Follow diffusion research in image and video generation, sampling, and controllability. Compare quality improvements with their computational costs.
58 papers · Latest edition October 1, 2026
Three of the latest briefs in this collection. Read the evidence and the original papers alongside them.
FoMo uses the moment when two images diverge during diffusion generation as an automated measure of how perceptually different they are.
LongLive-Plug distills key video-generation capabilities into reusable LoRA adapters that can accelerate and improve many downstream diffusion models without retraining each one.
Simplex Diffusion Models keep uncertainty alive during discrete denoising, enabling faster and stronger generation for text, code, and math tasks.
Newest editions first.
Jaihyun Lew, Mingi Jung, Minjun Park, Wooseok Song, Sungroh Yoon
FoMo uses the moment when two images diverge during diffusion generation as an automated measure of how perceptually different they are.
Shuai Yang, Luozhou Wang, Wei Huang, ZhiFei Chen, Bohan Zhang, Xiao Fu, Qianli Ma, Chen-Hsuan Lin, Weian Mao, Bryan Chu, Song Han, Yukang Chen
LongLive-Plug distills key video-generation capabilities into reusable LoRA adapters that can accelerate and improve many downstream diffusion models without retraining each one.
Justin Deschenaux, Alexandre Galashov, Andrew Campbell, Li Kevin Wenliang, James Thornton, Arnaud Doucet, Valentin De Bortoli
Simplex Diffusion Models keep uncertainty alive during discrete denoising, enabling faster and stronger generation for text, code, and math tasks.
Albert Ge, Chandan Singh, Yufan Zhuang, Xiaodong Liu, Jianfeng Gao, Frederic Sala
The work gives diffusion language models a small set of learned memory tokens so they can preserve reasoning across multiple chunks without retaining the generated text.
Gjergj Plepi, Sven Behnke
SlotDiT uses object-centric slots instead of conventional image latents to make diffusion-based robotic video prediction more structured, efficient, and task-effective.
Arman Adibi, Alireza Jafari, Mohammad Ghavamzadeh, Hadi Daneshmand
This work argues that transformers can generate data in context by internally simulating diffusion-like sampling processes.
Chuyan Chen, Haoxing Chen, Kun Chen, Zhenglin Cheng, Long Cui, Ruishan Fang, Zhangxuan Gu, Zhicheng Huang, Zhenzhong Lan, Yuanting Lei, Haoquan Li, Jianguo Li, Rongchuan Li, Sidu Li, Tao Lin, Deyuan Liu, Jiacheng Liu, Lin Liu, Yuxuan Lou, Zhisheng Lu, Yuxin Ma, Shuheng Shen, Peng Sun, Chaoyang Wang, Hongjun Wang, Xiaomei Wang, Yongxin Wang, Chengzhang Wu, Hongru Wu, Jun Xie
LLaDA-Image offers an open, diffusion-based recipe for building highly capable image generators and distills them into a model that can generate and edit images in just a few steps.
Igor Pavlovic, Thiemo Wandel, Anton Obukhov, Luca Bartolomei, Andrey Davydov, Fabio Tosi, Matteo Poggi, Sabine Süsstrunk, Dengxin Dai
Marigold V2 turns powerful diffusion models into fast, sharp, and more generalizable predictors of depth and other dense visual properties.
Subham Sekhar Sahoo, Lingjie Chen, Khiem Pham, Jonathan Geuter, Chaitanya Dwivedi, Varad Pimpalkhute, Yash Akhauri, Alexander Moreno, Mikhail Yurochkin, Zhenting Wang, Mostafa Elhoushi, Nolan Dey, Shane Bergsma, Joel Hestness, John Thickstun, Eric Xing, Zhengzhong Liu
Uno uses discrete diffusion to make standard autoregressive LLMs generate multiple tokens in parallel, delivering up to three times faster inference without sacrificing quality.
Shiqian Li, Chenguo Lin, Zhiguang Liu, Yu Tang, Jiarong Ou, Rui Chen, Yixin Zhu
4DStreamCtrl aims to make video diffusion interactive by generating long, realistic streams while jointly controlling camera motion, object movement, and depth in real time.
Yangshuai Liu, Zheming Li, Jiaao Li, Kang He, Ziliang Lai, Zhitai Liu, Chengru Song
This work makes multi-reference diffusion editing much faster by caching visual reference computations without sacrificing instruction following or image fidelity.
Shaojie Guo, Lichen Ma, Haoyang Tong, Yu He, Zipeng Guo, Xiaoan Liu, Feng Yan, Yu Guo, Fei Wang, Junshi Huang, Yan Wang
This work makes pixel-space diffusion easier to train by teaching models to recover images from coarse structures to fine details in the same order humans and vision systems naturally perceive them.
Kyle Chickering, Wei-An Lin, Swayam Bhanded, Dan Saunders, Akshat Tripathi, Jiaming Song, Shyamal Buch, Xinchen Yan
Abra shows that diffusion image models scale predictably, but reach their best performance with far more training data per parameter than language models.
Dengyang Jiang, Ruoyi Du, Zhennan Chen, Dongyang Liu, Zanyi Wang, Mingzhe Zheng, Xiangpeng Yang, Huanqia Cai, Aiming Hao, Yuming Jiang, Peng Gao, Harry Yang, Steven Hoi
This study shows how to efficiently train pixel-space image diffusion models by first learning in latent space, achieving competitive quality with substantially faster inference.
Jinshan Liu, Haoran Qin, Xiaobing Tu, Jiacheng Liu, Jiahui Hu, Zhengan Yan, Yukun Xie, Kerui Shen, Jinkui Ren, Yuqi Lin, Xiantao Zhang, Linfeng Zhang
LinCa speeds up diffusion-based image and video generation by learning which parts of intermediate features can be safely reused, achieving substantial acceleration with minimal quality loss.
DiffusionGemma Team, Adrien Ali Taïga, James Assiene, Daniele Calandriello, Rahma Chaabouni, João Gante, Tamara von Glehn, Nate Keating, Chris Knutsen, Martin Kukla, Tianlin Liu, Ivan Lobov, Ofir Nabati, João Gabriel Oliveira, Nicolas Perez-Nieves, Nastasia Prutianova, Bobak Shahriari, Jean Tarbouriech, Pavel Tyletski, Çağlar Ünlü, Cindy Wu, Glenn Cameron, Jerome Connor, Sertan Girgin, Maarten Grootendorst, Alon Levkovitch, Eliya Nachmani, Omar Sanseviero, Piotr Stanczyk, Quentin Berthet, Andrew Campbell, Clément Crepy, Valentin De Bortoli, Arnaud Doucet, Romuald Elie, Alexandre Galashov, Klaus Greff, Alexis Jacq, David Ruhe, Yu-Han Wu, Sebastian Flennerhag, Brendan O'Donoghue, George Scrivener, Shantanu Thakoor
DiffusionGemma turns a large language model into a fast parallel text generator, reaching roughly 1,500 tokens per second while preserving much of the model's reasoning and multimodal capability.
Yicheng Xiao, Wenxun Dai, Xinran Qin, Lin Song, Maoquan Zhang, Hang Xu, Yukang Chen, Yitong Li, Guohui Zhang, Yuan Zhang, Xuying Zhang, Tommy Zhang, Jianlong Yuan, Peihao Li, Shuai Lu, Siming Fu, Chuyang Zhao, Xin Han, Jie Huang, Wenbo Li, Guoqing Ma, Wei Huang, Xiaojuan Qi, Haoyang Huang, Nan Duan
JoyAI-Video-Edit aims to make high-quality, open-ended video editing run in real time by combining autoregressive generation with diffusion distillation.
Zilong Chen, Chaorui Deng, Kunchang Li, Hongyi Yuan, Haoqi Fan
This work shows that better-structured language can make visual diffusion models more capable, and uses scaling laws to build a stronger prompting system.
Yueyi Liu, Chi Zhang, Sen Cui, Miao Liu
ElasticTTT keeps diffusion models from forgetting their generative skills while adapting them for one-shot video editing.
Jiatong Li, Leo Liang, Linghe Kong, Yulun Zhang
FreqForcing uses low-frequency visual anchors to keep autoregressive video generation stable for minutes without additional training.
Rogerio Guimaraes, Pietro Perona
The method improves image generation by exploring many noise seeds early, pruning weak candidates progressively, and spending computation only on the most promising ones.
Rahul Sajnani, Yulia Gryaditskaya, Radomír Měch, Srinath Sridhar, Matheus Gadelha
A new diffusion-transformer interface lets users precisely control what appears in different image regions using text or image guidance.
Eungjune Shim, Hansol Lee, Eunjung Ju
DiffGI makes thin-shell 3D generation more precise and efficient by combining differentiable geometry images with latent diffusion.
Maohua Li, Qirui Li, Yanke Zhou, Yiduo Li, Zhaosheng Chi, Chao Xu, Cuifeng Shen, Yixuan Xu, Hanlin Tang, Kan Liu, Tao Lan, Lin Qu, Shao-Qun Zhang
This work shows that seemingly uninformative template tokens in diffusion transformers quietly preserve object identity, and that some attention heads can be pruned to save computation with modest quality loss.
Ruchit Rawal, Reza Shirkavand, Sayak Paul, Yuxin Wen, Heng Huang, Yizheng Chen, Tom Goldstein, Gowthami Somepalli
Flash-BoN speeds up diffusion model sampling by generating many cheap draft images first, then verifying and refining only the best one to get better results under the same compute budget.
Zezhong Qian, Xiaowei Chi, Chak-Wing Mak, Tianze Zhou, Ruibin Yuan, Yuhan Rui, Hengzhe Sun, Zhuoqun Wu, Yuming Li, Siyuan Qian, Sirui Han, Shanghang Zhang
HDR lets video diffusion models plan complex visual actions hierarchically before streaming them efficiently, improving reasoning while sharply reducing inference cost.
Yushi Huang, Xiangxin Zhou, Jun Zhang, Liefeng Bo, Tianyu Pang
MeanFlowNFT brings reinforcement-learning-based alignment to ultra-fast few-step image and video generators, achieving quality that can rival or exceed much slower diffusion models.
Anh Nguyen, Ngan Nguyen, Duc Vu, Trung Dao, Viet Nguyen, Quan Dao, Kien Nguyen, Chi Tran, Phong Nguyen, Khoi Nguyen, Cuong Pham, Dimitris Metaxas, Vishal M. Patel, Anh Tran
This paper shows how to distill powerful modern diffusion models into smaller one-step generators even when the teacher and student use incompatible latent spaces.
Xinyin Ma, Julius Berner, Chao Liu, Arash Vahdat, Weili Nie, Xinchao Wang
Flex-Forcing lets one video diffusion model switch between fast autoregressive generation and high-quality bidirectional generation, aiming to get the best of both worlds.
Xingyu Zheng, Xianglong Liu, Yifu Ding, Weilun Feng, Junqing Lin, Jinyang Guo, Haotong Qin
MrFlow speeds up text-to-image diffusion by generating a coarse image first and then refining it in stages, reaching large speedups without extra training.
Haoyou Deng, Keyu Yan, Chaojie Mao, Xiang Wang, Yu Liu, Changxin Gao, Nong Sang
SharpMoE improves diffusion model efficiency and image quality by routing computation to the most important tokens using cleaner saliency signals instead of noisy latents.
Shufan Li, Greg Heinrich, Hanrong Ye, Yonggan Fu, Aditya Grover, Jan Kautz, Pavlo Molchanov
This paper improves discrete diffusion image generation by letting the model revise earlier token choices and by training more effectively with large vocabularies, boosting both quality and efficiency.
Fangtai Wu, Hailong Guo, Shijie Huang, Jiayi Song, Yubo Huang, Mushui Liu, Zhao Wang, Yunlong Yu, Jiaming Liu, Ruihua Huang
CollectionLoRA condenses dozens of specialized image-editing LoRAs into a single efficient model, making diffusion-based customization easier to deploy without losing visual quality.
Ruoyu Wang, Boye Niu, Xiangxin Zhou, Yushi Huang, Tongliang Liu, Chi Zhang
This paper makes it cheaper and more stable to train image generators from human preferences by redesigning how reward gradients are backpropagated through flow-matching sampling trajectories.
Yujin Choi, Jaehong Yoon
This paper introduces a training-free way to make fast text-to-image generators safer by editing their velocity trajectories, cutting down harmful outputs while keeping benign images intact.
Kaiwen Zheng, Guande He, Min Zhao, Jintao Zhang, Huayu Chen, Jianfei Chen, Chen-Hsuan Lin, Ming-Yu Liu, Jun Zhu, Qianli Ma
This paper shows how to make streaming video and interactive world-model diffusion much faster and more practical by unifying two training styles into one distillation recipe.
Hadar Davidson, Noam Issachar, Sagie Benaim
This paper improves diffusion image generation by replacing uniform noise with frequency-aware colored noise during sampling, making the process more efficient and producing sharper images.
Yunyang Ge, Xianyi He, Zezhong Zhang, Bin Lin, Bin Zhu, Xinhua Cheng, Li Yuan
OSP-Next makes text-to-video generation faster and more hardware-friendly by mixing sparse attention, smarter parallelism, quantization, and post-training RL while keeping quality high.
Funing Fu, Tenghui Wang, Junyong Cen, Qichao Zhu, Guanyu Zhou
This paper shows that in latent diffusion models, predicting the clean latent directly can work better than predicting velocity, because the choice of training target changes the geometry of learning in compressed latent space.
Yifan Lu, Qi Wu, Jay Zhangjie Wu, Zian Wang, Huan Ling, Sanja Fidler, Xuanchi Ren
PiD replaces traditional latent decoders with a fast pixel-diffusion decoder that can turn compressed latents into high-resolution images much more efficiently and with better visual fidelity.
Mahesh Bhosale, Naresh Kumar Devulapally, Abdul Wasi, Chau Pham, Vishnu Suresh Lokhande, David Doermann
This paper tackles hallucinations in diffusion image generators by controlling the model’s score dynamics, improving reliability without sacrificing quality.
Weichen Fan, Haiwen Diao, Penghao Wu, Ziwei Liu
This paper boosts pixel-space diffusion by filtering out high-frequency noise early in the denoising process, helping the model focus on the signal where it matters most.
Georgios Tsoumplekas, Stella Bounareli, Vasileios Argyriou
This work makes it easier to combine multiple LoRA styles and concepts in diffusion models by weighting each adapter based on how important its trigger words are in the prompt.
Haojun Qiu, Kiriakos N. Kutulakos, David B. Lindell
This paper turns single-image diffusion into a fast, training-free process by using closed-form patch statistics, enabling high-quality image generation from one reference image in seconds rather than hours.
Yu Lu, Junjie Yang, Piotr Koniusz, YuXin Song, Yi Yang
FadeMem makes long video generation more memory-efficient by keeping recent details crisp while compressing older context into a hierarchical cache that preserves coherence.
Dongyang Liu, Ruoyi Du, David Liu, Dengyang Jiang, Liangchen Li, Qilong Wu, Zhen Li, Steven C. H. Hoi, Hongsheng Li, Peng Gao
This paper shows how to squeeze high-quality images out of a diffusion model in just two denoising steps by aligning the student to its teacher, separating step-specific parameters, and training end-to-end with regularization.
Zixin Jessie Chen, Zhuo Chen, Archer Wang, Jeff Gore, William T. Freeman, Congyue Deng, Marin Soljačić
This paper introduces a single diffusion model that can both generate images and perform super-resolution by starting the denoising process at different scales.
Boya Zeng, Tianze Luo, Shu Pu, Jucheng Shen, Taiming Lu, Gabriel Sarch, Zhuang Liu
This paper delivers a fully open, highly competitive text-to-image diffusion model plus a carefully tested training recipe that could become a new starting point for open generative AI research.
Agata Żywot, Iason Skylitsis, Thijmen Nijdam, Zoe Tzifa-Kratira, Derck Prinzhorn, Konrad Szewczyk, Aritra Bhowmik
This paper lets Stable Diffusion take both a text prompt and a reference image at inference time, so you can steer generation toward a desired style or visual concept without retraining.
Woojung Han, Seil Kang, Youngjun Jun, Min-Hung Chen, Fu-En Yang, Seong Jae Hwang
This paper finds that early diffusion steps can preserve better physics than later ones, and introduces a lightweight method to lock in that motion prior while still producing high-quality video.
Chao Xu, Maohua Li, Qirui Li, Yixuan Xu, Yanke Zhou, Yunhe Li, Cuifeng Shen, Hanlin Tang, Kan Liu, Tao Lan, Lin Qu, Shao-Qun Zhang
This paper argues that the way diffusion transformers pass information across layers is suboptimal, and introduces a new residual-routing method that can speed up training and improve image quality.
Xing Cong, Hanlin Tang, Kan Liu, Lan Tao, Lin Qu, Chenhao Xie
RT-Lynx speeds up diffusion transformers by sparsifying activations instead of weights, preserving image quality while making generation faster.
Ziyue Lin, Jiahe Hou, Hongyu Xia, Xinrui Xie, Feifei Wang, Yuyin Zhou, Wei Wang, Jiawei Liu, Liangqiong Qu
This paper improves image-to-image translation by splitting diffusion into two stages so one part aligns domains with noise and the other learns the actual semantic change more efficiently.
Ryan Fosdick
DEMON turns a diffusion music generator into a live, playable instrument by speeding up and restructuring how denoising controls take effect in real time.
Yuyang Zhao, Yicheng Pan, Qiyuan He, Jincheng Yu, Junsong Chen, Tian Ye, Haozhe Liu, Enze Xie, Song Han
This paper makes real-time video editing practical on a consumer GPU by pairing a diffusion-transformer model with a new consistency training trick and aggressive Blackwell-specific optimization.
Jaihoon Kim, Taehoon Yoon, Prin Phunyaphibarn, Seungjun Kim, Morteza Mardani, Minhyuk Sung
This paper speeds up reward-guided sampling in discrete diffusion models by learning a cheap twist function, making aligned generation more practical for text, DNA, proteins, and large language models.
Yuchen Liang, Ness Shroff, Yingbin Liang
This paper speeds up discrete diffusion sampling by using Gibbs-style corrections, making text and music generation much faster without extra training.
Yushi Huang, Xiangxin Zhou, Ruoyu Wang, Chi Zhang, Jun Zhang, Tianyu Pang
This paper makes fast image generators better at following human preferences by blending distillation and reinforcement learning into a new two-stage training recipe.