arXiv subjects

Weichen Fan

Publications and source records attributed to Weichen Fan.

18 recordsLinked to original sources

arXiv2026

SenseNova-U1.5: Towards Native Unified Visual Intelligence

We launch SenseNova-U1.5, an 8B-MoT native unified multimodal model that understands, reasons about, and generates visual content within an encoder-free and VAE-free architecture. We strengthen its visual interface through spatially coherent patch reconstruction and scale its training with carefully curated generation and editing data, improved task formulation, structural prompt enhancement, and native resolutions of up to 4K. For post-training, we optimize specialized experts for visual aesthetics, bilingual text rendering, infographic generation, and image editing, and consolidate their capabilities through multi-expert on-policy distillation. Across extensive evaluations, SenseNova-U1.5 largely advances image fidelity, text rendering, complex composition, multi-reference editing, and interleaved generation, while improving instruction following and preserving subject identity, geometry, and unmodified regions. Despite limited exposure to structured formats in its generation data, SenseNova-U1.5 generalizes effectively to long, complex, and structured visual instructions, further proving that multimodal understanding can transfer to visual planning and creation. Together, these findings position native unified modelling as a promising path towards systems that perceive, reason and create within a fully end-to-end framework. We will open-source training code, including supervised fine-tuning, reinforcement learning, and on-policy distillation.

cs.CV↗

arXiv2026

Uncovering Understanding-Generation Synergy in Native Unified Multimodal Models: From Representation, Task to System

While unified multimodal models (UMMs) jointly perform visual understanding and generation within a single model, functional unification does not guarantee learning synergy: the two objectives may reinforce each other, compete for capacity, or merely coexist. We investigate their relationship at the representation, task, and system levels in a controlled, structurally native setting without pretrained vision priors. At the representation level, we find that each objective provides useful signal to the other: generation enriches the visual features learned for understanding, while understanding strengthens vision--language alignment for generation. However, when both objectives are forced through the same computation path, one tends to dominate. A task-decoupled architecture that specializes conflicting visual computation while preserving semantic interaction avoids this asymmetric degradation. At the task level, through three case studies, we find positive bidirectional transfer when understanding and generation tasks rely on shared knowledge. At the system level, we show that an end-to-end UMM outperforms a matched planner--executor pipeline on complex tasks that explicitly require both image understanding and generation. Together, these results show that the value of UMMs extends beyond a unified interface: appropriate specialization, shared task knowledge, and end-to-end optimization can turn coexistence into synergy.

cs.CV↗

arXiv2026

Apple-$\pi$: Benchmarking Thinking with Video Towards Law-Grounded Physical Intelligence

Modern video generation models are increasingly hailed as emerging world models with an internalized grasp of physical law. Yet existing benchmarks largely evaluate physical plausibility only at the output level, without verifying whether the model arrives there through a faithful, law-grounded reasoning process. We introduce Apple-PI, the first benchmark that anchors video-model evaluation explicitly in physical laws. Apple-PI comprises three components. 1) Orchard: a dataset of 400 videos covering ten canonical tasks in classical mechanics. It separates single-law tasks for confounder-free diagnosis from multi-law tasks for probing generalization. 2) Benchmark Protocol: a three-stage protocol based on scientific reasoning, including Perception, Formulation, and Deduction. It uses chain-of-frames prompting on infographic-annotated first frames, treating the generated video as the model's visible reasoning trace. 3) Evaluation Suite: a hybrid evaluation suite that combines MLLM-based subjective scoring with physics-law-grounded objective measures. This enables stage-resolved diagnosis of not only whether a model fails, but where it fails. Benchmarking 11 models shows that current video models remain far from reliable law-grounded world simulators, with the best video model scoring only 0.473. Our stage-, pillar-, and source-resolved analyses further expose a Perception-to-Formulation-to-Deduction bottleneck, weak multi-law state transfer, and a persistent Sim-to-Real gap. These findings position Apple-PI as a diagnostic foundation for guiding future video models toward world models with law-grounded physical intelligence.

cs.CV↗

arXiv2026

Show the Signal, Hide the Noise: Spectral Forcing for Pixel-Space Diffusion

Pixel-space diffusion models are trained on full-bandwidth noisy images, yet the useful signal available to the denoiser is strongly frequency dependent. Under rectified-flow diffusion and natural-image power-law spectra, the per-band data-to-noise contour $k^{*}(t) = (1-t)^{-2/\alpha}$ separates a signal-bearing low-frequency region from a noise-dominated high-frequency region at each time $t$. We show that this implicit coarse-to-fine structure is not merely descriptive: it induces a capacity-allocation problem. A standard pixel-space denoiser must discover the moving bandwidth boundary internally and can spend computation on frequency-time regions where the optimal prediction collapses to deterministic baselines rather than data-distribution modeling. To make this boundary explicit, we introduce Spectral Forcing, a parameter-free, time-conditional 2D-DCT low-pass operator applied to the noisy input before the patch embedder. Its cutoff expands monotonically with the diffusion time and becomes the identity at the data endpoint. Through controlled synthetic experiments, we identify the regime in which the operator is beneficial: coarse patch tokenization and data whose high-frequency content is predominantly noise rather than essential signal. On ImageNet-256 with JiT-700M/32, Spectral Forcing consistently improves both FID and Inception Score across different training epochs, demonstrating robust gains throughout training; at finer tokenization, the spectral forcing is still competitive. We further insert the unchanged operator into SenseNova-U1, a unified text-to-image model, where it improves DPG-Bench and GenEval, showing that the input-side spectral prior transfers beyond class-conditional generation. These results suggest a route to capacity-efficient pixel-space diffusion by showing the signal and hiding the noise.

cs.CV↗

arXiv2026

From Pixels to Words -- Towards Native One-Vision Models at Scale

Current vision-language models (VLMs) typically stitch together separate image encoders and language decoders via multi-stage alignment, a modular framework that inevitably fragments pixel-level signals across frames and scatters early pixel-word interactions. In parallel, native VLMs, despite impressive performance on single images, remain largely unexplored in multi-image, video understanding, and spatial intelligence. Hence, we introduce NEO-ov, a native foundation model that learns cross-frame and pixel-word correspondence end-to-end, without any external encoders, auxiliary adapters, or post-hoc fusion. By eliminating module boundaries entirely, NEO-ov enables fine-grained and unified spatiotemporal modeling to emerge natively inside the model. Notably, NEO-ov largely narrows the gap to modular counterparts while excelling at fine-grained visual perception, validating that native "one-vision" architectures are not only feasible but competitive at scale. Beyond empirical performance, we unveil systematic architectural analyses and detailed training recipes to facilitate subsequent native multimodal modeling. Our code and models are publicly available at: https://github.com/EvolvingLMMs-Lab/NEO.

cs.CV↗

arXiv2026

SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture

Recent large vision-language models (VLMs) remain fundamentally constrained by a persistent dichotomy: understanding and generation are treated as distinct problems, leading to fragmented architectures, cascaded pipelines, and misaligned representation spaces. We argue that this divide is not merely an engineering artifact, but a structural limitation that hinders the emergence of native multimodal intelligence. Hence, we introduce SenseNova-U1, a native unified multimodal paradigm built upon NEO-unify, in which understanding and generation evolve as synergistic views of a single underlying process. We launch two native unified variants, SenseNova-U1-8B-MoT and SenseNova-U1-A3B-MoT, built on dense (8B) and mixture-of-experts (30B-A3B) understanding baselines, respectively. Designed from first principles, they rival top-tier understanding-only VLMs across text understanding, vision-language perception, knowledge reasoning, agentic decision-making, and spatial intelligence. Meanwhile, they deliver strong semantic consistency and visual fidelity, excelling in conventional or knowledge-intensive any-to-image (X2I) synthesis, complex text-rich infographic generation, and interleaved vision-language generation, with or without think patterns. Beyond performance, we show detailed model design, data preprocessing, pre-/post-training, and inference strategies to support community research. Last but not least, preliminary evidence demonstrates that our models extend beyond perception and generation, performing strongly in vision-language-action (VLA) and world model (WM) scenarios. This points toward a broader roadmap where models do not translate between modalities, but think and act across them in a native manner. Multimodal AI is no longer about connecting separate systems, but about building a unified one and trusting the necessary capabilities to emerge from within.

cs.CV↗

arXiv2026

Test of a 34 GHz EOM laser frequency comb at ESPRESSO

Laser frequency combs (LFCs) are a promising technology for wavelength calibration of astronomical high-resolution spectrographs requiring utmost accuracy and stability, since they directly translate the fundamental SI time standard from the radio frequency regime to optical frequencies. However, they have so far seen limited use in practice, due to their complexity, incomplete wavelength coverage, but also the challenges in the data analysis they imply. Here, we present a detailed test of a 34 GHz electro-optic modulation comb with the ESPRESSO spectrograph. Using thin-film lithum-niobate waveguides for broadening and harmonic generation, the setup provides partial coverage of the IR, visible, and near-UV spectral ranges. We focus on assessing the quality of the delivered spectra and their capability to facilitate accurate and stable wavelength calibration. We present a detailed analysis of the spectrally-diffuse background, the line width, and characterize the line-spread function over a broader width than possible with the ESPRESSO facility LFC. Comparing both combs, we find strong local discrepancies in the wavelength calibration accuracy up to 15m/s , which correlate with the echellogram structure. These do not originate from the lasers, but from misalignments in the ESPRESSO calibration unit, highlighting the strong need to make instrument fiber feeds more robust to light-injection effects. Nevertheless, we demonstrate excellent stability of the wavelength calibration, with a scatter of only 17cm/s . This, however, can only be achieved when accurately modeling the non-Gaussian line-spread function, showcasing the need for advanced data analysis techniques when dealing with LFC spectra.

astro-ph.IM↗

arXiv2025

The Prism Hypothesis: Harmonizing Semantic and Pixel Representations via Unified Autoencoding

Deep representations across modalities are inherently intertwined. In this paper, we systematically analyze the spectral characteristics of various semantic and pixel encoders. Interestingly, our study uncovers a highly inspiring and rarely explored correspondence between an encoder's feature spectrum and its functional role: semantic encoders primarily capture low-frequency components that encode abstract meaning, whereas pixel encoders additionally retain high-frequency information that conveys fine-grained detail. This heuristic finding offers a unifying perspective that ties encoder behavior to its underlying spectral structure. We define it as the Prism Hypothesis, where each data modality can be viewed as a projection of the natural world onto a shared feature spectrum, just like the prism. Building on this insight, we propose Unified Autoencoding (UAE), a model that harmonizes semantic structure and pixel details via an innovative frequency-band modulator, enabling their seamless coexistence. Extensive experiments demonstrate that UAE effectively unifies semantic abstraction and pixel-level fidelity within a single latent space, achieving state-of-the-art performance. Moreover, we show that UAE can be directly applied to pixel-space modeling, significantly improving both FID and IS over the vanilla JIT baseline. Our code is avaliable at: https://github.com/WeichenFan/UAE.

cs.CV↗

arXiv2025

Mid-infrared continua via spectral broadening and difference frequency generation in a nanophotonic lithium niobate waveguide

Periodically poled thin film lithium niobate waveguides provide simultaneous access to efficient second and third order nonlinear processes, enabling broadband generation of coherent laser light. Here, we demonstrate the generation of a broadband mid-infrared continuum in a nanophotonic lithium niobate waveguide pumped by a telecom-wavelength femtosecond source. Specifically, our dual-stage design includes both third-order nonlinear spectral broadening followed by a dedicated periodically poled waveguide section performing efficient broadband intrapulse difference frequency generation. Driven by sub-100 fs pulses with approximately 200 pJ pulse energy, the generated mid-infrared light covers wavelengths from 3200 to 4800 nm. Cascaded harmonic generation also extends the spectrum into the visible and ultraviolet domains, resulting in an overall spectral bandwidth ranging from 350 to 4800 nm.

physics.optics↗

arXiv2025

CFG-Zero*: Improved Classifier-Free Guidance for Flow Matching Models

Classifier-Free Guidance (CFG) is a widely adopted technique in diffusion/flow models to improve image fidelity and controllability. In this work, we first analytically study the effect of CFG on flow matching models trained on Gaussian mixtures where the ground-truth flow can be derived. We observe that in the early stages of training, when the flow estimation is inaccurate, CFG directs samples toward incorrect trajectories. Building on this observation, we propose CFG-Zero*, an improved CFG with two contributions: (a) optimized scale, where a scalar is optimized to correct for the inaccuracies in the estimated velocity, hence the * in the name; and (b) zero-init, which involves zeroing out the first few steps of the ODE solver. Experiments on both text-to-image (Lumina-Next, Stable Diffusion 3, and Flux) and text-to-video (Wan-2.1) generation demonstrate that CFG-Zero* consistently outperforms CFG, highlighting its effectiveness in guiding Flow Matching models. (Code is available at github.com/WeichenFan/CFG-Zero-star)

cs.CV↗

arXiv2025

Spectral dynamics in broadband frequency combs with overlapping harmonics

Optical frequency combs and their spectra of evenly spaced discrete laser lines are essential to modern time and frequency metrology. Recent advances in integrated photonic waveguides enable efficient nonlinear broadening of an initially narrowband frequency comb to multi-octave bandwidth. Here, we study the nonlinear dynamics in the generation of such ultra-broadband spectra where different harmonics of the comb can overlap. We show that a set of interleaved combs with different offset frequencies extending across the entire spectrum can emerge, which transform into a single evenly spaced ultra-broadband frequency comb when the initial comb is offset-free.

physics.optics↗

arXiv2025

RepVideo: Rethinking Cross-Layer Representation for Video Generation

Video generation has achieved remarkable progress with the introduction of diffusion models, which have significantly improved the quality of generated videos. However, recent research has primarily focused on scaling up model training, while offering limited insights into the direct impact of representations on the video generation process. In this paper, we initially investigate the characteristics of features in intermediate layers, finding substantial variations in attention maps across different layers. These variations lead to unstable semantic representations and contribute to cumulative differences between features, which ultimately reduce the similarity between adjacent frames and negatively affect temporal coherence. To address this, we propose RepVideo, an enhanced representation framework for text-to-video diffusion models. By accumulating features from neighboring layers to form enriched representations, this approach captures more stable semantic information. These enhanced representations are then used as inputs to the attention mechanism, thereby improving semantic expressiveness while ensuring feature consistency across adjacent frames. Extensive experiments demonstrate that our RepVideo not only significantly enhances the ability to generate accurate spatial appearances, such as capturing complex spatial relationships between multiple objects, but also improves temporal consistency in video generation.

cs.CV↗

arXiv2025

Vchitect-2.0: Parallel Transformer for Scaling Up Video Diffusion Models

We present Vchitect-2.0, a parallel transformer architecture designed to scale up video diffusion models for large-scale text-to-video generation. The overall Vchitect-2.0 system has several key designs. (1) By introducing a novel Multimodal Diffusion Block, our approach achieves consistent alignment between text descriptions and generated video frames, while maintaining temporal coherence across sequences. (2) To overcome memory and computational bottlenecks, we propose a Memory-efficient Training framework that incorporates hybrid parallelism and other memory reduction techniques, enabling efficient training of long video sequences on distributed systems. (3) Additionally, our enhanced data processing pipeline ensures the creation of Vchitect T2V DataVerse, a high-quality million-scale training dataset through rigorous annotation and aesthetic evaluation. Extensive benchmarking demonstrates that Vchitect-2.0 outperforms existing methods in video quality, training efficiency, and scalability, serving as a suitable base for high-fidelity video generation.

cs.CV↗

arXiv2024

Supercontinua from integrated gallium nitride waveguides

Supercontinua are broadband spectra that are essential to optical spectroscopy, sensing, imaging, and metrology. They are generated from ultrashort laser pulses through nonlinear frequency conversion in fibers, bulk media, and chip-integrated waveguides. For any generating platform, balancing the competing criteria of strong nonlinearity, transparency, and absence of multiphoton absorption is a key challenge. Here, we explore supercontinuum generation in integrated gallium nitride (GaN) waveguides, which combine a high Kerr-nonlinearity, mid-infrared transparency, and a large bandgap that prevents two- and three-photon absorption in the technologically important telecom C-band, where compact erbium-based pump lasers exist. Using this type of laser, we demonstrate tunable dispersive waves and gap-free spectra extending to almost 4 micron in wavelength, relevant to functional group chemical sensing. Additionally, leveraging the material's second-order nonlinearity, we implement on-chip f-2f interferometry to detect the pump laser's carrier-envelope offset frequency, which enables precision metrology. These results demonstrate the versatility of GaN-on-sapphire as a new platform for broadband nonlinear photonics.

physics.optics↗

arXiv2023

Link-Context Learning for Multimodal LLMs

The ability to learn from context with novel concepts, and deliver appropriate responses are essential in human conversations. Despite current Multimodal Large Language Models (MLLMs) and Large Language Models (LLMs) being trained on mega-scale datasets, recognizing unseen images or understanding novel concepts in a training-free manner remains a challenge. In-Context Learning (ICL) explores training-free few-shot learning, where models are encouraged to ``learn to learn" from limited tasks and generalize to unseen tasks. In this work, we propose link-context learning (LCL), which emphasizes "reasoning from cause and effect" to augment the learning capabilities of MLLMs. LCL goes beyond traditional ICL by explicitly strengthening the causal relationship between the support set and the query set. By providing demonstrations with causal links, LCL guides the model to discern not only the analogy but also the underlying causal associations between data points, which empowers MLLMs to recognize unseen images and understand novel concepts more effectively. To facilitate the evaluation of this novel approach, we introduce the ISEKAI dataset, comprising exclusively of unseen generated image-label pairs designed for link-context learning. Extensive experiments show that our LCL-MLLM exhibits strong link-context learning capabilities to novel concepts over vanilla MLLMs. Code and data will be released at https://github.com/isekai-portal/Link-Context-Learning.

cs.CV↗

arXiv2023

Hierarchy Flow For High-Fidelity Image-to-Image Translation

Image-to-image (I2I) translation comprises a wide spectrum of tasks. Here we divide this problem into three levels: strong-fidelity translation, normal-fidelity translation, and weak-fidelity translation, indicating the extent to which the content of the original image is preserved. Although existing methods achieve good performance in weak-fidelity translation, they fail to fully preserve the content in both strong- and normal-fidelity tasks, e.g. sim2real, style transfer and low-level vision. In this work, we propose Hierarchy Flow, a novel flow-based model to achieve better content preservation during translation. Specifically, 1) we first unveil the drawbacks of standard flow-based models when applied to I2I translation. 2) Next, we propose a new design, namely hierarchical coupling for reversible feature transformation and multi-scale modeling, to constitute Hierarchy Flow. 3) Finally, we present a dedicated aligned-style loss for a better trade-off between content preservation and stylization during translation. Extensive experiments on a wide range of I2I translation benchmarks demonstrate that our approach achieves state-of-the-art performance, with convincing advantages in both strong- and normal-fidelity tasks. Code and models will be at https://github.com/WeichenFan/HierarchyFlow.

cs.CV↗

arXiv2022

StyleFlow For Content-Fixed Image to Image Translation

Image-to-image (I2I) translation is a challenging topic in computer vision. We divide this problem into three tasks: strongly constrained translation, normally constrained translation, and weakly constrained translation. The constraint here indicates the extent to which the content or semantic information in the original image is preserved. Although previous approaches have achieved good performance in weakly constrained tasks, they failed to fully preserve the content in both strongly and normally constrained tasks, including photo-realism synthesis, style transfer, and colorization, etc. To achieve content-preserving transfer in strongly constrained and normally constrained tasks, we propose StyleFlow, a new I2I translation model that consists of normalizing flows and a novel Style-Aware Normalization (SAN) module. With the invertible network structure, StyleFlow first projects input images into deep feature space in the forward pass, while the backward pass utilizes the SAN module to perform content-fixed feature transformation and then projects back to image space. Our model supports both image-guided translation and multi-modal synthesis. We evaluate our model in several I2I translation benchmarks, and the results show that the proposed model has advantages over previous methods in both strongly constrained and normally constrained tasks.

cs.CV↗

arXiv2022

InvNorm: Domain Generalization for Object Detection in Gastrointestinal Endoscopy

Domain Generalization is a challenging topic in computer vision, especially in Gastrointestinal Endoscopy image analysis. Due to several device limitations and ethical reasons, current open-source datasets are typically collected on a limited number of patients using the same brand of sensors. Different brands of devices and individual differences will significantly affect the model's generalizability. Therefore, to address the generalization problem in GI(Gastrointestinal) endoscopy, we propose a multi-domain GI dataset and a light, plug-in block called InvNorm(Invertible Normalization), which could achieve a better generalization performance in any structure. Previous DG(Domain Generalization) methods fail to achieve invertible transformation, which would lead to some misleading augmentation. Moreover, these models would be more likely to lead to medical ethics issues. Our method utilizes normalizing flow to achieve invertible and explainable style normalization to address the problem. The effectiveness of InvNorm is demonstrated on a wide range of tasks, including GI recognition, GI object detection, and natural image recognition.

eess.IV↗