unified, scalable path toward next-generation audio-video synthesis.

Scaling multimodal alignment between video and audio is challenging, particularly due to limited data and the mismatch between text descriptions and frame-level video information. In this work, we tackle the scaling challenge in multimodal-to-audio generation, examining whether models trained on short instances can generalize to longer ones during testing. To tackle this challenge, we present multimodal hierarchical networks so-called MMHNet, an enhanced extension of state-of-the-art video-to-audio models. Our approach integrates a hierarchical method and non-causal Mamba to support long-form audio generation. Our proposed method significantly improves long audio generation up to more than 5 minutes. We also prove that training short and testing long is possible in the video-to-audio generation tasks without training on the longer durations. We show in our experiments that our proposed method could achieve remarkable results on long-video to audio benchmarks, beating prior works in video-to-audio tasks. Moreover, we showcase our model capability in generating more than 5 minutes, while prior video-to-audio methods fall short in generating with long durations.

Joint audio-video generation aims to synthesize synchronized multisensory content, yet current unified models struggle with fine-grained acoustic control, particularly for identity-preserving speech. Existing approaches either suffer from temporal misalignment due to cascaded generation or lack the capability to perform zero-shot voice cloning within a joint synthesis framework. In this work, we present MM-Sonate, a multimodal flow-matching framework that unifies controllable audio-video joint generation with zero-shot voice cloning capabilities. Unlike prior works that rely on coarse semantic descriptions, MM-Sonate utilizes a unified instruction-phoneme input to enforce strict linguistic and temporal alignment. To enable zero-shot voice cloning, we introduce a timbre injection mechanism that effectively decouples speaker identity from linguistic content. Furthermore, addressing the limitations of standard classifier-free guidance in multimodal settings, we propose a noise-based negative conditioning strategy that utilizes natural noise priors to significantly enhance acoustic fidelity. Empirical evaluations demonstrate that MM-Sonate establishes new state-of-the-art performance in joint generation benchmarks, significantly outperforming baselines in lip synchronization and speech intelligibility, while achieving voice cloning fidelity comparable to specialized Text-to-Speech systems.

We present Vx2Text, a framework for text generation from multimodal inputs consisting of video plus text, speech, or audio. In order to leverage transformer networks, which have been shown to be effective at modeling language, each modality is first converted into a set of language embeddings by a learnable tokenizer. This allows our approach to perform multimodal fusion in the language space, thus eliminating the need for ad-hoc cross-modal fusion modules. To address the non-differentiability of tokenization on continuous inputs (e.g., video or audio), we utilize a relaxation scheme that enables end-to-end training. Furthermore, unlike prior encoder-only models, our network includes an autoregressive decoder to generate open-ended text from the multimodal embeddings fused by the language encoder. This renders our approach fully generative and makes it directly applicable to different "video+x to text" problems without the need to design specialized network heads for each task. The proposed framework is not only conceptually simple but also remarkably effective: experiments demonstrate that our approach based on a single architecture outperforms the state-of-the-art on three video-based text-generation tasks -- captioning, question answering and audio-visual scene-aware dialog.

Large multimodal models have achieved remarkable progress in both understanding and generation. Recent efforts pursue unified multimodal models that integrate heterogeneous components to support both capabilities within a single framework. However, such unification introduces inference inefficiencies, e.g., specific tasks or samples may not require the full knowledge or capacity of the unified model. Yet, a systematic understanding of how these inefficiencies manifest across different components remains limited. In this work, we first conduct a systematic analysis of unified multimodal model components using training-free pruning as a probing methodology, considering both depth pruning and width reduction. Our study reveals that the understanding component exhibits notable compressibility in both understanding and generation tasks, which is more pronounced in the latter. In contrast, the generation components are highly sensitive to compression, with performance deteriorating sharply even under moderate compression ratios. To address this limitation, we propose the Mixture-of-Experts (MoE) Adaptation, inspired by the dynamic activation patterns observed across different samples. This approach partitions the generation module into multiple experts and enables sparse activation to restore generation quality. We validate the effectiveness of sparse activation through expert-frozen tuning and further demonstrate that a fully trainable adaptation delivers additional gains. As a result, the adapted BAGEL model achieves performance comparable to the full model while activating only about half of its parameters. The code is released at https://github.com/Shwai-He/SparseUnifiedModel{this link}.

The success of multi-modal large language models (MLLMs) has been largely attributed to the large-scale training data. However, the training data of many MLLMs is unavailable due to privacy concerns. The expensive and labor-intensive process of collecting multi-modal data further exacerbates the problem. Is it possible to synthesize multi-modal training data automatically without compromising diversity and quality? In this paper, we propose a new method, Oasis, to synthesize high-quality multi-modal data with only images. Oasis breaks through traditional methods by prompting only images to the MLLMs, thus extending the data diversity by a large margin. Our method features a delicate quality control method which ensures the data quality. We collected over 500k data and conducted incremental experiments on LLaVA-NeXT. Extensive experiments demonstrate that our method can significantly improve the performance of MLLMs. The image-based synthesis also allows us to focus on the specific-domain ability of MLLMs. Code and data will be publicly available.

Despite the success of multimodal contrastive learning in aligning visual and linguistic representations, a persistent geometric anomaly, the Modality Gap, remains: embeddings of distinct modalities expressing identical semantics occupy systematically offset regions. Prior approaches to bridge this gap are largely limited by oversimplified isotropic assumptions, hindering their application in large-scale scenarios. In this paper, we address these limitations by precisely characterizing the geometric shape of the modality gap and leveraging it for efficient model scaling. First, we propose the Fixed-frame Modality Gap Theory, which decomposes the modality gap within a frozen reference frame into stable biases and anisotropic residuals. Guided by this precise modeling, we introduce ReAlign, a training-free modality alignment strategy. Utilizing statistics from massive unpaired data, ReAlign aligns text representation into the image representation distribution via a three-step process comprising Anchor, Trace, and Centroid Alignment, thereby explicitly rectifying geometric misalignment. Building on ReAlign, we propose ReVision, a scalable training paradigm for Multimodal Large Language Models (MLLMs). ReVision integrates ReAlign into the pretraining stage, enabling the model to learn the distribution of visual representations from unpaired text before visual instruction tuning, without the need for large-scale, high-quality image-text pairs. Our framework demonstrates that statistically aligned unpaired data can effectively substitute for expensive image-text pairs, offering a robust path for the efficient scaling of MLLMs.

Conditional motion generation has been extensively studied in computer vision, yet two critical challenges remain. First, while masked autoregressive methods have recently outperformed diffusion-based approaches, existing masking models lack a mechanism to prioritize dynamic frames and body parts based on given conditions. Second, existing methods for different conditioning modalities often fail to integrate multiple modalities effectively, limiting control and coherence in generated motion. To address these challenges, we propose Motion Anything, a multimodal motion generation framework that introduces an Attention-based Mask Modeling approach, enabling fine-grained spatial and temporal control over key frames and actions. Our model adaptively encodes multimodal conditions, including text and music, improving controllability. Additionally, we introduce Text-Music-Dance (TMD), a new motion dataset consisting of 2,153 pairs of text, music, and dance, making it twice the size of AIST++, thereby filling a critical gap in the community. Extensive experiments demonstrate that Motion Anything surpasses state-of-the-art methods across multiple benchmarks, achieving a 15% improvement in FID on HumanML3D and showing consistent performance gains on AIST++ and TMD. See our project website https://steve-zeyu-zhang.github.io/MotionAnything

In this paper, we address the task of multimodal-to-speech generation, which aims to synthesize high-quality speech from multiple input modalities: text, video, and reference audio. This task has gained increasing attention due to its wide range of applications, such as film production, dubbing, and virtual avatars. Despite recent progress, existing methods still suffer from limitations in speech intelligibility, audio-video synchronization, speech naturalness, and voice similarity to the reference speaker. To address these challenges, we propose AlignDiT, a multimodal Aligned Diffusion Transformer that generates accurate, synchronized, and natural-sounding speech from aligned multimodal inputs. Built upon the in-context learning capability of the DiT architecture, AlignDiT explores three effective strategies to align multimodal representations. Furthermore, we introduce a novel multimodal classifier-free guidance mechanism that allows the model to adaptively balance information from each modality during speech synthesis. Extensive experiments demonstrate that AlignDiT significantly outperforms existing methods across multiple benchmarks in terms of quality, synchronization, and speaker similarity. Moreover, AlignDiT exhibits strong generalization capability across various multimodal tasks, such as video-to-speech synthesis and visual forced alignment, consistently achieving state-of-the-art performance. The demo page is available at https://mm.kaist.ac.kr/projects/AlignDiT.

In this work, we study the problem of generating novel images from complex multimodal prompt sequences. While existing methods achieve promising results for text-to-image generation, they often struggle to capture fine-grained details from lengthy prompts and maintain contextual coherence within prompt sequences. Moreover, they often result in misaligned image generation for prompt sequences featuring multiple objects. To address this, we propose a Multi-modal Generation via Cross-Modal In-Context Learning (MGCC) method that generates novel images from complex multimodal prompt sequences by leveraging the combined capabilities of large language models (LLMs) and diffusion models. Our MGCC comprises a novel Cross-Modal Refinement module to explicitly learn cross-modal dependencies between the text and image in the LLM embedding space, and a contextual object grounding module to generate object bounding boxes specifically targeting scenes with multiple objects. Our MGCC demonstrates a diverse range of multimodal capabilities, like novel image generation, the facilitation of multimodal dialogue, and generation of texts. Experimental evaluations on two benchmark datasets, demonstrate the effectiveness of our method. On Visual Story Generation (VIST) dataset with multimodal inputs, our MGCC achieves a CLIP Similarity score of 0.652 compared to SOTA GILL 0.641. Similarly, on Visual Dialogue Context (VisDial) having lengthy dialogue sequences, our MGCC achieves an impressive CLIP score of 0.660, largely outperforming existing SOTA method scoring 0.645. Code: https://github.com/VIROBO-15/MGCC

The rapid proliferation of Vision-Language Models (VLMs) is often framed as enabling unified multimodal knowledge discovery but rests on an under-examined assumption: that current VLMs faithfully synthesise multimodal data. We argue they often do not, and this gap reflects a trustworthiness problem in the dominant Vision Encoder-Projector-LLM paradigm. Rather than extracting grounded knowledge from visual inputs, state-of-the-art models frequently exhibit functional blindness, i.e., exploiting strong language priors to bypass severe visual representation bottlenecks. In this work, we challenge the conventional methodology of multimodal evaluation, which relies on data ablation or new dataset creation and therefore conflates dataset biases with architectural incapacity. We propose an information-theoretic departure: the Modality Translation Protocol, designed to quantify what we call the Expense of Seeing. By translating semantic payloads rather than ablating them, we formulate three novel metrics -- the Toll (ToS), Curse (CoS), and Fallacy (FoS) of Seeing -- culminating in the Semantic Sufficiency Criterion (SSC). Furthermore, we hypothesise a Divergence Law of Multimodal Scaling: as the underlying language engines scale to unprecedented reasoning capabilities, the penalty of the visual knowledge bottleneck may increase rather than diminish. We argue the community should move beyond "multimodal gain" as a primary evaluation target. By elevating the SSC from a passive diagnostic constraint to an active architectural blueprint, we provide a foundation for guiding the next generation of AI systems toward genuine multimodal reasoning.

There is growing interest in integrating high-fidelity visual synthesis capabilities into large language models (LLMs) without compromising their strong reasoning capabilities. Existing methods that directly train LLMs or bridge LLMs and diffusion models usually suffer from costly training since the backbone LLMs have not seen image representations during pretraining. We present Bifrost-1, a unified framework that bridges pretrained multimodal LLMs (MLLMs) and diffusion models using patch-level CLIP image embeddings as latent variables, which are natively aligned with the MLLM's CLIP visual encoder. These patch-level image embeddings are integrated into the diffusion model with a lightweight adaptation of its ControlNet. To retain the original multimodal reasoning capabilities of MLLMs, we equip the MLLM with a visual generation branch initialized from the original MLLM parameters when predicting the patch-level image embeddings. By seamlessly integrating pretrained MLLMs and diffusion models with patch-level CLIP latents, our framework enables high-fidelity controllable image generation with significant training efficiency. Our experiments demonstrate that Bifrost-1 achieves comparable or better performance than previous methods in terms of visual fidelity and multimodal understanding, with substantially lower compute during training. We also provide comprehensive ablation studies showing the effectiveness of our design choices.

Existing text-to-video diffusion models rely solely on text-only encoders for their pretraining. This limitation stems from the absence of large-scale multimodal prompt video datasets, resulting in a lack of visual grounding and restricting their versatility and application in multimodal integration. To address this, we construct a large-scale multimodal prompt dataset by employing retrieval methods to pair in-context examples with the given text prompts and then utilize a two-stage training strategy to enable diverse video generation tasks within the same model. In the first stage, we propose a multimodal conditional video generation framework for pretraining on these augmented datasets, establishing a foundational model for grounded video generation. Secondly, we finetune the model from the first stage on three video generation tasks, incorporating multi-modal instructions. This process further refines the model's ability to handle diverse inputs and tasks, ensuring seamless integration of multi-modal information. After this two-stage train-ing process, VIMI demonstrates multimodal understanding capabilities, producing contextually rich and personalized videos grounded in the provided inputs, as shown in Figure 1. Compared to previous visual grounded video generation methods, VIMI can synthesize consistent and temporally coherent videos with large motion while retaining the semantic control. Lastly, VIMI also achieves state-of-the-art text-to-video generation results on UCF101 benchmark.

Text-guided diffusion models have advanced image editing by enabling intuitive control through language. However, despite their strong capabilities, we surprisingly find that SOTA methods struggle with simple, everyday transformations such as rain or blur. We attribute this limitation to weak and inconsistent textual supervision during training, which leads to poor alignment between language and vision. Existing solutions often rely on extra finetuning or stronger text conditioning, but suffer from high data and computational requirements. We argue that diffusion-based editing capabilities aren't lost but merely hidden from text. The door to cost-efficient visual editing remains open, and the key lies in a vision-centric paradigm that perceives and reasons about visual change as humans do, beyond words. Inspired by this, we introduce Visual Diffusion Conditioning (VDC), a training-free framework that learns conditioning signals directly from visual examples for precise, language-free image editing. Given a paired example -one image with and one without the target effect- VDC derives a visual condition that captures the transformation and steers generation through a novel condition-steering mechanism. An accompanying inversion-correction step mitigates reconstruction errors during DDIM inversion, preserving fine detail and realism. Across diverse tasks, VDC outperforms both training-free and fully fine-tuned text-based editing methods. The code and models are open-sourced at https://omaralezaby.github.io/vdc/

Text-conditional diffusion models are able to generate high-fidelity images with diverse contents. However, linguistic representations frequently exhibit ambiguous descriptions of the envisioned objective imagery, requiring the incorporation of additional control signals to bolster the efficacy of text-guided diffusion models. In this work, we propose Cocktail, a pipeline to mix various modalities into one embedding, amalgamated with a generalized ControlNet (gControlNet), a controllable normalisation (ControlNorm), and a spatial guidance sampling method, to actualize multi-modal and spatially-refined control for text-conditional diffusion models. Specifically, we introduce a hyper-network gControlNet, dedicated to the alignment and infusion of the control signals from disparate modalities into the pre-trained diffusion model. gControlNet is capable of accepting flexible modality signals, encompassing the simultaneous reception of any combination of modality signals, or the supplementary fusion of multiple modality signals. The control signals are then fused and injected into the backbone model according to our proposed ControlNorm. Furthermore, our advanced spatial guidance sampling methodology proficiently incorporates the control signal into the designated region, thereby circumventing the manifestation of undesired objects within the generated image. We demonstrate the results of our method in controlling various modalities, proving high-quality synthesis and fidelity to multiple external signals.

How does audio describe the world around us? In this work, we propose a method for generating images of visual scenes from diverse in-the-wild sounds. This cross-modal generation task is challenging due to the significant information gap between auditory and visual signals. We address this challenge by designing a model that aligns audio-visual modalities by enriching audio features with visual information and translating them into the visual latent space. These features are then fed into the pre-trained image generator to produce images. To enhance image quality, we use sound source localization to select audio-visual pairs with strong cross-modal correlations. Our method achieves substantially better results on the VEGAS and VGGSound datasets compared to previous work and demonstrates control over the generation process through simple manipulations to the input waveform or latent space. Furthermore, we analyze the geometric properties of the learned embedding space and demonstrate that our learning approach effectively aligns audio-visual signals for cross-modal generation. Based on this analysis, we show that our method is agnostic to specific design choices, showing its generalizability by integrating various model architectures and different types of audio-visual data.

Reinforcement Learning (RL) has shown promise in improving the reasoning abilities of Large Language Models (LLMs). However, the specific challenges of adapting RL to multimodal data and formats remain relatively unexplored. In this work, we present Observe-R1, a novel framework aimed at enhancing the reasoning capabilities of multimodal large language models (MLLMs). We draw inspirations from human learning progression--from simple to complex and easy to difficult, and propose a gradual learning paradigm for MLLMs. To this end, we construct the NeuraLadder dataset, which is organized and sampled according to the difficulty and complexity of data samples for RL training. To tackle multimodal tasks, we introduce a multimodal format constraint that encourages careful observation of images, resulting in enhanced visual abilities and clearer and more structured responses. Additionally, we implement a bonus reward system that favors concise, correct answers within a length constraint, alongside a dynamic weighting mechanism that prioritizes uncertain and medium-difficulty problems, ensuring that more informative samples have a greater impact on training. Our experiments with the Qwen2.5-VL-3B and Qwen2.5-VL-7B models on 20k samples from the NeuraLadder dataset show that Observe-R1 outperforms a series of larger reasoning models on both reasoning and general benchmarks, achieving superior clarity and conciseness in reasoning chains. Ablation studies validate the effectiveness of our strategies, highlighting the robustness and generalization of our approach. The dataset and code will be released at https://github.com/zrguo/Observe-R1.

Existing text-to-image diffusion models primarily generate images from text prompts. However, the inherent conciseness of textual descriptions poses challenges in faithfully synthesizing images with intricate details, such as specific entities or scenes. This paper presents UNIMO-G, a simple multimodal conditional diffusion framework that operates on multimodal prompts with interleaved textual and visual inputs, which demonstrates a unified ability for both text-driven and subject-driven image generation. UNIMO-G comprises two core components: a Multimodal Large Language Model (MLLM) for encoding multimodal prompts, and a conditional denoising diffusion network for generating images based on the encoded multimodal input. We leverage a two-stage training strategy to effectively train the framework: firstly pre-training on large-scale text-image pairs to develop conditional image generation capabilities, and then instruction tuning with multimodal prompts to achieve unified image generation proficiency. A well-designed data processing pipeline involving language grounding and image segmentation is employed to construct multi-modal prompts. UNIMO-G excels in both text-to-image generation and zero-shot subject-driven synthesis, and is notably effective in generating high-fidelity images from complex multimodal prompts involving multiple image entities.

We present VideoPoet, a language model capable of synthesizing high-quality video, with matching audio, from a large variety of conditioning signals. VideoPoet employs a decoder-only transformer architecture that processes multimodal inputs -- including images, videos, text, and audio. The training protocol follows that of Large Language Models (LLMs), consisting of two stages: pretraining and task-specific adaptation. During pretraining, VideoPoet incorporates a mixture of multimodal generative objectives within an autoregressive Transformer framework. The pretrained LLM serves as a foundation that can be adapted for a range of video generation tasks. We present empirical results demonstrating the model's state-of-the-art capabilities in zero-shot video generation, specifically highlighting VideoPoet's ability to generate high-fidelity motions. Project page: http://sites.research.google/videopoet/

Multimodal learning combines multiple data modalities, broadening the types and complexity of data our models can utilize: for example, from plain text to image-caption pairs. Most multimodal learning algorithms focus on modeling simple one-to-one pairs of data from two modalities, such as image-caption pairs, or audio-text pairs. However, in most real-world settings, entities of different modalities interact with each other in more complex and multifaceted ways, going beyond one-to-one mappings. We propose to represent these complex relationships as graphs, allowing us to capture data with any number of modalities, and with complex relationships between modalities that can flexibly vary from one sample to another. Toward this goal, we propose Multimodal Graph Learning (MMGL), a general and systematic framework for capturing information from multiple multimodal neighbors with relational structures among them. In particular, we focus on MMGL for generative tasks, building upon pretrained Language Models (LMs), aiming to augment their text generation with multimodal neighbor contexts. We study three research questions raised by MMGL: (1) how can we infuse multiple neighbor information into the pretrained LMs, while avoiding scalability issues? (2) how can we infuse the graph structure information among multimodal neighbors into the LMs? and (3) how can we finetune the pretrained LMs to learn from the neighbor context in a parameter-efficient manner? We conduct extensive experiments to answer these three questions on MMGL and analyze the empirical results to pave the way for future MMGL research.

Integrating multiple generative foundation models, especially those trained on different modalities, into something greater than the sum of its parts poses significant challenges. Two key hurdles are the availability of aligned data (concepts that contain similar meaning but is expressed differently in different modalities), and effectively leveraging unimodal representations in cross-domain generative tasks, without compromising their original unimodal capabilities. We propose Zipper, a multi-tower decoder architecture that addresses these concerns by using cross-attention to flexibly compose multimodal generative models from independently pre-trained unimodal decoders. In our experiments fusing speech and text modalities, we show the proposed architecture performs very competitively in scenarios with limited aligned text-speech data. We also showcase the flexibility of our model to selectively maintain unimodal (e.g., text-to-text generation) generation performance by freezing the corresponding modal tower (e.g. text). In cross-modal tasks such as automatic speech recognition (ASR) where the output modality is text, we show that freezing the text backbone results in negligible performance degradation. In cross-modal tasks such as text-to-speech generation (TTS) where the output modality is speech, we show that using a pre-trained speech backbone results in superior performance to the baseline.

Visual program synthesis is a promising approach to exploit the reasoning abilities of large language models for compositional computer vision tasks. Previous work has used few-shot prompting with frozen LLMs to synthesize visual programs. Training an LLM to write better visual programs is an attractive prospect, but it is unclear how to accomplish this. No dataset of visual programs for training exists, and acquisition of a visual program dataset cannot be easily crowdsourced due to the need for expert annotators. To get around the lack of direct supervision, we explore improving the program synthesis abilities of an LLM using feedback from interactive experience. We propose a method where we exploit existing annotations for a vision-language task to improvise a coarse reward signal for that task, treat the LLM as a policy, and apply reinforced self-training to improve the visual program synthesis ability of the LLM for that task. We describe a series of experiments on object detection, compositional visual question answering, and image-text retrieval, and show that in each case, the self-trained LLM outperforms or performs on par with few-shot frozen LLMs that are an order of magnitude larger. Website: https://zaidkhan.me/ViReP

Multimodal models often experience a significant performance drop when one or more modalities are missing during inference. To address this challenge, we propose a simple yet effective approach that enhances robustness to missing modalities while maintaining strong performance when all modalities are available. Our method introduces cross-modal proxy tokens (CMPTs), which approximate the class token of a missing modality by attending only to the tokens of the available modality without requiring explicit modality generation or auxiliary networks. To efficiently learn these approximations with minimal computational overhead, we employ low-rank adapters in frozen unimodal encoders and jointly optimize an alignment loss with a task-specific loss. Extensive experiments on five multimodal datasets show that our method outperforms state-of-the-art baselines across various missing rates while achieving competitive results in complete-modality settings. Overall, our method offers a flexible and efficient solution for robust multimodal learning. The code and pretrained models will be released on GitHub.

In this paper, we propose a Vision-Audio-Language Omni-peRception pretraining model (VALOR) for multi-modal understanding and generation. Different from widely-studied vision-language pretraining models, VALOR jointly models relationships of vision, audio and language in an end-to-end manner. It contains three separate encoders for single modality representations, and a decoder for multimodal conditional text generation. We design two pretext tasks to pretrain VALOR model, including Multimodal Grouping Alignment (MGA) and Multimodal Grouping Captioning (MGC). MGA projects vision, language and audio to the same common space, building vision-language, audio-language and audiovisual-language alignment simultaneously. MGC learns how to generate text tokens in conditions of vision, audio or their both. To promote vision-audio-language pretraining research, we construct a large-scale high-quality tri-modality dataset named VALOR-1M, which contains 1M audiable videos with human annotated audiovisual captions. Extensive experiments show that VALOR can learn strong multimodal correlations and be generalized to various downstream tasks (e.g., retrieval, captioning and question answering), with different input modalities (e.g., vision-language, audio-language and audiovisual-language). VALOR achieves new state-of-the-art performances on series of public cross-modality benchmarks. Code and data are available at project page https://casia-iva-group.github.io/projects/VALOR.

Multimodal agents, which integrate a controller e.g., a vision language model) with external tools, have demonstrated remarkable capabilities in tackling complex multimodal tasks. Existing approaches for training these agents, both supervised fine-tuning and reinforcement learning, depend on extensive human-annotated task-answer pairs and tool trajectories. However, for complex multimodal tasks, such annotations are prohibitively expensive or impractical to obtain. In this paper, we propose an iterative tool usage exploration method for multimodal agents without any pre-collected data, namely SPORT, via step-wise preference optimization to refine the trajectories of tool usage. Our method enables multimodal agents to autonomously discover effective tool usage strategies through self-exploration and optimization, eliminating the bottleneck of human annotation. SPORT has four iterative components: task synthesis, step sampling, step verification, and preference tuning. We first synthesize multimodal tasks using language models. Then, we introduce a novel trajectory exploration scheme, where step sampling and step verification are executed alternately to solve synthesized tasks. In step sampling, the agent tries different tools and obtains corresponding results. In step verification, we employ a verifier to provide AI feedback to construct step-wise preference data. The data is subsequently used to update the controller for tool usage through preference tuning, producing a SPORT agent. By interacting with real environments, the SPORT agent gradually evolves into a more refined and capable system. Evaluation in the GTA and GAIA benchmarks shows that the SPORT agent achieves 6.41% and 3.64% improvements, underscoring the generalization and effectiveness introduced by our method. The project page is https://SPORT-Agents.github.io.

Unified multimodal models (UMMs) strive to consolidate visual understanding and visual generation within a single architecture. However, prevailing training paradigms independently optimize understanding via sparse text signals and generation through dense pixel objectives. Such a decoupled strategy yields misaligned representation spaces, isolating visual understanding from generation and hindering their mutual reinforcement. This work presents the first systematic investigation into generative post-training, where we formulate hierarchical visual tasks as generative proxies to bridge the isolation in UMMs. Our empirical investigation reveals that high-level semantic tasks, particularly image segmentation, serve as optimal proxies. Unlike low-level tasks that distract models with texture details, segmentation provides structural semantics that significantly enhance both vision-centric perception and generative layout fidelity. Building upon these insights, we introduce Semantic Generative Tuning (SGT), a novel paradigm that leverages segmentation as a generative proxy to align and synergize multimodal capabilities. Mechanistic analyses further demonstrate that SGT fundamentally improves feature linear separability and optimizes visual-textual attention allocation pattern. Extensive evaluations show that SGT consistently improves both multimodal comprehension and generative fidelity across mainstream benchmarks. Our code is available on the https://song2yu.github.io/SGT/.

Despite remarkable advancements in Multimodal Large Language Models (MLLMs), a fundamental question remains: are MLLMs robust to contradicting modalities? To rigorously study this, we introduce MMA-Bench comprising videos and tasks that probe a model's reliance on specific modalities. Using black-box and white-box interpretability techniques, we provide a critical analysis of the brittleness of both open- and closed-sourced MLLMs. We show that current MLLMs struggle under misaligned audio-visual pairs and simple misleading text, thereby lacking robust multi-modal reasoning. Building on these findings, we propose a modality alignment tuning strategy to teach the model when to prioritize, leverage, or ignore specific modality cues. Through extensive experiments and analysis, we show that our alignment tuning yields demonstrably stronger multimodal grounding. This work provides both interpretability tools and a clear path toward developing MLLMs with intrinsically reliable cross-modal reasoning. Code and dataset will be publicly available.

Generating combined visual and auditory sensory experiences is critical for the consumption of immersive content. Recent advances in neural generative models have enabled the creation of high-resolution content across multiple modalities such as images, text, speech, and videos. Despite these successes, there remains a significant gap in the generation of high-quality spatial audio that complements generated visual content. Furthermore, current audio generation models excel in either generating natural audio or speech or music but fall short in integrating spatial audio cues necessary for immersive experiences. In this work, we introduce SEE-2-SOUND, a zero-shot approach that decomposes the task into (1) identifying visual regions of interest; (2) locating these elements in 3D space; (3) generating mono-audio for each; and (4) integrating them into spatial audio. Using our framework, we demonstrate compelling results for generating spatial audio for high-quality videos, images, and dynamic images from the internet, as well as media generated by learned approaches.

Multimodal large language models (MLLMs) have achieved remarkable performance on complex tasks with multimodal context. However, it is still understudied whether they exhibit modality preference when processing multimodal contexts. To study this question, we first build a MC\textsuperscript{2} benchmark under controlled evidence conflict scenarios to systematically evaluate modality preference, which is the tendency to favor one modality over another when making decisions based on multimodal conflicting evidence. Our extensive evaluation reveals that all 18 tested MLLMs generally demonstrate clear modality bias, and modality preference can be influenced by external interventions. An in-depth analysis reveals that the preference direction can be captured within the latent representations of MLLMs. Built on this, we propose a probing and steering method based on representation engineering to explicitly control modality preference without additional fine-tuning or carefully crafted prompts. Our method effectively amplifies modality preference toward a desired direction and applies to downstream tasks such as hallucination mitigation and multimodal machine translation, yielding promising improvements.

Diffusion-based video generation can create realistic videos, yet existing image- and text-based conditioning fails to offer precise motion control. Prior methods for motion-conditioned synthesis typically require model-specific fine-tuning, which is computationally expensive and restrictive. We introduce Time-to-Move (TTM), a training-free, plug-and-play framework for motion- and appearance-controlled video generation with image-to-video (I2V) diffusion models. Our key insight is to use crude reference animations obtained through user-friendly manipulations such as cut-and-drag or depth-based reprojection. Motivated by SDEdit's use of coarse layout cues for image editing, we treat the crude animations as coarse motion cues and adapt the mechanism to the video domain. We preserve appearance with image conditioning and introduce dual-clock denoising, a region-dependent strategy that enforces strong alignment in motion-specified regions while allowing flexibility elsewhere, balancing fidelity to user intent with natural dynamics. This lightweight modification of the sampling process incurs no additional training or runtime cost and is compatible with any backbone. Extensive experiments on object and camera motion benchmarks show that TTM matches or exceeds existing training-based baselines in realism and motion control. Beyond this, TTM introduces a unique capability: precise appearance control through pixel-level conditioning, exceeding the limits of text-only prompting. Visit our project page for video examples and code: https://time-to-move.github.io/.

Recently, Omni-modal large language models (OLLMs) have sparked a new wave of research, achieving impressive results in tasks such as audio-video understanding and real-time environment perception. However, hallucination issues still persist. Similar to the bimodal setting, the priors from the text modality tend to dominate, leading OLLMs to rely more heavily on textual cues while neglecting visual and audio information. In addition, fully multimodal scenarios introduce new challenges. Most existing models align visual or auditory modalities with text independently during training, while ignoring the intrinsic correlations between video and its corresponding audio. This oversight results in hallucinations when reasoning requires interpreting hidden audio cues embedded in video content. To address these challenges, we propose OmniDPO, a preference-alignment framework designed to mitigate hallucinations in OLLMs. Specifically, OmniDPO incorporates two strategies: (1) constructing text-preference sample pairs to enhance the model's understanding of audio-video interactions; and (2) constructing multimodal-preference sample pairs to strengthen the model's attention to visual and auditory information. By tackling both challenges, OmniDPO effectively improves multimodal grounding and reduces hallucination. Experiments conducted on two OLLMs demonstrate that OmniDPO not only effectively mitigates multimodal hallucinations but also significantly enhances the models' reasoning capabilities across modalities. All code and datasets will be released upon paper acceptance.

Existing video-to-audio (V2A) generation methods predominantly rely on text prompts alongside visual information to synthesize audio. However, two critical bottlenecks persist: semantic granularity gaps in training data, such as conflating acoustically distinct sounds under coarse labels, and textual ambiguity in describing micro-acoustic features. These bottlenecks make it difficult to perform fine-grained sound synthesis using text-controlled modes. To address these limitations, we propose AC-Foley, an audio-conditioned V2A model that directly leverages reference audio to achieve precise and fine-grained control over generated sounds. This approach enables fine-grained sound synthesis, timbre transfer, zero-shot sound generation, and improved audio quality. By directly conditioning on audio signals, our approach bypasses the semantic ambiguities of text descriptions while enabling precise manipulation of acoustic attributes. Empirically, AC-Foley achieves state-of-the-art performance for Foley generation when conditioned on reference audio, while remaining competitive with state-of-the-art video-to-audio methods even without audio conditioning.

Various applications of voice synthesis have been developed independently despite the fact that they generate "voice" as output in common. In addition, the majority of voice synthesis models currently rely on annotated audio data, but it is crucial to scale them to self-supervised datasets in order to effectively capture the wide range of acoustic variations present in human voice, including speaker identity, emotion, and prosody. In this work, we propose Make-A-Voice, a unified framework for synthesizing and manipulating voice signals from discrete representations. Make-A-Voice leverages a "coarse-to-fine" approach to model the human voice, which involves three stages: 1) semantic stage: model high-level transformation between linguistic content and self-supervised semantic tokens, 2) acoustic stage: introduce varying control signals as acoustic conditions for semantic-to-acoustic modeling, and 3) generation stage: synthesize high-fidelity waveforms from acoustic tokens. Make-A-Voice offers notable benefits as a unified voice synthesis framework: 1) Data scalability: the major backbone (i.e., acoustic and generation stage) does not require any annotations, and thus the training data could be scaled up. 2) Controllability and conditioning flexibility: we investigate different conditioning mechanisms and effectively handle three voice synthesis applications, including text-to-speech (TTS), voice conversion (VC), and singing voice synthesis (SVS) by re-synthesizing the discrete voice representations with prompt guidance. Experimental results demonstrate that Make-A-Voice exhibits superior audio quality and style similarity compared with competitive baseline models. Audio samples are available at https://Make-A-Voice.github.io

Building upon large language models (LLMs), recent large multimodal models (LMMs) unify cross-model understanding and generation into a single framework. However, LMMs still struggle to achieve accurate image-text alignment, prone to generating text responses contradicting the visual input or failing to follow the text-to-image prompts. Current solutions require external supervision (e.g., human feedback or reward models) and only address unidirectional tasks-either understanding or generation. In this work, based on the observation that understanding and generation are inverse dual tasks, we introduce a self-supervised dual reward mechanism to reinforce the understanding and generation capabilities of LMMs. Specifically, we sample multiple outputs for a given input in one task domain, then reverse the input-output pairs to compute the dual likelihood of the model as self-rewards for optimization. Extensive experimental results on visual understanding and generation benchmarks demonstrate that our method can effectively enhance the performance of the model without any external supervision, especially achieving remarkable improvements in text-to-image tasks.

Designed to learn long-range interactions on sequential data, transformers continue to show state-of-the-art results on a wide variety of tasks. In contrast to CNNs, they contain no inductive bias that prioritizes local interactions. This makes them expressive, but also computationally infeasible for long sequences, such as high-resolution images. We demonstrate how combining the effectiveness of the inductive bias of CNNs with the expressivity of transformers enables them to model and thereby synthesize high-resolution images. We show how to (i) use CNNs to learn a context-rich vocabulary of image constituents, and in turn (ii) utilize transformers to efficiently model their composition within high-resolution images. Our approach is readily applied to conditional synthesis tasks, where both non-spatial information, such as object classes, and spatial information, such as segmentations, can control the generated image. In particular, we present the first results on semantically-guided synthesis of megapixel images with transformers and obtain the state of the art among autoregressive models on class-conditional ImageNet. Code and pretrained models can be found at https://github.com/CompVis/taming-transformers .

Audio-video generation has often relied on complex multi-stage architectures or sequential synthesis of sound and visuals. We introduce Ovi, a unified paradigm for audio-video generation that models the two modalities as a single generative process. By using blockwise cross-modal fusion of twin-DiT modules, Ovi achieves natural synchronization and removes the need for separate pipelines or post hoc alignment. To facilitate fine-grained multimodal fusion modeling, we initialize an audio tower with an architecture identical to that of a strong pretrained video model. Trained from scratch on hundreds of thousands of hours of raw audio, the audio tower learns to generate realistic sound effects, as well as speech that conveys rich speaker identity and emotion. Fusion is obtained by jointly training the identical video and audio towers via blockwise exchange of timing (via scaled-RoPE embeddings) and semantics (through bidirectional cross-attention) on a vast video corpus. Our model enables cinematic storytelling with natural speech and accurate, context-matched sound effects, producing movie-grade video clips. All the demos, code and model weights are published at https://aaxwaz.github.io/Ovi

While synthetic data has proven effective for improving scientific reasoning in the text domain, multimodal reasoning remains constrained by the difficulty of synthesizing scientifically rigorous images. Existing Text-to-Image (T2I) models often produce outputs that are visually plausible yet scientifically incorrect, resulting in a persistent visual-logic divergence that limits their value for downstream reasoning. Motivated by recent advances in next-generation T2I models, we conduct a systematic study of scientific image synthesis across generation paradigms, evaluation, and downstream use. We analyze both direct pixel-based generation and programmatic synthesis, and propose ImgCoder, a logic-driven framework that follows an explicit "understand - plan - code" workflow to improve structural precision. To rigorously assess scientific correctness, we introduce SciGenBench, which evaluates generated images based on information utility and logical validity. Our evaluation reveals systematic failure modes in pixel-based models and highlights a fundamental expressiveness-precision trade-off. Finally, we show that fine-tuning Large Multimodal Models (LMMs) on rigorously verified synthetic scientific images yields consistent reasoning gains, with potential scaling trends analogous to the text domain, validating high-fidelity scientific synthesis as a viable path to unlocking massive multimodal reasoning capabilities.

We consider the task of generating diverse and realistic videos guided by natural audio samples from a wide variety of semantic classes. For this task, the videos are required to be aligned both globally and temporally with the input audio: globally, the input audio is semantically associated with the entire output video, and temporally, each segment of the input audio is associated with a corresponding segment of that video. We utilize an existing text-conditioned video generation model and a pre-trained audio encoder model. The proposed method is based on a lightweight adaptor network, which learns to map the audio-based representation to the input representation expected by the text-to-video generation model. As such, it also enables video generation conditioned on text, audio, and, for the first time as far as we can ascertain, on both text and audio. We validate our method extensively on three datasets demonstrating significant semantic diversity of audio-video samples and further propose a novel evaluation metric (AV-Align) to assess the alignment of generated videos with input audio samples. AV-Align is based on the detection and comparison of energy peaks in both modalities. In comparison to recent state-of-the-art approaches, our method generates videos that are better aligned with the input sound, both with respect to content and temporal axis. We also show that videos produced by our method present higher visual quality and are more diverse.

Current multimodal and multitask foundation models like 4M or UnifiedIO show promising results, but in practice their out-of-the-box abilities to accept diverse inputs and perform diverse tasks are limited by the (usually rather small) number of modalities and tasks they are trained on. In this paper, we expand upon the capabilities of them by training a single model on tens of highly diverse modalities and by performing co-training on large-scale multimodal datasets and text corpora. This includes training on several semantic and geometric modalities, feature maps from recent state of the art models like DINOv2 and ImageBind, pseudo labels of specialist models like SAM and 4DHumans, and a range of new modalities that allow for novel ways to interact with the model and steer the generation, for example image metadata or color palettes. A crucial step in this process is performing discrete tokenization on various modalities, whether they are image-like, neural network feature maps, vectors, structured data like instance segmentation or human poses, or data that can be represented as text. Through this, we expand on the out-of-the-box capabilities of multimodal models and specifically show the possibility of training one model to solve at least 3x more tasks/modalities than existing ones and doing so without a loss in performance. This enables more fine-grained and controllable multimodal generation capabilities and allows us to study the distillation of models trained on diverse data and objectives into a unified model. We successfully scale the training to a three billion parameter model using tens of modalities and different datasets. The resulting models and training code are open sourced at 4m.epfl.ch.

There has been a recent explosion of computer vision models which perform many tasks and are composed of an image encoder (usually a ViT) and an autoregressive decoder (usually a Transformer). However, most of this work simply presents one system and its results, leaving many questions regarding design decisions and trade-offs of such systems unanswered. In this work, we aim to provide such answers. We take a close look at autoregressive decoders for multi-task learning in multimodal computer vision, including classification, captioning, visual question answering, and optical character recognition. Through extensive systematic experiments, we study the effects of task and data mixture, training and regularization hyperparameters, conditioning type and specificity, modality combination, and more. Importantly, we compare these to well-tuned single-task baselines to highlight the cost incurred by multi-tasking. A key finding is that a small decoder learned on top of a frozen pretrained encoder works surprisingly well. We call this setup locked-image tuning with decoder (LiT-decoder). It can be seen as teaching a decoder to interact with a pretrained vision model via natural language.

Connecting text and visual modalities plays an essential role in generative intelligence. For this reason, inspired by the success of large language models, significant research efforts are being devoted to the development of Multimodal Large Language Models (MLLMs). These models can seamlessly integrate visual and textual modalities, both as input and output, while providing a dialogue-based interface and instruction-following capabilities. In this paper, we provide a comprehensive review of recent visual-based MLLMs, analyzing their architectural choices, multimodal alignment strategies, and training techniques. We also conduct a detailed analysis of these models across a wide range of tasks, including visual grounding, image generation and editing, visual understanding, and domain-specific applications. Additionally, we compile and describe training datasets and evaluation benchmarks, conducting comparisons among existing models in terms of performance and computational requirements. Overall, this survey offers a comprehensive overview of the current state of the art, laying the groundwork for future MLLMs.

Memorization in generative models extends far beyond verbatim text reproduction--it manifests through non-literal patterns, semantic associations, and surprisingly, across modalities in transcript-conditioned generation tasks such as Lyrics-to-Song (L2S) and Text-to-Video (T2V) models. We reveal a new class of cross-modality memorization where models trained on these tasks leak copyrighted content through indirect, phonetic pathways invisible to traditional text-based analysis. In this work, we introduce Adversarial PhoneTic Prompting (APT), an attack that replaces iconic phrases with homophonic alternatives--e.g., "mom's spaghetti" becomes "Bob's confetti"--preserving the acoustic form while largely changing semantic content. We demonstrate that models can be prompted to regurgitate memorized songs using phonetically similar but semantically unrelated lyrics. Despite the semantic drift, black-box models like SUNO and open-source models like YuE generate outputs that are strikingly similar to the original songs--melodically, rhythmically, and vocally--achieving high scores on AudioJudge, CLAP, and CoverID. These effects persist across genres and languages. More surprisingly, we find that phonetic prompts alone can trigger visual memorization in text-to-video models: when given altered lyrics from Lose Yourself, Veo 3 generates scenes that mirror the original music video--complete with a hooded rapper and dim urban settings--despite no explicit visual cues in the prompt. This cross-modality leakage represents an unprecedented threat: models memorize deep, structural patterns that transcend their training modality, making traditional safety measures like copyright filters ineffective. Our findings reveal a fundamental vulnerability in transcript-conditioned generative models and raise urgent concerns around copyright, provenance, and secure deployment of multimodal generation systems.

While recent advancements in multimodal language models have enabled image generation from expressive multi-image instructions, existing methods struggle to maintain performance under complex interleaved instructions. This limitation stems from the structural separation of images and text in current paradigms, which forces models to bridge difficult long-range dependencies to match descriptions with visual targets. To address these challenges, we propose Images iN SEnTences (a.k.a, INSET), a unified generation model that seamlessly embeds images as native vocabulary within textual instructions. By positioning visual features directly at their corresponding semantic slots, INSET leverages the contextual locality of transformers for precise object binding, effectively treating images as dense, expressive language tokens. Furthermore, we introduce a scalable data engine that synthesizes 15M high-quality interleaved samples from standard image and video datasets, utilizing VLMs and LLMs to construct rich, long-horizon sequences. Evaluation results on InterleaveBench demonstrate that INSET significantly outperforms state-of-the-art methods in multi-image consistency and text alignment, with performance gaps widening as input complexity increases. Beyond standard generation, our approach inherently extends to multimodal image editing, integrating visual content as part of the instruction to facilitate highly expressive and creative visual manipulations.

Text-to-image (T2I) generation has seen significant progress with diffusion models, enabling generation of photo-realistic images from text prompts. Despite this progress, existing methods still face challenges in following complex text prompts, especially those requiring compositional and multi-step reasoning. Given such complex instructions, SOTA models often make mistakes in faithfully modeling object attributes, and relationships among them. In this work, we present an alternate paradigm for T2I synthesis, decomposing the task of complex multi-step generation into three steps, (a) Generate: we first generate an image using existing diffusion models (b) Plan: we make use of Multi-Modal LLMs (MLLMs) to identify the mistakes in the generated image expressed in terms of individual objects and their properties, and produce a sequence of corrective steps required in the form of an edit-plan. (c) Edit: we make use of an existing text-guided image editing models to sequentially execute our edit-plan over the generated image to get the desired image which is faithful to the original instruction. Our approach derives its strength from the fact that it is modular in nature, is training free, and can be applied over any combination of image generation and editing models. As an added contribution, we also develop a model capable of compositional editing, which further helps improve the overall accuracy of our proposed approach. Our method flexibly trades inference time compute with performance on compositional text prompts. We perform extensive experimental evaluation across 3 benchmarks and 10 T2I models including DALLE-3 and the latest -- SD-3.5-Large. Our approach not only improves the performance of the SOTA models, by upto 3 points, it also reduces the performance gap between weaker and stronger models. https://dair-iitd.github.io/GraPE/{https://dair-iitd.github.io/GraPE/}

Multimodal modeling represents a vital step from modality-agnostic reasoning toward world modeling. While early approaches predominantly rely on late-fusion that assembles encoders and frozen language backbones with output heads, recent efforts have shifted the paradigm toward native multimodal modeling (NMM) with the intrinsic integration of modalities for superior multimodal performance. Despite its potential, the design space of native architectures remains insufficiently defined. In this paper, we present the community with a formalized roadmap for this transition. Specifically, we formally define the architectural nativity, distinguishing mid-fusion and early-fusion from non-native paradigms. We further organize the existing native models through the lens of input-output duality into three categories: (i) Multi-to-Text for cross-modal comprehension with text-only output; (ii) Multi-to-Target for scenario-oriented generation, e.g., image, audio and video generation, and (iii) Multi-to-Multi for unified modeling with symmetric input-output. We deliver a comprehensive and industrial-grade investigation into the transition toward the definitive NMM framework, where understanding and generation seamlessly coexist within a unified transformer paradigm. We systematically unpack the end-to-end pipeline from industrial perspectives from architectural coordination, massive data curation, to full-stack training recipes, inference & deployment, and the comprehensive evaluation for truly native modeling.

Multimodal Contrastive Learning (MCL) advances in aligning different modalities and generating multimodal representations in a joint space. By leveraging contrastive learning across diverse modalities, large-scale multimodal data enhances representational quality. However, a critical yet often overlooked challenge remains: multimodal data is rarely collected in a single process, and training from scratch is computationally expensive. Instead, emergent multimodal data can be used to optimize existing models gradually, i.e., models are trained on a sequence of modality pair data. We define this problem as Continual Multimodal Contrastive Learning (CMCL), an underexplored yet crucial research direction at the intersection of multimodal and continual learning. In this paper, we formulate CMCL through two specialized principles of stability and plasticity. We theoretically derive a novel optimization-based method, which projects updated gradients from dual sides onto subspaces where any gradient is prevented from interfering with the previously learned knowledge. Two upper bounds provide theoretical insights on both stability and plasticity in our solution. Beyond our theoretical contributions, we conduct experiments on multiple datasets by comparing our method against advanced continual learning baselines. The empirical results further support our claims and demonstrate the efficacy of our method. Our codes are available at https://github.com/Xiaohao-Liu/CMCL.

Image degradation from blur, noise, compression, and poor illumination severely undermines multimodal understanding in real-world settings. Unified multimodal models that combine understanding and generation within a single architecture are a natural fit for this challenge, as their generative pathway can model the fine-grained visual structure that degradation destroys. Yet these models fail to leverage their own generative capacity on degraded inputs. We trace this disconnect to two compounding factors: existing training regimes never ask the model to invoke generation during reasoning, and the standard decode-reencode pathway does not support effective joint optimization. We present CLEAR, a framework that connects the two capabilities through three progressive steps: (1) supervised fine-tuning on a degradation-aware dataset to establish the generate-then-answer reasoning pattern; (2) a Latent Representation Bridge that replaces the decode-reencode detour with a direct, optimizable connection between generation and reasoning; (3) Interleaved GRPO, a reinforcement learning method that jointly optimizes text reasoning and visual generation under answer-correctness rewards. We construct MMD-Bench, covering three degradation severity levels across six standard multimodal benchmarks. Experiments show that CLEAR substantially improves robustness on degraded inputs while preserving clean-image performance. Our analysis further reveals that removing pixel-level reconstruction supervision leads to intermediate visual states with higher perceptual quality, suggesting that task-driven optimization and visual quality are naturally aligned.