arXiv subjects

Zecheng He

Publications and source records attributed to Zecheng He.

At least 19 recordsLinked to original sources

arXiv2026

Think in Strokes, Not Pixels: Process-Driven Image Generation via Interleaved Reasoning

Humans paint images incrementally: they plan a global layout, sketch a coarse draft, inspect, and refine details, and most importantly, each step is grounded in the evolving visual states. However, can unified multimodal models trained on text-image interleaved datasets also imagine the chain of intermediate states? In this paper, we introduce process-driven image generation, a multi-step paradigm that decomposes synthesis into an interleaved reasoning trajectory of thoughts and actions. Rather than generating images in a single step, our approach unfolds across multiple iterations, each consisting of 4 stages: textual planning, visual drafting, textual reflection, and visual refinement. The textual reasoning explicitly conditions how the visual state should evolve, while the generated visual intermediate in turn constrains and grounds the next round of textual reasoning. A core challenge of process-driven generation stems from the ambiguity of intermediate states: how can models evaluate each partially-complete image? We address this through dense, step-wise supervision that maintains two complementary constraints: for the visual intermediate states, we enforce the spatial and semantic consistency; for the textual intermediate states, we preserve the prior visual knowledge while enabling the model to identify and correct prompt-violating elements. This makes the generation process explicit, interpretable, and directly supervisable. To validate proposed method, we conduct experiments under various text-to-image generation benchmarks.

cs.CV↗

arXiv2026

UniT: Unified Multimodal Chain-of-Thought Test-time Scaling

Unified models can handle both multimodal understanding and generation within a single architecture, yet they typically operate in a single pass without iteratively refining their outputs. Many multimodal tasks, especially those involving complex spatial compositions, multiple interacting objects, or evolving instructions, require decomposing instructions, verifying intermediate results, and making iterative corrections. While test-time scaling (TTS) has demonstrated that allocating additional inference compute for iterative reasoning substantially improves language model performance, extending this paradigm to unified multimodal models remains an open challenge. We introduce UniT, a framework for multimodal chain-of-thought test-time scaling that enables a single unified model to reason, verify, and refine across multiple rounds. UniT combines agentic data synthesis, unified model training, and flexible test-time inference to elicit cognitive behaviors including verification, subgoal decomposition, and content memory. Our key findings are: (1) unified models trained on short reasoning trajectories generalize to longer inference chains at test time; (2) sequential chain-of-thought reasoning provides a more scalable and compute-efficient TTS strategy than parallel sampling; (3) training on generation and editing trajectories improves out-of-distribution visual reasoning. These results establish multimodal test-time scaling as an effective paradigm for advancing both generation and understanding in unified models.

cs.CV↗

arXiv2026

Non-Markov Multi-Round Conversational Image Generation with History-Conditioned MLLMs

Conversational image generation requires a model to follow user instructions across multiple rounds of interaction, grounded in interleaved text and images that accumulate as chat history. While recent multimodal large language models (MLLMs) can generate and edit images, most existing multi-turn benchmarks and training recipes are effectively Markov: the next output depends primarily on the most recent image, enabling shortcut solutions that ignore long-range history. In this work we formalize and target the more challenging non-Markov setting, where a user may refer back to earlier states, undo changes, or reference entities introduced several rounds ago. We present (i) non-Markov multi-round data construction strategies, including rollback-style editing that forces retrieval of earlier visual states and name-based multi-round personalization that binds names to appearances across rounds; (ii) a history-conditioned training and inference framework with token-level caching to prevent multi-round identity drift; and (iii) enabling improvements for high-fidelity image reconstruction and editable personalization, including a reconstruction-based DiT detokenizer and a multi-stage fine-tuning curriculum. We demonstrate that explicitly training for non-Markov interactions yields substantial improvements in multi-round consistency and instruction compliance, while maintaining strong single-round editing and personalization.

cs.CV↗

arXiv2025

Exploring MLLM-Diffusion Information Transfer with MetaCanvas

Multimodal learning has rapidly advanced visual understanding, largely via multimodal large language models (MLLMs) that use powerful LLMs as cognitive cores. In visual generation, however, these powerful core models are typically reduced to global text encoders for diffusion models, leaving most of their reasoning and planning ability unused. This creates a gap: current multimodal LLMs can parse complex layouts, attributes, and knowledge-intensive scenes, yet struggle to generate images or videos with equally precise and structured control. We propose MetaCanvas, a lightweight framework that lets MLLMs reason and plan directly in spatial and spatiotemporal latent spaces and interface tightly with diffusion generators. We empirically implement MetaCanvas on three different diffusion backbones and evaluate it across six tasks, including text-to-image generation, text/image-to-video generation, image/video editing, and in-context video generation, each requiring precise layouts, robust attribute binding, and reasoning-intensive control. MetaCanvas consistently outperforms global-conditioning baselines, suggesting that treating MLLMs as latent-space planners is a promising direction for narrowing the gap between multimodal understanding and generation.

cs.CV↗

arXiv2025

Improving Chain-of-Thought Efficiency for Autoregressive Image Generation

Autoregressive multimodal large language models have recently gained popularity for image generation, driven by advances in foundation models. To enhance alignment and detail, newer approaches employ chain-of-thought (CoT) reasoning, expanding user inputs into elaborated prompts prior to image synthesis. However, this strategy can introduce unnecessary redundancy -- a phenomenon we call visual overthinking -- which increases computational costs and can introduce details that contradict the original prompt. In this work, we explore how to generate more concise CoT sequences for more efficient image generation. We introduce ShortCoTI, a lightweight optimization framework that encourages more concise CoT while preserving output image quality. ShortCoTI rewards more concise prompts with an adaptive function that scales according to an estimated difficulty for each task. Incorporating this reward into a reinforcement learning paradigm reduces prompt reasoning length by 54% while maintaining or slightly improving quality metrics across multiple benchmarks (T2I-CompBench, GenEval). Qualitative analysis shows that our method eliminates verbose explanations and repetitive refinements, producing reasoning prompts that are both concise and semantically rich. As a result, ShortCoTI improves computational efficiency without compromising the fidelity or visual appeal of generated images.

cs.CV↗

arXiv2025

Populate-A-Scene: Affordance-Aware Human Video Generation

Can a video generation model be repurposed as an interactive world simulator? We explore the affordance perception potential of text-to-video models by teaching them to predict human-environment interaction. Given a scene image and a prompt describing human actions, we fine-tune the model to insert a person into the scene, while ensuring coherent behavior, appearance, harmonization, and scene affordance. Unlike prior work, we infer human affordance for video generation (i.e., where to insert a person and how they should behave) from a single scene image, without explicit conditions like bounding boxes or body poses. An in-depth study of cross-attention heatmaps demonstrates that we can uncover the inherent affordance perception of a pre-trained video model without labeled affordance datasets.

cs.CV↗

arXiv2025

Token-Shuffle: Towards High-Resolution Image Generation with Autoregressive Models

Autoregressive (AR) models, long dominant in language generation, are increasingly applied to image synthesis but are often considered less competitive than Diffusion-based models. A primary limitation is the substantial number of image tokens required for AR models, which constrains both training and inference efficiency, as well as image resolution. To address this, we present Token-Shuffle, a novel yet simple method that reduces the number of image tokens in Transformer. Our key insight is the dimensional redundancy of visual vocabularies in Multimodal Large Language Models (MLLMs), where low-dimensional visual codes from visual encoder are directly mapped to high-dimensional language vocabularies. Leveraging this, we consider two key operations: token-shuffle, which merges spatially local tokens along channel dimension to decrease the input token number, and token-unshuffle, which untangles the inferred tokens after Transformer blocks to restore the spatial arrangement for output. Jointly training with textual prompts, our strategy requires no additional pretrained text-encoder and enables MLLMs to support extremely high-resolution image synthesis in a unified next-token prediction way while maintaining efficient training and inference. For the first time, we push the boundary of AR text-to-image generation to a resolution of 2048x2048 with gratifying generation performance. In GenAI-benchmark, our 2.7B model achieves 0.77 overall score on hard prompts, outperforming AR models LlamaGen by 0.18 and diffusion models LDM by 0.15. Exhaustive large-scale human evaluations also demonstrate our prominent image generation ability in terms of text-alignment, visual flaw, and visual appearance. We hope that Token-Shuffle can serve as a foundational design for efficient high-resolution image generation within MLLMs.

cs.CV↗

arXiv2025

MoCha: Towards Movie-Grade Talking Character Synthesis

Recent advancements in video generation have achieved impressive motion realism, yet they often overlook character-driven storytelling, a crucial task for automated film, animation generation. We introduce Talking Characters, a more realistic task to generate talking character animations directly from speech and text. Unlike talking head, Talking Characters aims at generating the full portrait of one or more characters beyond the facial region. In this paper, we propose MoCha, the first of its kind to generate talking characters. To ensure precise synchronization between video and speech, we propose a speech-video window attention mechanism that effectively aligns speech and video tokens. To address the scarcity of large-scale speech-labeled video datasets, we introduce a joint training strategy that leverages both speech-labeled and text-labeled video data, significantly improving generalization across diverse character actions. We also design structured prompt templates with character tags, enabling, for the first time, multi-character conversation with turn-based dialogue-allowing AI-generated characters to engage in context-aware conversations with cinematic coherence. Extensive qualitative and quantitative evaluations, including human preference studies and benchmark comparisons, demonstrate that MoCha sets a new standard for AI-generated cinematic storytelling, achieving superior realism, expressiveness, controllability and generalization.

cs.CV↗

arXiv2025

Movie Weaver: Tuning-Free Multi-Concept Video Personalization with Anchored Prompts

Video personalization, which generates customized videos using reference images, has gained significant attention. However, prior methods typically focus on single-concept personalization, limiting broader applications that require multi-concept integration. Attempts to extend these models to multiple concepts often lead to identity blending, which results in composite characters with fused attributes from multiple sources. This challenge arises due to the lack of a mechanism to link each concept with its specific reference image. We address this with anchored prompts, which embed image anchors as unique tokens within text prompts, guiding accurate referencing during generation. Additionally, we introduce concept embeddings to encode the order of reference images. Our approach, Movie Weaver, seamlessly weaves multiple concepts-including face, body, and animal images-into one video, allowing flexible combinations in a single model. The evaluation shows that Movie Weaver outperforms existing methods for multi-concept video personalization in identity preservation and overall quality.

cs.CV↗

arXiv2024

Llama Learns to Direct: DirectorLLM for Human-Centric Video Generation

In this paper, we introduce DirectorLLM, a novel video generation model that employs a large language model (LLM) to orchestrate human poses within videos. As foundational text-to-video models rapidly evolve, the demand for high-quality human motion and interaction grows. To address this need and enhance the authenticity of human motions, we extend the LLM from a text generator to a video director and human motion simulator. Utilizing open-source resources from Llama 3, we train the DirectorLLM to generate detailed instructional signals, such as human poses, to guide video generation. This approach offloads the simulation of human motion from the video generator to the LLM, effectively creating informative outlines for human-centric scenes. These signals are used as conditions by the video renderer, facilitating more realistic and prompt-following video generation. As an independent LLM module, it can be applied to different video renderers, including UNet and DiT, with minimal effort. Experiments on automatic evaluation benchmarks and human evaluations show that our model outperforms existing ones in generating videos with higher human motion fidelity, improved prompt faithfulness, and enhanced rendered subject naturalness.

cs.CV↗

arXiv2024

Movie Gen: A Cast of Media Foundation Models

We present Movie Gen, a cast of foundation models that generates high-quality, 1080p HD videos with different aspect ratios and synchronized audio. We also show additional capabilities such as precise instruction-based video editing and generation of personalized videos based on a user's image. Our models set a new state-of-the-art on multiple tasks: text-to-video synthesis, video personalization, video editing, video-to-audio generation, and text-to-audio generation. Our largest video generation model is a 30B parameter transformer trained with a maximum context length of 73K video tokens, corresponding to a generated video of 16 seconds at 16 frames-per-second. We show multiple technical innovations and simplifications on the architecture, latent spaces, training objectives and recipes, data curation, evaluation protocols, parallelization techniques, and inference optimizations that allow us to reap the benefits of scaling pre-training data, model size, and training compute for training large scale media generation models. We hope this paper helps the research community to accelerate progress and innovation in media generation models. All videos from this paper are available at https://go.fb.me/MovieGenResearchVideos.

cs.CV↗

arXiv2024

Imagine yourself: Tuning-Free Personalized Image Generation

Diffusion models have demonstrated remarkable efficacy across various image-to-image tasks. In this research, we introduce Imagine yourself, a state-of-the-art model designed for personalized image generation. Unlike conventional tuning-based personalization techniques, Imagine yourself operates as a tuning-free model, enabling all users to leverage a shared framework without individualized adjustments. Moreover, previous work met challenges balancing identity preservation, following complex prompts and preserving good visual quality, resulting in models having strong copy-paste effect of the reference images. Thus, they can hardly generate images following prompts that require significant changes to the reference image, \eg, changing facial expression, head and body poses, and the diversity of the generated images is low. To address these limitations, our proposed method introduces 1) a new synthetic paired data generation mechanism to encourage image diversity, 2) a fully parallel attention architecture with three text encoders and a fully trainable vision encoder to improve the text faithfulness, and 3) a novel coarse-to-fine multi-stage finetuning methodology that gradually pushes the boundary of visual quality. Our study demonstrates that Imagine yourself surpasses the state-of-the-art personalization model, exhibiting superior capabilities in identity preservation, visual quality, and text alignment. This model establishes a robust foundation for various personalization applications. Human evaluation results validate the model's SOTA superiority across all aspects (identity preservation, text faithfulness, and visual appeal) compared to the previous personalization models.

cs.CV↗

arXiv2023

Adversarial Medical Image with Hierarchical Feature Hiding

Deep learning based methods for medical images can be easily compromised by adversarial examples (AEs), posing a great security flaw in clinical decision-making. It has been discovered that conventional adversarial attacks like PGD which optimize the classification logits, are easy to distinguish in the feature space, resulting in accurate reactive defenses. To better understand this phenomenon and reassess the reliability of the reactive defenses for medical AEs, we thoroughly investigate the characteristic of conventional medical AEs. Specifically, we first theoretically prove that conventional adversarial attacks change the outputs by continuously optimizing vulnerable features in a fixed direction, thereby leading to outlier representations in the feature space. Then, a stress test is conducted to reveal the vulnerability of medical images, by comparing with natural images. Interestingly, this vulnerability is a double-edged sword, which can be exploited to hide AEs. We then propose a simple-yet-effective hierarchical feature constraint (HFC), a novel add-on to conventional white-box attacks, which assists to hide the adversarial feature in the target feature distribution. The proposed method is evaluated on three medical datasets, both 2D and 3D, with different modalities. The experimental results demonstrate the superiority of HFC, \emph{i.e.,} it bypasses an array of state-of-the-art adversarial medical AE detectors more efficiently than competing adaptive attacks, which reveals the deficiencies of medical reactive defense and allows to develop more robust defenses in future.

eess.IV↗

arXiv2023

Trainable Projected Gradient Method for Robust Fine-tuning

Recent studies on transfer learning have shown that selectively fine-tuning a subset of layers or customizing different learning rates for each layer can greatly improve robustness to out-of-distribution (OOD) data and retain generalization capability in the pre-trained models. However, most of these methods employ manually crafted heuristics or expensive hyper-parameter searches, which prevent them from scaling up to large datasets and neural networks. To solve this problem, we propose Trainable Projected Gradient Method (TPGM) to automatically learn the constraint imposed for each layer for a fine-grained fine-tuning regularization. This is motivated by formulating fine-tuning as a bi-level constrained optimization problem. Specifically, TPGM maintains a set of projection radii, i.e., distance constraints between the fine-tuned model and the pre-trained model, for each layer, and enforces them through weight projections. To learn the constraints, we propose a bi-level optimization to automatically learn the best set of projection radii in an end-to-end manner. Theoretically, we show that the bi-level optimization formulation could explain the regularization capability of TPGM. Empirically, with little hyper-parameter search cost, TPGM outperforms existing fine-tuning methods in OOD performance while matching the best in-distribution (ID) performance. For example, when fine-tuned on DomainNet-Real and ImageNet, compared to vanilla fine-tuning, TPGM shows $22\%$ and $10\%$ relative OOD improvement respectively on their sketch counterparts. Code is available at \url{https://github.com/PotatoTian/TPGM}.

cs.CV↗

arXiv2023

SoK: Hardware Defenses Against Speculative Execution Attacks

Speculative execution attacks leverage the speculative and out-of-order execution features in modern computer processors to access secret data or execute code that should not be executed. Secret information can then be leaked through a covert channel. While software patches can be installed for mitigation on existing hardware, these solutions can incur big performance overhead. Hardware mitigation is being studied extensively by the computer architecture community. It has the benefit of preserving software compatibility and the potential for much smaller performance overhead than software solutions. This paper presents a systematization of the hardware defenses against speculative execution attacks that have been proposed. We show that speculative execution attacks consist of 6 critical attack steps. We propose defense strategies, each of which prevents a critical attack step from happening, thus preventing the attack from succeeding. We then summarize 20 hardware defenses and overhead-reducing features that have been proposed. We show that each defense proposed can be classified under one of our defense strategies, which also explains why it can thwart the attack from succeeding. We discuss the scope of the defenses, their performance overhead, and the security-performance trade-offs that can be made.

cs.CR↗

arXiv2022

Medical Aegis: Robust adversarial protectors for medical images

Deep neural network based medical image systems are vulnerable to adversarial examples. Many defense mechanisms have been proposed in the literature, however, the existing defenses assume a passive attacker who knows little about the defense system and does not change the attack strategy according to the defense. Recent works have shown that a strong adaptive attack, where an attacker is assumed to have full knowledge about the defense system, can easily bypass the existing defenses. In this paper, we propose a novel adversarial example defense system called Medical Aegis. To the best of our knowledge, Medical Aegis is the first defense in the literature that successfully addresses the strong adaptive adversarial example attacks to medical images. Medical Aegis boasts two-tier protectors: The first tier of Cushion weakens the adversarial manipulation capability of an attack by removing its high-frequency components, yet posing a minimal effect on classification performance of the original image; the second tier of Shield learns a set of per-class DNN models to predict the logits of the protected model. Deviation from the Shield's prediction indicates adversarial examples. Shield is inspired by the observations in our stress tests that there exist robust trails in the shallow layers of a DNN model, which the adaptive attacks can hardly destruct. Experimental results show that the proposed defense accurately detects adaptive attacks, with negligible overhead for model inference.

cs.CV↗

arXiv2021

CloudShield: Real-time Anomaly Detection in the Cloud

In cloud computing, it is desirable if suspicious activities can be detected by automatic anomaly detection systems. Although anomaly detection has been investigated in the past, it remains unsolved in cloud computing. Challenges are: characterizing the normal behavior of a cloud server, distinguishing between benign and malicious anomalies (attacks), and preventing alert fatigue due to false alarms. We propose CloudShield, a practical and generalizable real-time anomaly and attack detection system for cloud computing. Cloudshield uses a general, pretrained deep learning model with different cloud workloads, to predict the normal behavior and provide real-time and continuous detection by examining the model reconstruction error distributions. Once an anomaly is detected, to reduce alert fatigue, CloudShield automatically distinguishes between benign programs, known attacks, and zero-day attacks, by examining the prediction error distributions. We evaluate the proposed CloudShield on representative cloud benchmarks. Our evaluation shows that CloudShield, using model pretraining, can apply to a wide scope of cloud workloads. Especially, we observe that CloudShield can detect the recently proposed speculative execution attacks, e.g., Spectre and Meltdown attacks, in milliseconds. Furthermore, we show that CloudShield accurately differentiates and prioritizes known attacks, and potential zero-day attacks, from benign programs. Thus, it significantly reduces false alarms by up to 99.0%.

cs.CR↗

arXiv2021

A Hierarchical Feature Constraint to Camouflage Medical Adversarial Attacks

Deep neural networks (DNNs) for medical images are extremely vulnerable to adversarial examples (AEs), which poses security concerns on clinical decision making. Luckily, medical AEs are also easy to detect in hierarchical feature space per our study herein. To better understand this phenomenon, we thoroughly investigate the intrinsic characteristic of medical AEs in feature space, providing both empirical evidence and theoretical explanations for the question: why are medical adversarial attacks easy to detect? We first perform a stress test to reveal the vulnerability of deep representations of medical images, in contrast to natural images. We then theoretically prove that typical adversarial attacks to binary disease diagnosis network manipulate the prediction by continuously optimizing the vulnerable representations in a fixed direction, resulting in outlier features that make medical AEs easy to detect. However, this vulnerability can also be exploited to hide the AEs in the feature space. We propose a novel hierarchical feature constraint (HFC) as an add-on to existing adversarial attacks, which encourages the hiding of the adversarial representation within the normal feature distribution. We evaluate the proposed method on two public medical image datasets, namely {Fundoscopy} and {Chest X-Ray}. Experimental results demonstrate the superiority of our adversarial attack method as it bypasses an array of state-of-the-art adversarial detectors more easily than competing attack methods, supporting that the great vulnerability of medical features allows an attacker more room to manipulate the adversarial representations.

cs.CV↗