D-OPSD presents the first online self‑distillation framework for step‑distilled diffusion models, allowing them to continuously fine‑tune with only image‑text pairs, retain their fast few‑step sampling, and acquire new concepts, styles, or domain preferences without reward models.
UniVidX, a unified multimodal framework for video generation and understanding accepted at SIGGRAPH 2026, reformulates diverse video graphics tasks as conditional generation, achieving or surpassing state‑of‑the‑art performance while demonstrating strong data efficiency and cross‑domain generalization.
Omni2Sound tackles the long‑standing “generalist” dilemma of unified audio generation by constructing a high‑quality V‑T‑A dataset (SoundAtlas), employing a three‑stage progressive training pipeline, and using a simple Diffusion Transformer backbone, ultimately achieving state‑of‑the‑art performance on T2A, V2A and VT2A tasks and strong robustness on off‑screen scenarios.
The paper introduces CFG‑Ctrl, a control‑theoretic redesign of classifier‑free diffusion guidance that treats the generation process as a dynamic system, achieving more stable and accurate text‑to‑image results across multiple model scales and evaluation metrics.
This article reviews eight Kuaishou‑authored papers accepted at ICLR 2026, summarizing their problem statements, novel methods such as front‑door causal attribution, visual table retrieval, denoising rerankers, difficulty‑adaptive reasoning, diffusion code infilling, generative ordinal regression, multimodal video retrieval, e‑commerce dialogue benchmarks, and a new LLM creativity evaluator, together with reported experimental gains.
This article analyzes the rapid evolution of multimodal video generation models from separated visual‑audio pipelines to unified diffusion Transformers, detailing VAE compression, MoE scaling, cross‑modal alignment techniques, comprehensive evaluation metrics, real‑world applications, and the remaining technical challenges.
This roundup highlights recent AI breakthroughs such as NVIDIA‑MIT’s Sol‑RL framework for faster diffusion model training, Peking University’s CPL++ visual localization improvement, DeepMind’s TIPSv2 for image recognition, Boston Dynamics Spot’s AI upgrade, Anthropic’s safety paper, a major MCP protocol vulnerability, OpenAI’s GPT‑5.4 release, and the shifting AI video landscape.
NVIDIA, Hong Kong University, and MIT introduced the Sol‑RL framework, which uses reinforcement‑learning‑guided sampling to cut diffusion model training time by several‑fold without sacrificing image quality, potentially lowering entry barriers for small teams and shifting the AIGC industry toward an efficiency‑driven competition.
Sol‑RL, a framework from NVIDIA, Hong Kong University and MIT, integrates NVFP4 inference for large‑scale rollout exploration and BF16 precision for high‑fidelity regeneration, delivering up to 4.64× faster convergence at equivalent reward levels while preserving BF16 training fidelity across SANA, FLUX.1 and SD3.5‑L models.
Researchers from HKUST, CUHK and XiaoHongShu introduced TDM‑R1, a reinforcement‑learning‑based method that enables 4‑step diffusion image generation to surpass 80‑step models in speed, fidelity, and complex instruction adherence, as demonstrated on the GenEval benchmark and multiple quality metrics.
This article presents concise summaries of three recent AI‑driven finance papers—Diffolio’s diffusion‑based risk‑aware portfolio optimization, STORM’s dual‑vector‑quantized VAE factor model, and AutoHypo‑Fin’s autonomous web‑mined hypothesis generation—highlighting their motivations, methods, and experimental gains.
MIT researchers introduce DRiffusion, a draft‑and‑refine parallel framework that uncovers intrinsic parallelism in diffusion models, delivering 1.4–3.7× speedup on three GPUs while preserving near‑lossless image quality across Stable Diffusion 2.1, SDXL and SD3 evaluated on MS‑COCO.
This article introduces C²FG, a training‑free, plug‑and‑play time‑adaptive exponential control function that replaces the fixed classifier‑free guidance scale, theoretically justifies its superiority with score discrepancy bounds, and demonstrates significant FID and IS improvements across multiple diffusion architectures on ImageNet.
This article curates and summarizes recent AI research papers presented at AAAI 2026 that advance quantitative finance, covering controllable market generation, LLM‑powered alpha factor mining, risk‑aware multi‑agent portfolio management, foundation models for market data, and reinforcement‑learning trading policies.
The ACCORD framework formalizes the concept‑coupling issue in text‑to‑image diffusion models as a statistical dependency problem and resolves it with two plug‑and‑play regularization losses, dramatically improving fidelity and text control without altering model architecture.
The article introduces NVIDIA’s Physical AI suite announced at GTC 2026—including Isaac GR00T, SOMA‑X, Kimodo, and FDFO—explains each model’s architecture and purpose, and provides one‑click online tutorials that let developers experiment with humanoid robotics, human‑body modeling, motion generation, and diffusion model fine‑tuning at minimal cost.
This article reviews the DigMA model, which uses a diffusion‑guided meta‑agent to generate high‑fidelity, controllable order‑flow data for financial markets, details its problem formulation, architecture, training on Chinese stock datasets, extensive experiments—including reinforcement‑learning‑based high‑frequency trading evaluation—and demonstrates its superior accuracy and ultra‑low latency generation.
The STABLE framework integrates a conditional diffusion generator with a Black‑Litterman mean‑variance optimizer to produce style‑aware return forecasts and risk‑aware portfolio weights, achieving up to a 122.9% Sharpe‑ratio boost, lower drawdowns, and a 15.7% MSE reduction across major equity markets.
The paper introduces MIGM-Shortcut, a self‑supervised method that learns controlled latent‑state dynamics to bypass redundant bidirectional attention in Masked Image Generation Models, achieving over 4× speed‑up on state‑of‑the‑art multimodal diffusion models like Lumina‑DiMOO while preserving image quality.
On March 9 2026, AI breakthroughs ranged from Claude Opus solving a 30‑year math problem and Tesla unveiling embodied AGI to Apple’s M4 chip limit being cracked, a new 30B open‑source model surpassing Gemini, and advances in diffusion and multimodal research, reflecting rapid industry evolution.
FireRed-Image-Edit, an open‑source instruction‑driven diffusion model, combines massive high‑quality data, a dual‑stream multimodal architecture, progressive training, and a comprehensive multi‑dimensional benchmark to achieve unprecedented pixel‑level control and human‑like editing performance across diverse visual tasks.
The new Latent Forcing technique from Fei‑Fei Li’s team reorders the diffusion trajectory, first generating a latent structural sketch and then refining pixel details, which restores efficiency of latent‑space models while preserving 100 % pixel fidelity, achieving state‑of‑the‑art FID scores on ImageNet‑256.
The Pixel Mean Flow (pMF) method eliminates multi‑step sampling and latent‑space encoding, generating high‑quality images in a single step and achieving state‑of‑the‑art FID scores on ImageNet while drastically reducing computational cost.
The article reviews Black Forest Labs' FLUX.2 Klein model, highlighting its sub‑second 1024×1024 image generation, low‑VRAM requirements, four‑step inference speedups, and competitive quality versus SD3 and Midjourney V6, while also sharing Midjourney‑style prompt examples for creative design.
Alibaba’s Qwen‑Image‑Layered model, now fully open‑source, automatically separates a single image into editable RGBA layers using diffusion, offering Photoshop‑level editing, prompt‑controlled layer counts, and deep decomposition, with applications ranging from PPT de‑construction to game asset extraction, while noting limitations on realistic photos.
AI video generation now turns textual prompts into high‑quality clips using diffusion models and transformer‑based architectures; this article explains the underlying mathematics, training objectives, spatio‑temporal encoding, breakthroughs like consistent motion and physical realism, and discusses the technology’s opportunities and inherent risks.
This technical report details the end‑to‑end workflow for integrating an AIGC video generation module into a social app, covering requirement analysis, model and hardware selection, dataset construction, LoRA and full‑parameter training, multiple acceleration techniques such as Sage Attention, TeaCache, XDiT, gradient‑checkpointing offload, tiled VAE, and quantization, followed by extensive performance evaluation and metric‑based ranking of the final models.
This article presents a novel paradigm—Latent Reward Model (LRM) and Latent Preference Optimization (LPO)—that repurposes diffusion models as noise‑aware latent reward models for step‑level preference optimization, addressing the shortcomings of pixel‑level reward models, introducing multi‑preference consistent filtering, and demonstrating significant performance and efficiency gains on benchmarks such as PickScore and T2I‑CompBench++.
The LocDiff framework introduces a spherical‑harmonics Dirac‑delta encoding and a conditional Siren‑UNet diffusion model that enables accurate worldwide image geolocation without relying on predefined grids or external image libraries, outperforming prior methods in precision, generalization, and computational efficiency.
This article summarizes five recent arXiv papers that apply advanced AI techniques such as diffusion models, hierarchical attention, and stochastic differential equations to multivariate financial time‑series forecasting, portfolio selection, volatility surface generation, and gold‑futures alpha strategies, presenting their core methods and experimental results.
The paper identifies imbalanced concept distributions as the main obstacle to arbitrary concept‑combination in text‑to‑image/video generation, proposes the token‑level IMBA Distance and a lightweight IMBA Loss that adaptively re‑weights training tokens, and demonstrates through extensive experiments and a new Inert‑CompBench benchmark that this loss dramatically improves compositional ability without extra data.
This article breaks down the ICCV 2025 paper "Zero‑Shot Image Anomaly Detection Using Generative Foundation Models," explaining how DiffPathV2 leverages diffusion model denoising trajectories, six‑dimensional score errors, and SSIM weighting to detect out‑of‑distribution images without any task‑specific training, achieving state‑of‑the‑art AUROC scores across multiple benchmarks.
The paper introduces ERDM, an enhanced rolling diffusion model that integrates progressive noise scheduling and time‑loss weighting from EDM, demonstrates superior CRPS scores on Navier‑Stokes and ERA5 mid‑term weather forecasts, and achieves comparable accuracy with far lower computational cost.
This paper proposes a reinforcement‑learning framework that simultaneously ensures safety, sample efficiency, and robustness, applying a contextual‑bandit perspective to ranking/recommendation systems and text‑to‑image diffusion models, and introduces novel algorithms for safe deployment, variance‑reduced off‑policy estimation, and a LOOP method for generative RL.
This article walks through the theory behind flow‑matching generative models, contrasting them with diffusion models, detailing the velocity‑field formulation, training objective, and sampling procedure, and includes visual illustrations of the core concepts.
BranchGRPO introduces a tree‑structured branching, reward‑fusion, and lightweight pruning framework that dramatically speeds up diffusion and flow model training while delivering denser, more stable reward signals, achieving up to five‑fold faster convergence and higher alignment scores on image and video generation benchmarks.
This technical article provides a comprehensive comparison of diffusion models and flow matching, covering their intuitive explanations, underlying mathematics, training objectives, sampling efficiency, theoretical guarantees, practical examples, and code implementations to illustrate how each generative approach works.
A recent MIT‑Toyota study shows that the locality observed in image diffusion models emerges from the statistical structure of training data rather than from architectural biases, and a simple linear denoiser can replicate this behavior, reshaping how we think about model design.
FantasyWorld introduces a geometry‑enhanced framework that augments a frozen video diffusion model with a trainable geometry branch, enabling simultaneous video representation and implicit 3D field generation, achieving spatially consistent, high‑quality virtual worlds and outperforming recent baselines in multi‑view coherence and geometric fidelity.
The article surveys the evolution of video generation models—from early GANs and DCGAN to diffusion‑based approaches like Stable Diffusion and DiT—highlighting how stability, high quality, massive compute, and multimodal data pipelines are shaping the current and future paths of dynamic multimodal video generation.
This digest summarizes three recent arXiv papers that propose novel diffusion‑based generation, a channel‑independent convolution for multivariate forecasting, and a style‑guided diffusion framework, each demonstrating improved realism, coherence, and diversity of synthetic time‑series data through extensive experiments.
The article explains the core challenges of real‑time bidding, reviews Kuaishou's evolution from PID to MPC to reinforcement learning, and introduces generative reinforcement‑learning methods (GAVE and CBD) that combine decision transformers or diffusion models with value‑guided exploration and score‑based RTG, achieving significant offline and online performance gains.
InfGen introduces a resolution‑agnostic generation paradigm that replaces the VAE decoder in diffusion models, allowing any‑size image synthesis with up to ten‑fold speed gains, achieving 4K outputs in under 7 seconds while improving visual quality.
This article summarizes four recent AI‑driven finance papers: a quantum‑annealing asset clustering algorithm, a causal machine‑learning model for predicting market troughs, a multimodal large‑model approach to financial time‑series forecasting, and a diffusion‑model method for generating stochastic‑differential‑equation sample paths.
This article systematically explores the technical evolution, core principles, and emerging innovations of AI‑generated video, covering generation methods, GAN and diffusion models, transformer‑based DiT architectures, efficiency‑boosting NCR, audio‑visual V2A integration, and real‑world applications across media, education, and commerce.
This digest reviews four recent AI‑driven finance papers: a robust MCVaR portfolio optimizer with ellipsoidal support and RKHS uncertainty, a PPO‑based adaptive weighting system for LLM‑generated alphas, an empirical comparison of price‑based, GICS‑based, and LLM‑embedding stock clustering, and a diffusion‑model approach that generates future financial chart images from current charts and text prompts.
This article introduces a 1.5‑hour tutorial presented by Tsinghua researchers at IJCAI 2025, covering the latest advances in multimodal generative AI, including multimodal large language models, diffusion models, post‑training generalization techniques, and unified understanding‑generation frameworks.
DynamicFace introduces a diffusion‑based framework that explicitly decouples identity, pose, expression, illumination and background using composable 3D facial priors, achieving superior identity preservation, motion consistency and visual fidelity in both image and video face‑swapping tasks.
The paper introduces Single Trajectory Distillation (STD), a novel training framework that aligns full PF‑ODE trajectories from a fixed noisy state, uses a Trajectory Bank to cut training cost, and adds an Asymmetric Adversarial Loss to markedly improve style consistency and aesthetic quality while accelerating image and video style‑transfer diffusion models.
This article provides a critical, up‑to‑date overview of visual reinforcement learning, formalizes the problem, traces policy‑optimization evolution, categorizes over 200 recent works into four pillars, analyzes algorithms, reward design, benchmarks, and highlights open challenges and future research directions.
FLUX-Lightning, introduced by PaddleMIX, combines phased consistency distillation, adversarial learning, distribution‑matching distillation, and reflow loss to reduce diffusion model inference to just four steps while preserving image quality, and leverages the CINN compiler to achieve over 30% speed gains on A800 GPUs, surpassing existing SOTA acceleration methods.
The paper introduces SADA (Stability‑guided Adaptive Diffusion Acceleration), a novel paradigm that dynamically allocates sparsity per token using a unified stability criterion, enabling efficient ODE‑based sampling for diffusion and flow‑matching models, achieving up to 1.8× speedup with negligible fidelity loss across SD‑2, SDXL, Flux, ControlNet and MusicLDM.
DualReal introduces a novel adaptive joint training framework that simultaneously customizes subject identity and motion dynamics in video generation, overcoming the conflicts of traditional isolated approaches by using a dual-domain perception adapter and stage-fusion controller, achieving up to 31.8% improvement on CLIP‑I and DINO‑I metrics.
This article presents a differentiable solver search method that quickly finds high‑quality sampling paths for diffusion models, demonstrating significant FID improvements across Rectified‑Flow, DDPM/VP, and text‑to‑image models while requiring no model parameter changes.
Orthus, a new unified multimodal model presented at ICML 2025, leverages an autoregressive Transformer backbone with separate language and diffusion heads to enable lossless image‑text interleaved generation, outperforming existing models on both understanding and generation benchmarks while remaining computationally efficient.
HarmoniCa, a new feature‑caching framework co‑designed by HKUST, Beihang University, and SenseTime, tackles diffusion model inference bottlenecks by aligning training and inference through Step‑Wise Denoising Training and an Image Error Proxy Objective, achieving up to 2× speedup while preserving image quality.
The USP framework introduces masked latent modeling within a VAE space to pre‑train ViT encoders, enabling seamless weight transfer to both image classification, segmentation, and diffusion‑based generation tasks, dramatically speeding up DiT and SiT models while preserving strong visual representations.
The USP framework introduces masked latent modeling within a VAE space to pretrain ViT encoders, enabling seamless weight transfer to both image classification and diffusion‑based generation tasks, dramatically accelerating training while preserving strong performance across multiple benchmarks.
This article introduces VMBench, the first perception‑aligned video motion generation benchmark that defines a five‑dimensional metric suite and a meta‑guided prompt generation pipeline, and presents LD‑RPS, a zero‑shot unified image restoration framework based on latent diffusion recurrent posterior sampling, together with extensive experiments validating both systems.
This article introduces VMBench, a perception‑aligned video motion generation benchmark with a five‑dimensional metric suite and meta‑guided prompt generation, and LD‑RPS, a zero‑shot unified image restoration framework using latent diffusion and recurrent posterior sampling, detailing their motivations, innovations, experiments, and future directions.
The EvoSearch method introduced by HKUST and Kuaishou’s KuaLing team leverages test‑time scaling to dramatically improve diffusion‑based image and video generation without training, using evolutionary search along the denoising trajectory, achieving state‑of‑the‑art results on SD2.1, Flux‑1‑dev and other models.
Tencent’s TEG team secured first place in the NTIRE 2025 UGC Video Enhancement competition by introducing a progressive, three‑stage AI framework that decomposes enhancement tasks into expert models for color correction, denoising, and temporal stability, incorporates advanced loss functions, extensive hardware‑level optimizations, INT8 quantization techniques, and outlines future diffusion‑based generative enhancements.
EvoSearch, a test‑time evolutionary search method, dramatically improves image and video generation by increasing inference compute without extra training, outperforming existing scaling techniques on diffusion and flow models while maintaining robustness and diversity across multiple benchmarks.
CVPR 2025 in Nashville will feature 12 Kuaishou papers spanning large‑scale video datasets, quality assessment, 3D/4D reconstruction, controllable generation, diffusion scaling laws, multimodal simulation, and novel benchmarks, highlighting the company's cutting‑edge contributions to video AI research.
The article examines the current dominance of diffusion models in commercial video generation, contrasts them with autoregressive methods, and details how the open‑source MAGI‑1 model combines both paradigms to achieve longer, more controllable video synthesis while addressing scalability and quality challenges.
MVPainter introduces a fully open‑source pipeline that generates high‑quality, PBR‑compatible 3D textures from a single reference image and a white model by leveraging multi‑view diffusion, geometric control, and a human‑aligned evaluation framework, dramatically improving texture fidelity, alignment, and detail.
This article presents the LLaDA series of diffusion‑based large language models, explains how their generative‑modeling principle yields language intelligence comparable to autoregressive models, and details the multimodal LLaDA‑V architecture, training methods, experimental results, and broader implications for AI research.
SuperEdit introduces rectified instruction generation and contrastive supervision to fix noisy supervision in instruction‑based image editing, achieving up to 9.19% performance gains on Real‑Edit benchmarks without extra model parameters or pre‑training, and releases all data and code publicly.
SuperEdit introduces rectified instruction generation and contrastive supervision to fix noisy training signals in instruction‑based image editing, achieving up to 9.19% performance gains without extra parameters or pre‑training, as demonstrated on the Real‑Edit benchmark.
GPDiT, a novel autoregressive diffusion transformer, unifies diffusion and autoregressive modeling for video generation, introducing lightweight causal attention and a parameter‑free rotation‑based time conditioning that boost temporal consistency and cut training/inference costs, achieving state‑of‑the‑art results on multiple benchmarks.
This article explains the fundamental principles of diffusion models, using physics and machine‑learning analogies to describe forward and reverse diffusion, the role of Gaussian noise, iteration trade‑offs, U‑Net architecture, and shared‑weight training for image generation.
This article provides a comprehensive, step‑by‑step analysis of Diffusion Policy for robot visuomotor control, covering its motivation, task characteristics, model design, dataset preparation, training pipeline, inference procedure, experimental results, and open research questions.
PixelHacker introduces a latent class guidance (LCG) paradigm that injects foreground and background embeddings into a diffusion model, training on 14 million image‑mask pairs and achieving state‑of‑the‑art structural and semantic consistency across Places2, CelebA‑HQ and FFHQ benchmarks.
This comprehensive survey reviews the rapid progress of multimodal understanding and text‑to‑image generation models, categorises existing unified architectures into diffusion‑based, autoregressive, and hybrid paradigms, analyses their tokenisation strategies, datasets and benchmarks, and highlights current challenges and future research directions.
DP-Recon leverages generative diffusion priors and a visibility‑guided SDS loss to achieve high‑fidelity, compositional 3D scene reconstruction from extremely sparse images, delivering superior geometry, texture, and text‑driven editing capabilities demonstrated on benchmark datasets and real‑world indoor scenarios.
DiffDenoise introduces a three‑stage self‑supervised pipeline that combines a blind‑spot network, conditional diffusion modeling, and stabilized reverse diffusion sampling to dramatically improve medical image denoising performance on both synthetic and real datasets, while also offering a fast distilled version for practical deployment.
PosterMaker leverages a ControlNet‑based TextRenderNet with character‑level visual features and a reward‑driven foreground‑extension detector to generate high‑quality product posters that accurately render Chinese text (over 90% sentence accuracy) while preserving product fidelity, and is already deployed in Alibaba’s AI creative tool.
FloED introduces a dual‑branch, flow‑guided diffusion framework that dramatically improves spatio‑temporal consistency and computational efficiency for video restoration, outperforming existing text‑guided diffusion methods on both object removal and background repair benchmarks.
FlexIP introduces a dual‑adapter architecture and a dynamic weight‑gating mechanism that decouple identity preservation from personalized editing, allowing continuous control over image generation and outperforming prior SOTA methods in both fidelity and flexibility.
DCEdit introduces a precise semantic localization strategy and a dual-level control mechanism for text‑guided image editing, delivering superior background preservation and editing quality, as demonstrated on the new RW‑800 benchmark and extensive comparisons with state‑of‑the‑art diffusion models.
TurboFill introduces a fast image‑inpainting model that trains a repair adapter on a few‑step text‑to‑image diffusion backbone, achieving state‑of‑the‑art results with only four diffusion steps while dramatically reducing computational cost.
Flow Matching reformulates generative modeling by learning a time‑dependent vector field that deterministically transports Gaussian noise to data, using a neural network trained with an analytically derived L2 loss, yielding simpler training, faster convergence, and deterministic sampling that matches or exceeds diffusion model quality.
MagicColor introduces a novel multi‑instance sketch‑coloring framework that uses a two‑stage self‑play training strategy, instance guidance, and edge‑aware pixel‑level color matching to automatically produce high‑quality, consistent colors for multiple line‑art instances, outperforming prior GAN and diffusion‑based methods.
The article explains how knowledge distillation transfers capabilities from large to smaller diffusion models, covering hard and soft labels, temperature scaling, and contrasting it with data distillation, while detailing techniques such as consistency models, progressive distillation, adversarial distillation, and adversarial post‑training for model compression and step reduction.
OmniPaint introduces a unified diffusion‑based framework that achieves physically consistent object removal and insertion by leveraging a pre‑trained FLUX‑1 diffusion prior, a progressive CycleFlow training pipeline, and a novel reference‑free CFD metric for high‑fidelity image editing.
ImageRAG introduces a retrieval‑augmented generation framework that dynamically fetches relevant images to guide diffusion models, dramatically improving the synthesis of rare and fine‑grained concepts across multiple text‑to‑image systems, as demonstrated by extensive quantitative and user studies.
The article analyzes Mercury Coder, a diffusion‑based language model that generates text and code in parallel, compares its speed and quality against traditional autoregressive LLMs like GPT‑4o‑mini using a ball‑collision benchmark, and discusses the underlying score‑entropy training, current limitations, and future multimodal potential.
VideoPainter introduces a plug‑and‑play dual‑branch framework for video inpainting and editing, featuring a lightweight context encoder, ID‑consistent resampling, and the large VPData/VPBench datasets, and achieves state‑of‑the‑art results across eight quantitative and qualitative metrics.
This article explains how recent MIT research uses stochastic differential equations to model diffusion and flow processes, defines training objectives, explores conditional guidance, compares U‑Net and diffusion transformers, addresses memory challenges with latent diffusion, and surveys applications ranging from robotics to protein design.
IMAGPose introduces a unified conditional diffusion framework that combines feature‑level, image‑level, and cross‑view attention modules to generate high‑fidelity, photo‑realistic person images under diverse pose and multi‑view scenarios, outperforming prior SOTA methods on DeepFashion and Market‑1501.
The paper introduces a novel attention‑distillation loss and a guided‑sampling scheme that together enable diffusion models to faithfully transfer visual features from reference images, dramatically speeding synthesis and surpassing prior plug‑and‑play attention methods across style transfer, text‑to‑image generation, and texture synthesis tasks.
This paper introduces a lightweight Diffusion Restoration Adapter that integrates into pre‑trained diffusion priors such as StableDiffusion XL and StableDiffusion 3, dramatically reduces parameter overhead compared with ControNet, and delivers superior quantitative and visual results on real‑world image restoration benchmarks through a novel sampling strategy.
JD Retail’s self‑built AIGC platform ‘JingDianDian’ leverages multimodal diffusion models, ControlNet, RAG and reinforcement learning to automatically generate high‑quality product images, videos and marketing copy, cutting production time from days to seconds, slashing costs by over 99% for more than 350 k merchants.
DC‑AE introduces Residual Autoencoding and Decoupled High‑Resolution Adaptation to achieve up to 128× spatial compression in autoencoders, preserving reconstruction quality while delivering roughly 19× inference and 18× training speedups for high‑resolution diffusion models, as demonstrated on ImageNet and other benchmarks.
DC-ControlNet introduces intra‑ and inter‑element controllers that decouple global conditions into separate content and layout signals, enabling finer‑grained, conflict‑aware control of multi‑condition image generation and achieving higher flexibility and accuracy than traditional ControlNet approaches.
The TPAMI 2025 paper by researchers from Peking University, KAUST, and ByteDance introduces Invertible Diffusion Models (IDM), an end‑to‑end trainable, memory‑efficient diffusion framework that narrows the gap between noise estimation and image reconstruction, reduces sampling steps from 100 to 3, boosts PSNR by 2 dB, and speeds inference up to 15×, with open‑source code available.
ImageRAG introduces a dynamic retrieval‑augmented generation framework that integrates visual language models and CLIP‑based similarity search to supply reference images, enabling diffusion models like OmniGen and SDXL to better render rare and fine‑grained concepts, as demonstrated through extensive quantitative and qualitative experiments.
FlashVideo introduces a two‑stage diffusion framework that first ensures prompt fidelity at low resolution with a 5‑billion‑parameter DiT, then efficiently adds fine details at high resolution using flow matching, achieving state‑of‑the‑art quality with dramatically lower compute cost.
SnapGen is a 379 M‑parameter text‑to‑image diffusion model that produces 1024 px images on mobile devices in about 1.4 seconds, using a compact U‑Net design, multi‑stage knowledge distillation, step distillation, and optimized training tricks to outperform much larger models on standard benchmarks.
The paper introduces DiC, a fully convolutional diffusion model that rethinks 3×3 convolutions, adds sparse skip connections, stage‑specific embeddings and conditional gating, and demonstrates superior FID/IS scores and faster inference compared to diffusion Transformers across multiple scales.
