Research on learning to generate new data, from images and video to structured outputs. Explore modeling choices, evaluation methods, and practical tradeoffs.
63 papers · Latest edition October 1, 2026
Three of the latest briefs in this collection. Read the evidence and the original papers alongside them.
RULER uses instruction-specific visual rubrics as reinforcement-learning rewards to make SVG generation more faithful, stylish, and resistant to reward hacking.
A visual reward model that first decides what matters in each image-generation case, then scores outputs with detailed rubrics to provide better training signals.
WanPE turns ordinary text prompts into director-level cinematic plans, substantially improving the quality and consistency of long-form AI-generated videos.
Newest editions first.
Hangyu Ran, Yuhao Zheng, Yingying Zhang, Kevin Qinghong Lin, Han Peng
RULER uses instruction-specific visual rubrics as reinforcement-learning rewards to make SVG generation more faithful, stylish, and resistant to reward hacking.
Xuehai Bai, Zhenchen Tang, Yang Shi, Dianyi Wang, Tengfei Liu, Wanshun Su, Xuanyu Zhu, Ruohui Wang, Haiwen Diao, Haotian Wang, Xiaoling Gu, Yuanxing Zhang
A visual reward model that first decides what matters in each image-generation case, then scores outputs with detailed rubrics to provide better training signals.
Yubo Zhu, Yawen Shao, Ziyun Dai, Zixun Fang, Kai Zhu, Siyang Sun, Haolan Xue, Chuxin Wang, Tingyu Weng, Jingming Luo, Chen Shi, Lianghua Huang, Yufeng Ai, Yuzheng Wang, Wenyuan Zhang, Yu Shang, Yuxiang Bao, Zoubin Bi, Jie Xiao, Jinbo Xing, Jiaxing Zhao, Chongyang Zhong, Hengjian Chen, Chenwei Xie, Akide Liu, Zhehan Kan, Yu Liu, Wei Zhai, Sheng Zhong, Wei Tong
WanPE turns ordinary text prompts into director-level cinematic plans, substantially improving the quality and consistency of long-form AI-generated videos.
Jiawei Mao, Haoqin Tu, Hardy Chen, Yuhan Wang, Keyang Xu, Jieru Mei, Hongliang Fei, Ruogu Fang, Wei Shao, Cihang Xie, Yuyin Zhou
MovieGrid turns long videos into jointly modeled spatial grids, helping generative models produce more coherent stories with many connected shots.
Jiahao Chang, Dong Du, Wanhu Sun, Yujian Zheng, Chuanyu Pan, Bowen Zhao, Chongjie Ye, Yuanming Hu, Xiaoguang Han
SAM3D-Part lets users ask for specific components of a 3D object and generates complete, correctly placed meshes rather than decomposing the entire object.
Jintao Zhang, Kai Jiang, Jintao Chen, Xu Wang, Deyuan Liu, Jungang Li, Dechuang Chen, Ming Lin, Jingjiang Zhou, Haopeng Jin, Qi Jia, Xiaohang Wang, Yaole Wang, Zhanqiang Zhang, Ran Li, Zhengkun Huang, Shuyue Xiong, Yuji Wang, Zikun Dai, Hui He, Yang Luo, Mang Ning, Weiqi Feng, Chengyang Ye, Xinyue Lin, Min Zhao, Hongzhou Zhu, Hengkai Tan, Zeyuan Wang, Chendong Xiang, Kaiwen Zheng, Zhijie Deng, Fan Bao, Jianfei Chen, Jun Zhu
Vidu S2 brings real-time, interactive, editable, and potentially spatial AI-generated video to live applications.
Brian Nlong Zhao, Zhuoyang Pan, James M. Rehg, Jiajun Wu, Shangzhe Wu
Kirin turns large collections of animal videos into a multimodal system that can generate and animate realistic movements for diverse species.
Adheesh Sunil Juvekar, Onkar Kishor Susladkar, Kiet A. Nguyen, Muntasir Wahed, Nabeel Bashir, Xiaona Zhou, Tianjiao Yu, Vedant Shah, Ismini Lourentzou
EditVid is a training-free video editor that combines several clever mechanisms to preserve motion, identity, and locality across many kinds of edits.
TGR Team, Lei Cheng, Haonan Hu, Beibei Kong, Yudong Li, Zang Li, Yunsheng Pang, Hongyang Su, Jianchao Tu, Yunlong Wang, Bing Wen, Junzhang Zhu, Shaojie Zhu, Chengxiang Zhuo
Tencent’s TGR framework replaces fragmented recommendation pipelines with generative ranking, slate generation, and offline reasoning, delivering measurable gains at massive production scale.
Zhiyuan Li, Chi-Man Pun, Peng-Tao Jiang, Bo Li, Xiaodong Cun
EditaLive turns human video editing into a fast, live-streaming experience by adapting and distilling a pretrained animation model while preserving facial expressions and character appearance.
Guangting Zheng, Yiyuan Zhang, Tao Yang, Yunpeng Chen, Rui Zhu, Jiajun Deng, Yanyong Zhang
GenFirst stabilizes end-to-end latent generative models by letting generation shape the latent space before progressively adding reconstruction pressure.
Haiyang Xu, Zheng Ding, Zhuowen Tu
RECAP-Forcing helps long video generators remember newly appearing subjects and scenes instead of simply favoring the latest frames.
Yunze Tong, Mushui Liu, Canyu Zhao, Shiyi Zhang, Didi Zhu, Peng Zhang, Wanggui He, Jinlong Liu, Ying Chen, Hao Jiang, Pipei Huang, Bo Zheng
InfinityEdit lets video generators apply ongoing edits to an endless stream while maintaining temporal continuity and stability over time.
Jianjie Luo, Yiming Zhong, Haoming Shen, Yupeng Xiao, Zhenguo Yang
KeyID keeps video characters recognizable by drafting the action first and then repairing identity at strategically chosen keyframes.
Manwen Liao, Xinyu Lian, Jian Mao, Kaixu Chen, Li Luo, Jinghao Yan, Wanshui Gan, Qiao Yu, Weitian Zhang, Chunhua Shen, Guang Chen, Bo Dai, Xudong Xu, Zhaoyang Lyu
MegaParts uses compact discrete geometry tokens and long-context autoregressive modeling to generate highly complex 3D objects with hundreds of coherent parts.
Zihao Liu, Xiaolong Shen, Zhenglin Zhou, Ruijie Quan, Yi Yang
Latent-to-4D turns the hidden representations of video generators directly into stable dynamic 3D worlds without passing through RGB.
Kuan Xing, Ye Wang, Changyi Gan, Yuheng Li, Thao Nguyen, Yi Chang, Yilin Wang
Atelier aims to make artist-inspired image generation more faithful by planning around artistic intent while avoiding the visual shortcuts that models commonly associate with famous artists.
Minghui Guo, Shengqiong Wu, Hao Fei
V-RAE makes video generation more efficient and semantically meaningful by compressing videos into latents derived from pretrained vision models rather than pixel-focused autoencoders.
Smriti Joshi, Apostolia Tsirikoglou, Daniel M. Lang, Richard Osuala, Noah Márquez Varaa, Alejandro Guzman, Grzegorz Skorupko, Sebastian Ibarra Arregui, Lidia Garrucho, Akane Ohashi, Dimitra Ntoula, Eugen Divjak, Oğuz Lafcı, Jan C. Peeken, Julia A. Schnabel, Fredrik Strand, Oliver Diaz, Karim Lekadir
A single-pass generative model synthesizes realistic, time-resolved breast MRI contrast without gadolinium and shows promising robustness and clinical utility.
Songchun Zhang, Sitong Guo, Xianghao Kong, Pengwei Liu, Yuwei Guo, Lvmin Zhang, Anyi Rao
FlexComposer makes it easier to place still or moving objects into videos while preserving realistic motion, appearance, and user-defined trajectories.
Xiangbo Gao, Siyuan Yang, Ping He, Mingyang Wu, Yuheng Wu, Yushen Zuo, Jiongze Yu, Ryan Cui, Hongyuan Hua, Devin Ma, Xiao Jin, Yubo Yuan, Qing Yin, Jie Yang, Zhengzhong Tu
Visko Orbis aims to make long, high-quality video generation interactive in real time, letting users change the story or visual direction while the video is still being produced.
Alexi Gladstone, Heng Ji, Yilun Du
Explorative Modeling trains generative systems by trying multiple candidate outputs and learning from the best match, promising better scaling and much faster end-to-end generation.
Amber Yijia Zheng, Lu Liu, Raymond A. Yeh, Xi Yin
A controlled study using animated letters shows that balanced video data and accurate captions are crucial for training effective text-to-video models.
Haodong Li, Tianfei Ren, Xiaoxiao Ma, Chunmei Qing, Zhen Fang, Sipeng He, Ziyu Guo, Haoyu Wu, Juanxi Tian, Yihang Zou, Ruichuan An, Dongzhi Jiang, Boxue Yang, Ji Xie, Xu Huang, Wenhao Yan, Jialv Zou, Zhengrong Yue, Yaxin Luo, Xiaotong Li, Yuzhu Wang, Junyan Ye, Jinjing Zhao, Zehui Chen, Lin Chen, Renye Yan, Feng Zhao, Pheng-Ann Heng
VideoCoCo uses executable Blender code as chain-of-thought to plan physical dynamics before a generative video model turns them into realistic videos.
Tuhin Chakrabarty, Xinyue Liu, Jane C. Ginsburg, Paramveer Dhillon
AI-generated books may be lower quality, but their sheer volume is reshaping Amazon’s market and squeezing revenue from human-authored books.
Jin Xu, Kangdi Wang, Ruibin Yuan, Shun Lei, Xiong Wang, Xize Cheng, Xueyao Zhang, Yang Zhang, Yiheng Chen, Yongqi Wang, Yue Wang, Zhifang Guo, Zihan Liu, Zijian Lin, Dake Guo, Hangrui Hu, Lei Xie, Linhan Ma, Wei Xue, Wenxiang Guo, Xinfa Zhu, Xipin Wei, Yangze Li, Yuanjun Lv, Yuxuan Wang, Yunfei Chu, Zhiyong Wu
Qwen-Music combines melody planning, large-scale multilingual training, and waveform rendering to generate and reinterpret high-quality songs with vocals.
Zhihao Xie, Junfeng Wu, Xinting Hu, Junchao Huang, Li Jiang
VideoRAE turns video foundation-model features into efficient, semantically rich latent spaces that can accelerate both diffusion and autoregressive video generation.
Salah Eddine Bekhouche, Abdellah Zakaria Sellam, Fadi Dornaika, Abdenour Hadid
AffectFlow-DINO uses conditional rectified flow to represent ambiguity in facial emotion predictions while improving valence-arousal, expression, and Action Unit estimation.
Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert
This paper trains a 3D latent generative model to synthesize realistic chest CT scans with controllable clinical findings, then uses reinforcement learning to make the requested attributes appear more reliably.
Paul Engstler, Iro Laina, Christian Rupprecht, Andrea Vedaldi
SynCity 3000 turns image-to-3D generation into a scalable scene-level diffusion system that can build large, coherent 3D worlds from prompts and layouts.
Bin Lin, Zheyuan Liu, Chenguo Lin, Sixiang Chen, Yunyang Ge, Yunlong Lin, Jianwei Zhang, Miles Yang, Zhao Zhong, Liefeng Bo, Li Yuan
GEAR lets an image generator and its tokenizer learn together, making autoregressive image synthesis train faster and produce better visual representations.
Jiayi Xu, Di He, Guolin Ke
This paper makes pixel-by-pixel image generators train more like they infer, delivering a new state of the art without needing a separate tokenizer.
Chunshi Wang, Haohan Weng, Junliang Ye, Biwen Lei, Yang Li, Zibo Zhao, Zeqiang Lai, Kaiyi Zhang, Yunhan Yang, Zhuo Chen, Chunchao Guo, Yawei Luo
PolyFlow turns mesh generation into a parallel continuous process by embedding topology into a learnable state space, making artist-style 3D mesh synthesis faster and more controllable.
Wei Zhou, Xiongwei Zhu, Zelin Xu, Bo Dong, Lixue Gong, Yongyuan Liang, Meng Chu, Leigang Qu, Lingdong Kong, Wei Liu, Tat-Seng Chua
DanceOPD teaches a single image generator to juggle text-to-image creation, editing, and guidance signals by learning from its own rollout states, helping one model do many generation tasks without the usual capability conflicts.
Chuyang Zhao, Yifei Song, Hongfa Wang, Jianlong Yuan, Yuan Zhang, Siming Fu, Zhineng Chen, Huilin Deng, Haoyang Huang, Nan Duan
This paper makes generative models faster by training them in perceptual feature space, letting them produce good samples in just a few steps instead of dozens.
Lan Feng, Wuyang Li, Eloi Zablocki, Matthieu Cord, Alexandre Alahi
This paper shows how to train strong one-step image generators by matching feature distributions from multiple frozen encoders, achieving state-of-the-art quality and even compressing a four-step model into a single step.
Xin Jin, Huanqia Cai, Zhen Li, Zechao Zhan, Dengyang Jiang, Aiming Hao, Yuming Jiang, Chunle Guo, Peng Gao, Ming-Ming Cheng, Steven C. H. Hoi
This paper upgrades reward models for image generation by teaching them to reason about score distributions instead of predicting a single scalar, making them both more accurate and easier to deploy.
Shuai Yang, Bingjie Gao, Ziwei Liu, Jiaqi Wang, Dahua Lin, Tong Wu
PermaVid helps edited videos stay coherent over time by using separate memory for appearance and geometry so later generations don’t drift after changes.
Haobo Li, Yanhong Zeng, Yunhong Lu, Jiapeng Zhu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Yujun Shen, Zhipeng Zhang
AAD-1 improves one-step video generation by using a smarter, non-mirrored discriminator and a staged training recipe to reduce motion collapse and make outputs more realistic.
Jiale Xu, Wang Zhao, Ying Shan
MeshWeaver makes 3D mesh generation more efficient and accurate by weaving vertices through sparse voxel geometry instead of relying on long, brittle token sequences.
Yanming Zhang, Yihan Bian, Jingyuan Qi, Yuguang Yao, Lifu Huang, Tianyi Zhou
This paper teaches mask diffusion models to “think again” by revisiting and locally revising their own outputs across multiple turns, enabling stronger reasoning and refinement without starting over.
Daniel Gilo, Sven Elflein, Ido Sobol, Or Litany
FlowBender teaches generative models to look at their own mistakes and correct them on the fly, improving both constraint satisfaction and sample quality.
Zhida Zhang, Jie Ma, Zhan Peng, Haoxue Wu, Yang Han, Jun Liang, Jie Cao, Jing Li
SmartDirector lets video models generate more cinematic, better-paced stories by using multiple keyframes as narrative anchors across single-shot, multi-shot, and video-extension settings.
Mengyu Ye, Keito Kudo, Wataru Ikeda, Ryosuke Matsuda, Keisuke Sakaguchi, Jun Suzuki
Sumi is a fully open 7B diffusion language model trained from scratch on 1.5T tokens, offering the first large-scale reference point for studying uniform diffusion generation in practice.
Zhenhao Yang, Xiaoshi Wu, Zhengyao Lv, Xiaoyu Shi, Xintao Wang, Pengfei Wan, Kun Gai, Kwan-Yee K. Wong
DecMem introduces a split memory system to help video generators remember long-range details, aiming to produce more consistent minute-long worlds.
Jiwen Liu, Shujuan Li, Zhixue Fang, Xiaohan Li, Yan Zhou, Zijie Meng, Zhimin Zhang, Yawen Luo, Guoxin Zhang, Yu-Shen Liu, Pengfei Wan
OmniDirector makes video generation follow complex multi-shot camera moves by turning camera paths into visual motion grids and training a large controllable diffusion system on them.
Jiaqi Yan, Xiangyu Chen, Xinlin Zhong, Haibin Huang, Chi Zhang, Jie Liu, Jiantao Zhou, Xuelong Li
SwiftVR makes high-quality video restoration run in real time on consumer GPUs by redesigning one-step generative modeling for efficient streaming inference.
Wei Song, Tianhang Wang, Yitong Chen, Tong Zhang, Zuxuan Wu, Ming Li, Jiaqi Wang, Kaicheng Yu
This paper proposes a new way to turn images into discrete tokens by quantizing channels instead of patches, enabling a fresh autoregressive image generator that paints details progressively from coarse structure to fine texture.
Zizun Li, Haoyu Guo, Runzhe Teng, Chunhua Shen, Tong He
This paper teaches video generators to follow camera motion more faithfully by rewarding them with geometry-based feedback instead of only supervised examples.
Huaisong Zhang, Hao Yu, Yuxuan Zhang, Jiahe Wang, Xinrui Chen, Haoxiang Cao, Feng Lu, Wendong Zhang, Changqian Yu, Chun Yuan
This paper turns image-quality debugging for text-to-image models into a structured task that pinpoints what is wrong, where it is wrong, why, and how important it is, then uses those defects to improve the generator.
Tim Merino, Sam Earle, Ryunosuke Iwai, Julian Togelius, Edoardo Cetin
Dream-Cubed turns Minecraft worlds into a massive training ground for learning controllable 3D generative models that can fill in, extend, and create block-based environments interactively.
Hanhui Wang, Yiming Xie, Haiwen Feng, Zhaoyang Lv, Shenlong Wang, Huaizu Jiang
StreamForce is a real-time video generation system that can instantly react to continuous force inputs, aiming to make generated motion both physically responsive and visually realistic.
Maitreya Patel, Jingtao Li, Weiming Zhuang, Yezhou Yang, Lingjuan Lv
VibeToken is a new image tokenization and generation approach that lets autoregressive models create high-resolution images at arbitrary aspect ratios with far less compute.
Yuxuan Bian, Zeyue Xue, Songchun Zhang, Shiyi Zhang, Weiyang Jin, Yaowei Li, Junhao Zhuang, Haoran Li, Jie Huang, Haoyang Huang, Nan Duan, Qiang Xu
This paper introduces a new memory system that lets video generators keep producing coherent frames for extremely long durations in real time, pushing toward truly infinite video generation.
Dong Chen, Fangyun Wei, Ziyu Wan, Dongdong Chen, Jiawei Zhang, Jinjing Zhao, Sirui Zhang, Yang Yue, Zhiyang Liang, Baining Guo, Chong Luo, Jianmin Bao, Ji Li, Lei Shi, Qinhong Yang, Xiuyu Wu, Xuelu Feng, Yan Lu, Yanchen Dong, Yitong Wang, Yunuo Chen
Lens is a more compute-efficient text-to-image model that uses richer data, smarter batching, and post-training tricks to match or beat larger systems while generating images faster.
Jianzong Wu, Hao Lian, Jiongfan Yang, Dachao Hao, Ye Tian, Yunhai Tong, Jingyuan Zhu, Biaolong Chen, Qiaosong Qi, Aixi Zhang, Wanggui He, Mushui Liu, Jinlong Liu, Hao Jiang
LoomVideo is a faster, smaller video generation-and-editing model that uses a clever conditioning trick to avoid expensive token concatenation while still handling multimodal inputs.
Hidir Yesiltepe, Jiazhen Hu, Tuna Han Salih Meral, Adil Kaan Akan, Kaan Oktay, Hoda Eldardiry, Pinar Yanardag
VideoMLA makes minute-scale autoregressive video diffusion much cheaper in memory by compressing the KV cache with a low-rank latent representation, enabling longer video rollouts with better throughput and strong quality.
Krishnakumar Balasubramanian
This paper gives a mathematical guarantee for a new one-step generative modeling method by showing how particle-based drifting can converge and how its errors can be explicitly bounded.
Stanislav R. Kirpichenko, Andrei V. Konstantinov, Lev V. Utkin
This paper uses diffusion models to generate survival times and censoring outcomes, aiming to make continuous-time survival analysis more flexible and accurate without discretizing time.
Samson Gourevitch, Yazid Janati, Dario Shariatian, Umut Simsekli, Eric Moulines, Eric P. Xing, Alain Durmus
This paper shows that a subtle mismatch in how uniform diffusion models are trained and sampled can be fixed with a leave-one-out view, leading to better language generation and a cleaner alternative formulation.
Pablo Moreno-Muñoz, Adrian Müller, Gergely Neu
This paper turns generative modeling into a control problem and uses value functions to learn fast, straight transport paths for sampling data efficiently.
Yifan Wang, Yijia Ma, Wen Li, Chenyu You
This paper teaches a transformer to generate EEG as a continuous signal, using flow matching and signal-aware constraints to better preserve the brain-wave patterns that traditional denoising methods often miss.
Antoine Maillard, Sebastian Goldt
This paper shows that generative models can look converged without truly learning the underlying factors, separating surface-level matching from deeper representation recovery.