Ziyan Yang, Kushal Kafle, Franck Dernoncourt, Vicente Ordonez
We propose a margin-based loss for tuning joint vision-language models so that their gradient-based explanations are consistent with region-level annotations provided by humans for relatively smaller grounding datasets. We refer to this objective as Attention Mask Consistency (AMC) and demonstrate that it produces superior visual grounding results than previous methods that rely on using vision-language models to score the outputs of object detectors. Particularly, a model trained with AMC on top of standard vision-language modeling objectives obtains a state-of-the-art accuracy of 86.49% in the Flickr30k visual grounding benchmark, an absolute improvement of 5.38% when compared to the best previous model trained under the same level of supervision. Our approach also performs exceedingly well on established benchmarks for referring expression comprehension where it obtains 80.34% accuracy in the easy test of RefCOCO+, and 64.55% in the difficult split. AMC is effective, easy to implement, and is general as it can be adopted by any vision-language model, and can use any type of region annotations.
Ziyan Yang, Kushal Kafle, Zhe Lin, Scott Cohen, Zhihong Ding, Vicente Ordonez
We propose Subject-Conditional Relation Detection SCoRD, where conditioned on an input subject, the goal is to predict all its relations to other objects in a scene along with their locations. Based on the Open Images dataset, we propose a challenging OIv6-SCoRD benchmark such that the training and testing splits have a distribution shift in terms of the occurrence statistics of $\langle$subject, relation, object$\rangle$ triplets. To solve this problem, we propose an auto-regressive model that given a subject, it predicts its relations, objects, and object locations by casting this output as a sequence of tokens. First, we show that previous scene-graph prediction methods fail to produce as exhaustive an enumeration of relation-object pairs when conditioned on a subject on this benchmark. Particularly, we obtain a recall@3 of 83.8% for our relation-object predictions compared to the 49.75% obtained by a recent scene graph detector. Then, we show improved generalization on both relation-object and object-box predictions by leveraging during training relation-object pairs obtained automatically from textual captions and for which no object-box annotations are available. Particularly, for $\langle$subject, relation, object$\rangle$ triplets for which no object locations are available during training, we are able to obtain a recall@3 of 33.80% for relation-object pairs and 26.75% for their box locations.
Leticia Pinto-Alva, Ian K. Torres, Rosangel Garcia, Ziyan Yang, Vicente Ordonez
Over the years, datasets and benchmarks have had an outsized influence on the design of novel algorithms. In this paper, we introduce ChairSegments, a novel and compact semi-synthetic dataset for object segmentation. We also show empirical findings in transfer learning that mirror recent findings for image classification. We particularly show that models that are fine-tuned from a pretrained set of weights lie in the same basin of the optimization landscape. ChairSegments consists of a diverse set of prototypical images of chairs with transparent backgrounds composited into a diverse array of backgrounds. We aim for ChairSegments to be the equivalent of the CIFAR-10 dataset but for quickly designing and iterating over novel model architectures for segmentation. On Chair Segments, a U-Net model can be trained to full convergence in only thirty minutes using a single GPU. Finally, while this dataset is semi-synthetic, it can be a useful proxy for real data, leading to state-of-the-art accuracy on the Object Discovery dataset when used as a source of pretraining.
Jinghui Chen, Dongruo Zhou, Yiqi Tang, Ziyan Yang, Yuan Cao, Quanquan Gu
Adaptive gradient methods, which adopt historical gradient information to automatically adjust the learning rate, despite the nice property of fast convergence, have been observed to generalize worse than stochastic gradient descent (SGD) with momentum in training deep neural networks. This leaves how to close the generalization gap of adaptive gradient methods an open problem. In this work, we show that adaptive gradient methods such as Adam, Amsgrad, are sometimes "over adapted". We design a new algorithm, called Partially adaptive momentum estimation method, which unifies the Adam/Amsgrad with SGD by introducing a partial adaptive parameter $p$, to achieve the best from both worlds. We also prove the convergence rate of our proposed algorithm to a stationary point in the stochastic nonconvex optimization setting. Experiments on standard benchmarks show that our proposed algorithm can maintain a fast convergence rate as Adam/Amsgrad while generalizing as well as SGD in training deep neural networks. These results would suggest practitioners pick up adaptive gradient methods once again for faster training of deep neural networks.
Qi Zhao, Xingyu Ni, Ziyu Wang, Feng Cheng, Ziyan Yang, Lu Jiang, Bohan Wang
We investigate how to enhance the physical fidelity of video generation models by leveraging synthetic videos derived from computer graphics pipelines. These rendered videos respect real-world physics, such as maintaining 3D consistency, and serve as a valuable resource that can potentially improve video generation models. To harness this potential, we propose a solution that curates and integrates synthetic data while introducing a method to transfer its physical realism to the model, significantly reducing unwanted artifacts. Through experiments on three representative tasks emphasizing physical consistency, we demonstrate its efficacy in enhancing physical fidelity. While our model still lacks a deep understanding of physics, our work offers one of the first empirical demonstrations that synthetic video enhances physical fidelity in video synthesis. Website: https://kevinz8866.github.io/simulation/
Jaywon Koo, Jefferson Hernandez, Moayed Haji-Ali, Ziyan Yang, Vicente Ordonez
Evaluating text-to-image and text-to-video models is challenging due to a fundamental disconnect: established metrics fail to jointly measure visual quality and semantic alignment with text, leading to a poor correlation with human judgments. To address this critical issue, we propose cFreD, a general metric based on a Conditional Fréchet Distance that unifies the assessment of visual fidelity and text-prompt consistency into a single score. Existing metrics such as Fréchet Inception Distance (FID) capture image quality but ignore text conditioning while alignment scores such as CLIPScore are insensitive to visual quality. Furthermore, learned preference models require constant retraining and are unlikely to generalize to novel architectures or out-of-distribution prompts. Through extensive experiments across multiple recently proposed text-to-image models and diverse prompt datasets, cFreD exhibits a higher correlation with human judgments compared to statistical metrics , including metrics trained with human preferences. Our findings validate cFreD as a robust, future-proof metric for the systematic evaluation of text conditioned models, standardizing benchmarking in this rapidly evolving field. We release our evaluation toolkit and benchmark.
Team Seawead, Ceyuan Yang, Zhijie Lin, Yang Zhao, Shanchuan Lin, Zhibei Ma, Haoyuan Guo, Hao Chen, Lu Qi, Sen Wang, Feng Cheng, Feilong Zuo, Xuejiao Zeng, Ziyan Yang, Fangyuan Kong, Meng Wei, Zhiwu Qing, Fei Xiao, Tuyen Hoang, Siyu Zhang, Peihao Zhu, Qi Zhao, Jiangqiao Yan, Liangke Gui, Sheng Bi, Jiashi Li, Yuxi Ren, Rui Wang, Huixia Li, Xuefeng Xiao, Shu Liu, Feng Ling, Heng Zhang, Houmin Wei, Huafeng Kuang, Jerry Duncan, Junda Zhang, Junru Zheng, Li Sun, Manlin Zhang, Renfei Sun, Xiaobin Zhuang, Xiaojie Li, Xin Xia, Xuyan Chi, Yanghua Peng, Yuping Wang, Yuxuan Wang, Zhongkai Zhao, Zhuo Chen, Zuquan Song, Zhenheng Yang, Jiashi Feng, Jianchao Yang, Lu Jiang
This technical report presents a cost-efficient strategy for training a video generation foundation model. We present a mid-sized research model with approximately 7 billion parameters (7B) called Seaweed-7B trained from scratch using 665,000 H100 GPU hours. Despite being trained with moderate computational resources, Seaweed-7B demonstrates highly competitive performance compared to contemporary video generation models of much larger size. Design choices are especially crucial in a resource-constrained setting. This technical report highlights the key design decisions that enhance the performance of the medium-sized diffusion model. Empirically, we make two observations: (1) Seaweed-7B achieves performance comparable to, or even surpasses, larger models trained on substantially greater GPU resources, and (2) our model, which exhibits strong generalization ability, can be effectively adapted across a wide range of downstream applications either by lightweight fine-tuning or continue training. See the project page at https://seaweed.video/
Yuang Ai, Jiaming Han, Shaobin Zhuang, Weijia Mao, Xuefeng Hu, Ziyan Yang, Zhenheng Yang, Yali Wang, Huaibo Huang, Xiangyu Yue, Hao Chen
We present BitDance, a scalable autoregressive (AR) image generator that predicts binary visual tokens instead of codebook indices. With high-entropy binary latents, BitDance lets each token represent up to $2^{256}$ states, yielding a compact yet highly expressive discrete representation. Sampling from such a huge token space is difficult with standard classification. To resolve this, BitDance uses a binary diffusion head: instead of predicting an index with softmax, it employs continuous-space diffusion to generate the binary tokens. Furthermore, we propose next-patch diffusion, a new decoding method that predicts multiple tokens in parallel with high accuracy, greatly speeding up inference. On ImageNet 256x256, BitDance achieves an FID of 1.24, the best among AR models. With next-patch diffusion, BitDance beats state-of-the-art parallel AR models that use 1.4B parameters, while using 5.4x fewer parameters (260M) and achieving 8.7x speedup. For text-to-image generation, BitDance trains on large-scale multimodal tokens and generates high-resolution, photorealistic images efficiently, showing strong performance and favorable scaling. When generating 1024x1024 images, BitDance achieves a speedup of over 30x compared to prior AR models. We release code and models to facilitate further research on AR foundation models. Code and models are available at: https://github.com/shallowdream204/BitDance.
Dongruo Zhou, Jinghui Chen, Yuan Cao, Ziyan Yang, Quanquan Gu
Adaptive gradient methods are workhorses in deep learning. However, the convergence guarantees of adaptive gradient methods for nonconvex optimization have not been thoroughly studied. In this paper, we provide a fine-grained convergence analysis for a general class of adaptive gradient methods including AMSGrad, RMSProp and AdaGrad. For smooth nonconvex functions, we prove that adaptive gradient methods in expectation converge to a first-order stationary point. Our convergence rate is better than existing results for adaptive gradient methods in terms of dimension. In addition, we also prove high probability bounds on the convergence rates of AMSGrad, RMSProp as well as AdaGrad, which have not been established before. Our analyses shed light on better understanding the mechanism behind adaptive gradient methods in optimizing nonconvex objectives.
Ruozhen He, Paola Cascante-Bonilla, Ziyan Yang, Alexander C. Berg, Vicente Ordonez
Vision-and-language models trained to match images with text can be combined with visual explanation methods to point to the locations of specific objects in an image. Our work shows that the localization --"grounding"-- abilities of these models can be further improved by finetuning for self-consistent visual explanations. We propose a strategy for augmenting existing text-image datasets with paraphrases using a large language model, and SelfEQ, a weakly-supervised strategy on visual explanation maps for paraphrases that encourages self-consistency. Specifically, for an input textual phrase, we attempt to generate a paraphrase and finetune the model so that the phrase and paraphrase map to the same region in the image. We posit that this both expands the vocabulary that the model is able to handle, and improves the quality of the object locations highlighted by gradient-based visual explanation methods (e.g. GradCAM). We demonstrate that SelfEQ improves performance on Flickr30k, ReferIt, and RefCOCO+ over a strong baseline method and several prior works. Particularly, comparing to other methods that do not use any type of box annotations, we obtain 84.07% on Flickr30k (an absolute improvement of 4.69%), 67.40% on ReferIt (an absolute improvement of 7.68%), and 75.10%, 55.49% on RefCOCO+ test sets A and B respectively (an absolute improvement of 3.74% on average).
Ruozhen He, Ziyan Yang, Paola Cascante-Bonilla, Alexander C. Berg, Vicente Ordonez
This paper extensively investigates the effectiveness of synthetic training data to improve the capabilities of vision-and-language models for grounding textual descriptions to image regions. We explore various strategies to best generate image-text pairs and image-text-box triplets using a series of pretrained models under different settings and varying degrees of reliance on real data. Through comparative analyses with synthetic, real, and web-crawled data, we identify factors that contribute to performance differences, and propose SynGround, an effective pipeline for generating useful synthetic data for visual grounding. Our findings show that SynGround can improve the localization capabilities of off-the-shelf vision-and-language models and offers the potential for arbitrarily large scale data generation. Particularly, data generated with SynGround improves the pointing game accuracy of a pretrained ALBEF and BLIP models by 4.81% and 17.11% absolute percentage points, respectively, across the RefCOCO+ and the Flickr30k benchmarks.
Ruozhen He, Moayed Haji-Ali, Ziyan Yang, Vicente Ordonez
Text-to-image diffusion models trained on a fixed set of resolutions often fail to generalize, even when asked to generate images at lower resolutions than those seen during training. High-resolution text-to-image generators are currently unable to easily offer an out-of-the-box budget-efficient alternative to their users who might not need high-resolution images. We identify a key technical insight in diffusion models that when addressed can help tackle this limitation: Noise schedulers have unequal perceptual effects across resolutions. The same level of noise removes disproportionately more signal from lower-resolution images than from high-resolution images, leading to a train-test mismatch. We propose NoiseShift, a training-free method that recalibrates the noise level of the denoiser conditioned on resolution size. NoiseShift requires no changes to model architecture or sampling schedule and is compatible with existing models. When applied to Stable Diffusion 3, Stable Diffusion 3.5, and Flux-Dev, quality at low resolutions is significantly improved. On LAION-COCO, NoiseShift improves SD3.5 by 15.89%, SD3 by 8.56%, and Flux-Dev by 2.44% in FID on average. On CelebA, NoiseShift improves SD3.5 by 10.36%, SD3 by 5.19%, and Flux-Dev by 3.02% in FID on average. These results demonstrate the effectiveness of NoiseShift in mitigating resolution-dependent artifacts and enhancing the quality of low-resolution image generation.
Team Seedance, Heyi Chen, Siyan Chen, Xin Chen, Yanfei Chen, Ying Chen, Zhuo Chen, Feng Cheng, Tianheng Cheng, Xinqi Cheng, Xuyan Chi, Jian Cong, Jing Cui, Qinpeng Cui, Qide Dong, Junliang Fan, Jing Fang, Zetao Fang, Chengjian Feng, Han Feng, Mingyuan Gao, Yu Gao, Dong Guo, Qiushan Guo, Boyang Hao, Qingkai Hao, Bibo He, Qian He, Tuyen Hoang, Ruoqing Hu, Xi Hu, Weilin Huang, Zhaoyang Huang, Zhongyi Huang, Donglei Ji, Siqi Jiang, Wei Jiang, Yunpu Jiang, Zhuo Jiang, Ashley Kim, Jianan Kong, Zhichao Lai, Shanshan Lao, Yichong Leng, Ai Li, Feiya Li, Gen Li, Huixia Li, JiaShi Li, Liang Li, Ming Li, Shanshan Li, Tao Li, Xian Li, Xiaojie Li, Xiaoyang Li, Xingxing Li, Yameng Li, Yifu Li, Yiying Li, Chao Liang, Han Liang, Jianzhong Liang, Ying Liang, Zhiqiang Liang, Wang Liao, Yalin Liao, Heng Lin, Kengyu Lin, Shanchuan Lin, Xi Lin, Zhijie Lin, Feng Ling, Fangfang Liu, Gaohong Liu, Jiawei Liu, Jie Liu, Jihao Liu, Shouda Liu, Shu Liu, Sichao Liu, Songwei Liu, Xin Liu, Xue Liu, Yibo Liu, Zikun Liu, Zuxi Liu, Junlin Lyu, Lecheng Lyu, Qian Lyu, Han Mu, Xiaonan Nie, Jingzhe Ning, Xitong Pan, Yanghua Peng, Lianke Qin, Xueqiong Qu, Yuxi Ren, Kai Shen, Guang Shi, Lei Shi, Yan Song, Yinglong Song, Fan Sun, Li Sun, Renfei Sun, Yan Sun, Zeyu Sun, Wenjing Tang, Yaxue Tang, Zirui Tao, Feng Wang, Furui Wang, Jinran Wang, Junkai Wang, Ke Wang, Kexin Wang, Qingyi Wang, Rui Wang, Sen Wang, Shuai Wang, Tingru Wang, Weichen Wang, Xin Wang, Yanhui Wang, Yue Wang, Yuping Wang, Yuxuan Wang, Ziyu Wang, Guoqiang Wei, Wanru Wei, Di Wu, Guohong Wu, Hanjie Wu, Jian Wu, Jie Wu, Ruolan Wu, Xinglong Wu, Yonghui Wu, Ruiqi Xia, Liang Xiang, Fei Xiao, XueFeng Xiao, Pan Xie, Shuangyi Xie, Shuang Xu, Jinlan Xue, Shen Yan, Bangbang Yang, Ceyuan Yang, Jiaqi Yang, Runkai Yang, Tao Yang, Yang Yang, Yihang Yang, ZhiXian Yang, Ziyan Yang, Songting Yao, Yifan Yao, Zilyu Ye, Bowen Yu, Jian Yu, Chujie Yuan, Linxiao Yuan, Sichun Zeng, Weihong Zeng, Xuejiao Zeng, Yan Zeng, Chuntao Zhang, Heng Zhang, Jingjie Zhang, Kuo Zhang, Liang Zhang, Liying Zhang, Manlin Zhang, Ting Zhang, Weida Zhang, Xiaohe Zhang, Xinyan Zhang, Yan Zhang, Yuan Zhang, Zixiang Zhang, Fengxuan Zhao, Huating Zhao, Yang Zhao, Hao Zheng, Jianbin Zheng, Xiaozheng Zheng, Yangyang Zheng, Yijie Zheng, Jiexin Zhou, Jiahui Zhu, Kuan Zhu, Shenhan Zhu, Wenjia Zhu, Benhui Zou, Feilong Zuo
Hanyu Wang, Jiaming Han, Ziyan Yang, Qi Zhao, Shanchuan Lin, Xiangyu Yue, Abhinav Shrivastava, Zhenheng Yang, Hao Chen
Multimodal large language models (MLLMs) extend the success of language models to visual understanding, and recent efforts have sought to build unified MLLMs that support both understanding and generation. However, constructing such models remains challenging: hybrid approaches combine continuous embeddings with diffusion or flow-based objectives, producing high-quality images but breaking the autoregressive paradigm, while pure autoregressive approaches unify text and image prediction over discrete visual tokens but often face trade-offs between semantic alignment and pixel-level fidelity. In this work, we present Bridge, a pure autoregressive unified MLLM that augments pre-trained visual understanding models with generative ability through a Mixture-of-Transformers architecture, enabling both image understanding and generation within a single next-token prediction framework. To further improve visual generation fidelity, we propose a semantic-to-pixel discrete representation that integrates compact semantic tokens with fine-grained pixel tokens, achieving strong language alignment and precise description of visual details with only a 7.9% increase in sequence length. Extensive experiments across diverse multimodal benchmarks demonstrate that Bridge achieves competitive or superior results in both understanding and generation benchmarks, while requiring less training data and reduced training time compared to prior unified MLLMs.
Weichen Yu, Ziyan Yang, Shanchuan Lin, Qi Zhao, Jianyi Wang, Liangke Gui, Matt Fredrikson, Lu Jiang
In text-to-image (T2I) generation, a prevalent training technique involves utilizing Vision Language Models (VLMs) for image re-captioning. Even though VLMs are known to exhibit hallucination, generating descriptive content that deviates from the visual reality, the ramifications of such caption hallucinations on T2I generation performance remain under-explored. Through our empirical investigation, we first establish a comprehensive dataset comprising VLM-generated captions, and then systematically analyze how caption hallucination influences generation outcomes. Our findings reveal that (1) the disparities in caption quality persistently impact model outputs during fine-tuning. (2) VLMs confidence scores serve as reliable indicators for detecting and characterizing noise-related patterns in the data distribution. (3) even subtle variations in caption fidelity have significant effects on the quality of learned representations. These findings collectively emphasize the profound impact of caption quality on model performance and highlight the need for more sophisticated robust training algorithm in T2I. In response to these observations, we propose a approach leveraging VLM confidence score to mitigate caption noise, thereby enhancing the robustness of T2I models against hallucination in caption.
Jiaming Han, Hao Chen, Yang Zhao, Hanyu Wang, Qi Zhao, Ziyan Yang, Hao He, Xiangyu Yue, Lu Jiang
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
Leigang Qu, Feng Cheng, Ziyan Yang, Qi Zhao, Shanchuan Lin, Yichun Shi, Yicong Li, Wenjie Wang, Tat-Seng Chua, Lu Jiang
In-context image editing aims to modify images based on a contextual sequence comprising text and previously generated images. Existing methods typically depend on task-specific pipelines and expert models (e.g., segmentation and inpainting) to curate training data. In this work, we explore whether an in-context image editing model can be learned directly from videos. We introduce a scalable approach to annotate videos as interleaved multimodal sequences. To effectively learn from this data, we design a block-causal diffusion transformer trained on three proxy tasks: next-image prediction, current segmentation prediction, and next-segmentation prediction. Additionally, we propose a novel multi-turn image editing benchmark to advance research in this area. Extensive experiments demonstrate that our model exhibits strong in-context image editing capabilities and achieves state-of-the-art results on two multi-turn image editing benchmarks. Despite being trained exclusively on videos, our model also shows promising abilities in multi-concept composition, story generation, and chain-of-editing applications.
Yuwei Guo, Ceyuan Yang, Ziyan Yang, Zhibei Ma, Zhijie Lin, Zhenheng Yang, Dahua Lin, Lu Jiang
Recent advances in video generation can produce realistic, minute-long single-shot videos with scalable diffusion transformers. However, real-world narrative videos require multi-shot scenes with visual and dynamic consistency across shots. In this work, we introduce Long Context Tuning (LCT), a training paradigm that expands the context window of pre-trained single-shot video diffusion models to learn scene-level consistency directly from data. Our method expands full attention mechanisms from individual shots to encompass all shots within a scene, incorporating interleaved 3D position embedding and an asynchronous noise strategy, enabling both joint and auto-regressive shot generation without additional parameters. Models with bidirectional attention after LCT can further be fine-tuned with context-causal attention, facilitating auto-regressive generation with efficient KV-cache. Experiments demonstrate single-shot models after LCT can produce coherent multi-shot scenes and exhibit emerging capabilities, including compositional generation and interactive shot extension, paving the way for more practical visual content creation. See https://guoyww.github.io/projects/long-context-video/ for more details.
Jaywon Koo, Ziyan Yang, Paola Cascante-Bonilla, Baishakhi Ray, Vicente Ordonez
Visual Programming has recently emerged as an alternative to end-to-end black-box visual reasoning models. This type of method leverages Large Language Models (LLMs) to generate the source code for an executable computer program that solves a given problem. This strategy has the advantage of offering an interpretable reasoning path and does not require finetuning a model with task-specific data. We propose PropTest, a general strategy that improves visual programming by further using an LLM to generate code that tests for visual properties in an initial round of proposed solutions. Our method generates tests for data-type consistency, output syntax, and semantic properties. PropTest achieves comparable results to state-of-the-art methods while using publicly available LLMs. This is demonstrated across different benchmarks on visual question answering and referring expression comprehension. Particularly, PropTest improves ViperGPT by obtaining 46.1\% accuracy (+6.0\%) on GQA using Llama3-8B and 59.5\% (+8.1\%) on RefCOCO+ using CodeLlama-34B.
Shengqu Cai, Ceyuan Yang, Lvmin Zhang, Yuwei Guo, Junfei Xiao, Ziyan Yang, Yinghao Xu, Zhenheng Yang, Alan Yuille, Leonidas Guibas, Maneesh Agrawala, Lu Jiang, Gordon Wetzstein
Long video generation is fundamentally a long context memory problem: models must retain and retrieve salient events across a long range without collapsing or drifting. However, scaling diffusion transformers to generate long-context videos is fundamentally limited by the quadratic cost of self-attention, which makes memory and computation intractable and difficult to optimize for long sequences. We recast long-context video generation as an internal information retrieval task and propose a simple, learnable sparse attention routing module, Mixture of Contexts (MoC), as an effective long-term memory retrieval engine. In MoC, each query dynamically selects a few informative chunks plus mandatory anchors (caption, local windows) to attend to, with causal routing that prevents loop closures. As we scale the data and gradually sparsify the routing, the model allocates compute to salient history, preserving identities, actions, and scenes over minutes of content. Efficiency follows as a byproduct of retrieval (near-linear scaling), which enables practical training and synthesis, and the emergence of memory and consistency at the scale of minutes.