多模态生成与编辑

Multiplayer world models must generate independently controlled views with consistent representations of both players and their shared environment. Most existing approaches coordinate multiple players through joint multi-view generation, whose cost grows with each additional player. We present WorldCast, a distributed multiplayer world model in which each player runs a local client comprising a video generator and a state model. Using recorded player positions and map geometry during training, the state model estimates the player's position from generated video and control inputs. Clients exchange player states and project them into camera-aligned player state fields that guide where and how other players are rendered. Shared scene state enables clients to reuse one another's generated observations to maintain consistent scene appearance across views. Experiments on Counter-Strike 2 demonstrate WorldCast's consistency, real-time performance, and distributed scalability. The camera-aligned player state field improves player rendering rates by over an order of magnitude over joint-generation methods, while shared scene state improves visual consistency over whole rounds. Each client runs in real time and exchanges only player and scene states, enabling scalable multiplayer generation without a centralized computational bottleneck. Image quality remains stable over hour-long rollouts.

Identity-preserving video generation aims to maintain a subject's identity while synthesizing realistic videos. Yet a single reference portrait captures the subject's appearance under only one facial configuration. As expressions change, facial appearance can vary in highly identity-specific ways, leaving the subject's appearance under unseen expressions underdetermined by the reference alone. This expression-dependent variation also complicates evaluation: similarity to a neutral reference may decrease under strong expressions even for real images of the same person. We investigate this limitation from both generation and evaluation perspectives. First, we quantify how face-recognition similarity varies with expression intensity using controlled photographs and MEAD videos. We then construct a compact yet expressive reference gallery that captures diverse expression-dependent facial configurations. Matching against this gallery provides a more robust measure of identity similarity under expressive motion. To further expose performance degradation with expression intensity, we report identity similarity separately for mild, intense, and extreme expressions. For generation, we extend Stand-In to condition on our expression-diverse reference sets and develop a data-curation pipeline that extracts consistent yet diverse face crops from training videos. In practical settings where only a single portrait is available, we construct the reference set by synthesizing additional expressions with a pretrained facial reenactment model. On our controlled benchmark, both real and synthesized reference sets outperform the evaluated baselines in identity similarity across all three expression-intensity regimes, with the largest improvements for extreme expressions.

Dense correspondence matching has historically been bounded by simplifying spatio-temporal priors, such as smooth motion and rigid geometry. While effective for classical tasks, these assumptions break down in image editing and reference-guided generation (IEG), where transformations can preserve visual identity while breaking physical continuity. To establish identity-preserving correspondence across such transformations, we introduce FreeMatching, a generalizable framework combining generative and semantic foundation representations with heterogeneous supervision from classical datasets, tracked videos, and synthetic scenes. Teacher-guided iterative refinement further improves correspondence in IEG without dense correspondence annotations. Experimentally, a single FreeMatching model substantially improves correspondence quality on challenging IEG image pairs while retaining competitive performance on classical benchmarks. Furthermore, we demonstrate its utility as a quantitative metric for evaluating identity preservation, with scores that correlate with human judgment. The code is available at https://github.com/luping-liu/FreeMatching.

Faithful visual world simulation requires generated videos to maintain 4D world consistency, encompassing both static and dynamic consistency. Static consistency requires coherent 3D structure in static environments across viewpoints, while dynamic consistency requires plausible subject motion and consistent appearance over time. Geometry-aware post-training offers a promising way to improve world consistency. However, existing methods often rely on a static-scene assumption. Even those that accommodate dynamic scenes struggle to provide reliable static-consistency feedback, while dynamic consistency is often overlooked or inadequately assessed. To address these limitations, we introduce WorldAlign, a decoupled 4D reward framework that semantically separates static regions and dynamic subjects and provides feedback by aligning each with a world prior suited to its assumptions. For static regions, WorldAlign aligns static geometry with a geometric world prior through semantically guided masked reprojection, enabling more reliable static-consistency evaluation; an auxiliary camera-motion reward discourages nearly static solutions. For dynamic subjects, WorldAlign uses a strong vision-language model (VLM) as a dynamic world prior and constructs a VLM-as-a-judge reward based on sample-specific checklists that assess dynamicity, physical plausibility, shape, and texture consistency. This decoupled design enables more effective online post-training without requiring human preference annotations. Across two pretrained image-to-video generators, Wan2.1 and Wan2.2, WorldAlign jointly improves static and dynamic consistency over existing methods without suppressing overall or subject motion. These results support decoupled world-prior alignment for more faithful visual world simulation. Project page: https://worldalign.github.io/.

Current 3D generation methods lack explicit local control: geometric adherence is often defined by a global control strength, and appearance cannot be specified locally. We present SpaceFlow, a training-free pipeline for locally controllable 3D generation from text descriptions and a collection of geometric primitives. Each primitive serves as a proxy for an object part and is assigned a local control level, enabling users to specify whether regions should strictly follow the input shape or allow generative completion. During structure generation, we enforce these spatial constraints within the generative flow process. For appearance synthesis, the generated structure is segmented and matched to the primitives. Each generated part is conditioned only on its assigned text or image cue, thereby limiting cross-part leakage. Regional geometry metrics demonstrate that SpaceFlow preserves the specified geometry in high-control regions and enables plausible shape variation in low-control areas. A user study further indicates that the resulting balance between geometric fidelity and generative freedom remains competitive in overall quality. When evaluating appearance on fixed geometry, text-conditioned routing achieves state-of-the-art prompt faithfulness and color/material accuracy. Qualitative results additionally show localized routing of image cues. The project page is available at SpaceFlow3D.github.io.

Recent image editing systems have achieved impressive semantic understanding, visual fidelity, and instruction-following ability, while video editing remains substantially more difficult and costly. In this paper, we present a simple alternative to end-to-end video editing: instead of training a monolithic video editor, we transform a strong image editor into a video editor through anchor-based generation. Our key insight is that video editing can be decomposed into two subproblems: editing a sparse set of keyframes and propagating those edits across time. Based on this observation, we propose Anchor-based Video Editing (AVE), a two-stage framework in which a powerful image editor first performs composed editing on selected keyframes, and a motion-guided image-to-video diffusion model then generates the final video by treating the edited keyframes as fixed anchors. This design directly inherits the strengths of modern image editors while avoiding expensive end-to-end video editing training. Experiments on IVEBench and VIE-Bench show that AVE achieves strong performance in instruction following, temporal consistency, and content fidelity. Further ablations reveal that final video editing quality is strongly correlated with the quality of the image editor, suggesting that future progress in video editing may come from stronger image editing foundations and lightweight transfer to video. Code is available at https://github.com/wangf3014/AVE.

Modern video generators can realize increasingly complex visual narratives, positioning the prompt enhancer (PE) as a critical bridge from concise user instructions and multimodal references to structured cinematic plans. However, existing PE evaluation relies on rendered videos, imposing substantial computational and human costs, slowing PE training and iteration, and conflating PE quality with downstream generator behavior. To address this gap, we introduce PEBench, the first unified benchmark for direct PE evaluation across text-to-video, image-to-video, and reference-to-video prompt enhancement. It comprises 1,100 expert-verified cases and 1,005 visual assets, spanning 35 fine-grained tasks with diverse temporal, cinematic, audiovisual, and multi-reference requirements. In addition, we develop PEBench evaluation, an evidence-grounded framework that combines modality-aware fact extraction with rubric-based assessment across 24 criteria. Our systematic evaluation of representative open- and closed-source PE methods reveals an emerging shift from fine-grained descriptive expansion toward intent-preserving cinematic planning, while the caption-reconstruction and forward-refinement methods show complementary strengths in cinematic coverage and semantic fidelity or internal coherence, respectively. Human validation shows that PEBench scores align closely with expert judgments of enhanced prompts and downstream videos from Wan3.0 and MiniMax-H3, indicating that prompt-level evaluation reliably reflects downstream utility.

Recent motion generative models have demonstrated strong capabilities in synthesizing physically plausible character motion, but often overlook established animation principles used by professional animators to ground and design their animation work. Understanding and incorporating these principles into motion generative pipelines is essential for producing motions that serve not only physically grounded applications but also the needs of the character animation community. This enables the creation of characters that not only move in physically plausible ways but also feel alive, expressive, and engaging. To close this gap, we focus on the Exaggeration principle of animation and investigate how it can be incorporated into modern motion generative pipelines to produce more expressive character motions. To this end, we introduce a framework that operates at two stages of existing motion generative pipelines. The first stage introduces exaggeration during training, where we perform supervised fine-tuning of pre-trained text-to-motion models on our curated exaggeration dataset. The second stage operates at inference time, where we: (i) introduce a mathematical formulation of exaggeration based on dynamic movement primitives (DMPs); and (ii) leverage this formulation as an exaggeration guidance signal to guide existing diffusion and flow-matching text-to-motion generation models toward exaggerated motion without additional training. Through qualitative and quantitative evaluations against three strong motion generation models, we show that our methods generate more exaggerated and expressive motions while preserving neutral reference motion intent and physical plausibility.

Causal video diffusion models generate video autoregressively, which suits streaming, interactive, and long-video generation. Under standard training, however, they often yield lower generation quality than bidirectional models of the same size. Many existing approaches address this gap by initializing from or distilling a pretrained bidirectional teacher. We instead train a causal model from an image-model initialization, with no bidirectional video model at any stage. Because this path requires neither a large bidirectional teacher nor a complex distillation pipeline, it is simpler and more scalable. On this path, we find that a causal model trained on ground-truth history becomes strongly dependent on it, so that at inference errors in its own generated history propagate forward. We hypothesize that much of this dependence is unnecessary, because the current input already determines much of what the history provides. We propose Conditional Residual Prediction (CRP), a simple recipe for reducing a model's reliance on a condition: the model first predicts the target without the condition, and the condition may only add a residual on top of this prediction. Applied to history, CRP makes the model predict each chunk from the present as far as it can and use the past only for what the present cannot supply. In controlled experiments, CRP nearly closes the 6.14-point gap to a bidirectional model trained under the same setup. Scaling this recipe, we train Optica, a 2B-parameter causal video model that autoregressively generates 5-second 480p videos and reaches 82.78 on VBench with only about 15M training videos.

Diffusion and autoregression (AR) have long been seen as different categories of generative models, with diffusion specialising in continuous fields and AR specialising in discrete tokens. Recent work seeks to combine the advantages of the two models, and each hybrid fixes its decoding schedule by design. In this paper, we ask whether the performance of decoding schedules of one model can be predicted before decoding at a fixed number of steps. We describe diffusion, AR, and models in between as paths on one corruption lattice, and define the cost of a schedule as the dependence its parallel steps discard. The cost shows that the fewest steps of a zero-cost schedule are set by the geometry of the data, in the same way for tokens and for continuous fields. In particular, for data that are Markov on a graph and dependent along its paths, the fewest steps equal the graph's treedepth, which is logarithmic in the length of a sequence and linear in the side length of a grid. With fewer steps than the treedepth, every schedule pays a positive cost, whose ranking we predict before decoding with a kernel of pairwise dependence estimated from pretrained weights. Across text generation, image generation, and video generation, we verify most of the predictions about the rankings of different schedules under different metrics and benchmarks. This work therefore provides a design principle for decoding for future AR models, diffusion models, and anything in between. Our code is available at https://github.com/TSUITUENYUE/The-Lattice-of-Transition-Laws.

Prompt learning is a popular method for adapting foundation models, but learned prompts are typically task-specific and fail to generalize to new classes, domains, or compositions of tasks. In this paper, we introduce a Diffusion Meta-Prompt (DMP) model , a framework that models the distribution of learned prompts using diffusion models. Given a repository of previously learned prompts, DMP is trained and sampled without access to the original task examples or task losses, and synthesizes new prompts conditioned on natural language task descriptions. To improve the sampling stability, we introduce a test-time steering strategy for DMP, which uses the best training-selected prompt in the repository as a latent anchor during diffusion sampling, without retraining the DMP or accessing test classes. DMP improves generalization across classification, retrieval and text-to-image generation tasks, supports concept composition and negative prompting without explicit training. It reduces storage and inference costs by over 90% compared to prompt retrieval methods. For composite classification, DMP achieves upto 2.0% average gain over prior meta-learning methods across 55 pairs of datasets with gains as high as 8.5% on specific pairs such as Eurosat and Flowers. DMP also enhances cross-task generalization with ~2-9% improvement for hierarchical classification task. We further provide a theoretical guarantee bounding the expected task loss of prompts sampled from a DMP. Code is available: https://github.com/DeepakSridhar/dmp

Egocentric world models predict first-person observations conditioned on an agent's actions, but most focus on a single agent. Real embodied settings often involve multiple agents that act and interact within a shared environment. Existing multi-agent world models rely on coarse actions like locomotion, camera control, or discrete commands, leaving fine-grained embodied interactions underexplored. We formulate multi-agent egocentric world modeling as synchronized ego-stream generation for multiple agents interacting through fine-grained actions in a shared world. This requires cross-view action consistency, shared-environment consistency, and consistent propagation of interaction-induced state updates. We propose Multi-agent Egocentric World Model (ME-World), which jointly denoises multiple ego streams in a shared token sequence, conditions each stream on all agents' target-view poses, and grounds generation with shared environment memory. We train and evaluate on real and synthetic multi-agent data and introduce shared-world consistency metrics for environment, update, and identity consistency. Experiments show ME-World improves shared-world consistency, action control, identity preservation, and video quality over existing methods.

Instruction-based text-to-speech (TTS) offers control over voice characteristics and speech expression through interfaces including voice cloning and text-based voice design. Voice cloning reproduces a reference voice, whereas text-based voice design creates a voice from a natural-language description. However, neither interface directly enables users to modify the timbre of a given reference and synthesize speech with the modified voice. Meanwhile, utterance-level expressive instructions leave changes across individual text segments underspecified. We introduce EDICT, a framework that unifies global timbre editing and local expressive control by using an edited acoustic reference to anchor voice identity across segments. To enable synthesis with an instruction-edited voice, EDICT combines reference audio with structured timbre edits to generate an edited reference in codec-token space. This representation serves as a shared voice anchor for a frozen TTS backbone, allowing segment-specific natural-language instructions to guide expression. To accommodate instruction changes while supporting acoustic continuity, EDICT rebuilds the KV cache at each segment boundary, refreshing instruction conditioning while retaining bounded acoustic context from previously generated speech. Evaluations on our proposed TimbreEdit-Bench and IntraTTS-Bench demonstrate improved timbre editing and a favorable balance between local instruction adherence, speaker consistency, and transition quality. Audio demos are available.

World Action Models (WAMs) enable generalist robot manipulation by conditioning an action expert on representations from a pretrained video Diffusion Transformer (DiT). In closed-loop control, the video DiT runs at every chunk to encode the current observation into layerwise key-value (KV) pairs that the action expert queries. This prefill dominates the per-chunk computational cost, yet existing training-free accelerations leave it fully dense. We present WAM-Cache, a training-free framework that retains layerwise key-value representations across chunks and recomputes only a sparse refresh set of tokens. Crucially, we find that the intuitive heuristic of refreshing visually drifted tokens plateaus far below the dense baseline, even with an oracle predicting ground-truth KV drift. Downstream action accuracy is instead governed by where the action expert attends, not by what moved. WAM-Cache therefore selects the refresh set by uniting the action expert's cross-attention with visual latent surprise, complemented by a strict age bound that suppresses compounding error. On Fast-WAM, WAM-Cache cuts video DiT prefill FLOPs by 32-42% across RoboTwin 2.0, LIBERO, and real-world experiments, while staying within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot.

AI research progress can be viewed as the interaction between two processes: benchmark creation and method discovery. Historically, both were driven by human intelligence. However, recent advances in AI have accelerated automated method discovery, while automated benchmark creation has received comparatively less attention. To enable self-advancing systems, we propose Generative Adversarial Loop (GAL), a generator-discriminator framework alternating between two agentic searches: (1) a discriminator that generates adversarial data to expose weaknesses in current systems, and (2) a generator that discovers algorithms to overcome them. We apply this framework to approximation algorithms for efficient inference. Unlike existing auto research systems, which primarily focus on algorithm discovery, GAL introduces a discriminator agent that automates goalpost setting by continually searching for weaknesses in the current algorithm. We demonstrate adversarial data generation across four tasks: KV compression, sparse video generation, sparse attention, and context extension, where the discriminator identifies weaknesses in state of the art techniques. We further show that GAL enables autonomous improvement, with newly discovered algorithms improving not only on adversarially generated data, but also on established benchmarks. Specifically, GAL improves CompactorPress on KV compression with Qwen3-4B at 4x, raising performance on the discriminator dataset from 0.35 to 0.97, while also outperforming RULER-HARD (+0.77 pts). For context extension, GAL boosts Dual Chunk Attention from 0.20 to 0.90 on the discriminator dataset, while yielding gains on standard benchmarks(ScienceFiction (+6 pts) and PG19 32K (-0.33 PPL)). GAL thus provides a path toward autonomous goalpost setting and algorithmic improvement, where AI systems continually discover their own weaknesses and develop methods to overcome them.

Autoregressive video diffusion enables causal video streaming without a bidirectional pass over the full clip, but existing few-step systems usually retain only the opening and most recent frames in a fixed-size KV cache. Once an event leaves this window, later frames can no longer attend to it, a failure we term mid-horizon forgetting. We present Memory Forcing, a few-step streaming method that preserves this missing history without increasing the cache size. Its Archive & Working Banks partition the cache into sink, archive, and working regions, retaining diverse intermediate events alongside recent motion under fixed memory. Because absolute temporal indices drift outside the training range, Bank-aware RoPE reassigns indices at attention time so each bank remains distinguishable. At 1.3B, Memory Forcing leads on longer clips, shows the smallest drop from 5s to 60s among methods reporting all four lengths, and preserves subjects and scenes through leave-and-return. The same design scales to Wan2.2 5B, producing more physically plausible, realistic, and dynamic videos and, to our knowledge, the first public 5B model on this forcing line.

Training-free sparse attention offers a practical acceleration solution to Diffusion Transformers (DiTs) via reducing computations without fine-tuning. It typically involves estimating the importance of query-key regions and deriving sparse masks to compute only the important candidates, which inevitably introduces approximation errors that may degrade generation quality. To better balance the efficiency-quality trade-off, we propose iCATS, integrating improved importance estimation and sparse mask construction with an efficient hardware execution strategy. Specifically, for importance estimation, unlike previous works that perform independent clustering over query and key tokens based on feature similarity to estimate attention scores, iCATS demonstrates that clustering based on query-key dot-product interactions is more accurate and further reformulates this objective as a simple quadratic form for low-cost computation. For sparse mask construction, instead of using a fixed top-p rule, we observe that tolerance to sparse approximation errors varies across denoising timesteps and therefore introduce an SNR-guided sparsity schedule to adjust sparsity dynamically, leading to higher accuracy. Finally, for hardware execution, we devise a tail-merging strategy to reduce padding overhead caused by irregular cluster sizes, improving GPU kernel utilization. Extensive experiments show that iCATS achieves $2.03\times$ acceleration with 31.017 dB PSNR on HunyuanVideo-T2V-13B and $1.55\times$ acceleration with 29.301 dB PSNR on Wan2.1-T2V-14B, delivering a state-of-the-art efficiency-quality trade-off.

Large language model agents are being increasingly deployed as autonomous scientists, designing experiments and inferring mechanistic world models with minimal human oversight. Yet identifiability is often overlooked: when a plateau is reached, the agent needs to know whether it is not yet capable enough or the model simply is not identifiable from the data, in which case no amount of further experimentation of the same kind can help. We propose the Identifiability-Driven Experimental Agent (LLM-IDEA) for closed-loop discovery with an identifiability engine that returns a three-way plateau verdict: capability limit, resolvable within the design class, or certified exhausted. On ODEBench, 60 of the 62 systems with free constants are identifiable at round 0; the RC circuit is certified exhausted for every experiment that protocol can run, and a harvesting model is resolvable by one added initial condition. The identifiability engine reproduces known verdicts on Lotka-Volterra, Van der Pol, Lorenz, and a pharmacokinetic model, where it recommends the intravenous arm pharmacologists use, and it ranks the depth scorer of our own benchmark last among four observation designs. On the DiscoverPhysics benchmark, it finds two public worlds whose explanation rubric rewards a distinction no legal experiment can make, and every model there with accurate trajectories failed the explanation grade (15 of 15, against 5 of 9 in identifiable worlds, p = 0.012). On the Alien Universe, a two-body testbed we propose in which a force law switches between a provably non-identifiable and an identifiable protocol, LLM-IDEA on the identifiable protocol reaches discovery depth at least three on 8/8 seeds versus 1/8 without it. An autonomous discovery agent can thus compute, rather than guess, whether a plateau calls for more search, a better experiment of the same kind, or a different kind of experiment.

Audio generation models can translate natural-language descriptions into sound, but their outputs are typically waveforms. Their audio quality is constrained by audio compression, and their outputs do not readily support direct edits to notes, timbral parameters, or modulation relationships. We present AutoSynth, which represents MIDI performance events, fixed synthesizer parameters, and variable-length modulation routes as a unified sequence for a synthesizer, and learns their dependencies with an audio-conditioned autoregressive model. A single model supports both tasks. Given reference audio, the model directly predicts a synthesizer program; given text, it uses a pretrained audio generation model and converts the generated audio into a program. Training consists of two stages: supervised learning on large-scale audio-program pairs automatically constructed from a small set of native presets, followed by group-relative policy optimization with a mixed reward combining semantic similarity, pitch-related features, acoustic similarity, and sound usefulness. The pipeline requires neither paired text-target-program annotations nor a differentiable synthesizer. Experiments show that AutoSynth produces complete, editable synthesizer programs and achieves competitive results in both synthesizer inversion and text-driven generation. Audio demos and source code are available at https://auto-synth.github.io/.

We present LVSPM, a generalizable model that jointly estimates camera poses and synthesizes novel views from uncalibrated image collections. Trained with only RGB images and pose supervision, LVSPM avoids dense 3D ground truth and employs test-time training (TTT) layers to scale seamlessly to hundreds of input views. On RealEstate10k, Co3Dv2, and DL3DV, LVSPM surpasses VGGT in pose estimation across 16-256 views, with especially large margins at strict thresholds. For novel view synthesis under a practical protocol where more views cover larger scenes, LVSPM achieves state-of-the-art pose-free quality---surpassing even pose-dependent models in PSNR---and still maintains high quality as scene scale grows, while baselines collapse. The code is available at https://burningdust21.github.io/Projects/LVSPM .

Reinforcement learning with verifiable rewards (RLVR) has become a standard recipe for post-training vision-language models (VLMs), but it typically assumes a static training environment. As the actor improves, fixed tasks drift out of its learning frontier: many become trivial, others remain unsolvable; and the learning signal collapses. We argue that VLM post-training should evolve the visual environment alongside the actor, not just the actor itself. We propose VICO, a co-evolutionary framework in which an actor and an Environment-as-Rewriter (EnvRewriter) are trained jointly: the EnvRewriter edits verifiable image-side structures, such as scene graphs, chart tables, or protected region masks, and re-renders them to produce label-valid training samples whose difficulty is calibrated to the actor's current ability through a pass-rate-based reward. This loop continuously realigns task difficulty with actor capability without any additional human annotation. Across nine multimodal benchmarks spanning mathematical reasoning and visually grounded understanding, VICO-8B improves over its base model by up to +5.0% on out-of-domain tasks, surpasses the strongest self-evolution and text-editing co-evolution baselines by +4.3% and +8.4% respectively, and stays comparable to chart-specialized RLVR methods using 16-160 times fewer labeled samples. By shifting from human-labeled supervision to image-editing co-evolution, VICO offers a scalable path beyond static-corpus RLVR for visual reasoning.

Prompt optimization for text-to-image (T2I) generation has been pursued almost entirely as text rewriting, in which a short user brief is expanded into a longer, model-preferred token sequence. We argue that such a language-space formulation is ill-suited to structured visual design tasks such as logo creation, where a one-line brief leaves most design decisions unspecified. These decisions depend on relational priors that a linear sequence cannot encode, and they leave an uncontrolled channel through which protected marks may be reproduced. We therefore recast logo prompting as sampling within a structured design space, and instantiate this idea as DOGS (Design-space prompting with an Originality-aware GFlowNet Sampler). From a large corpus of real-world logos, we mine a typed, graph-structured design grammar whose edges record empirical co-occurrence. A GFlowNet sampler then generates design graphs with probability proportional to a terminal reward that combines recognizability, aesthetics, and corpus-relative originality. Every slot draws only from a closed design-level vocabulary, and any infringement-inducing or harmful token is removed during parsing. The originality reward further penalizes proximity to existing logos, thereby incorporating infringement avoidance into the method by construction. On two open-source renderers and against nine baselines, DOGS produces logos that are more recognizable and aesthetic, substantially more diverse, and far less prone to trademark infringement.

Unified multimodal diffusion large language models (dLLMs) offer a single architecture for both image generation and multimodal understanding, but their iterative decoding requires tens to hundreds of forward passes. Existing few-step distillation methods largely focus on either image generation or text generation, making it unclear how to compress a fully discrete multimodal dLLM into a single efficient student while preserving both generation and understanding. We introduce Omni-Diffusion-Distill, a unified two-stage distillation framework that retains strong generation and understanding capabilities while substantially reducing the inference cost of a unified multimodal dLLM. Omni-Diffusion-Distill aligns the distillation of both generation and understanding, for both images and text, in the discrete token space. In the first stage, the student is trained to skip decoding steps by replaying cached teacher trajectories, and in the second stage the student is refined on intermediate states along its own rollouts. We further remedy two sources of degradation in unified distillation with a pairwise collision penalty that reduces repetition under parallel text decoding, and entropy-matched guidance that prevents entropy collapse caused by fitting the sharpened teacher distribution in image generation. Omni-Diffusion-Distill achieves state-of-the-art trade-offs between decoding efficiency and generation and understanding performance for multimodal dLLMs, reducing image generation from 128 to 8 decoding steps and multimodal understanding from 512 to 64, giving 18.2x and 21.2x wall-clock speedups. Under these budgets, it scores 0.828 on GenEval and 83.0 on DPG-Bench for text-to-image generation, while reaching GPT judge scores of 20.0 on MM-Vet and 57.2 on COCO captioning (twice the teacher's 28.4 at the same steps) for multimodal understanding.

Gaussian splatting has emerged as a flexible representation for 3D reconstruction from posed images. However, existing methods are optimized primarily using rasterization-based losses, which supervise a splat only when it contributes to sampled camera rays. Gaussians that are occluded or contribute little to the sampled view therefore receive weak or no geometric gradients and may drift away from the underlying surface, producing undesired floaters. We introduce PCAsplat, a geometry-aware regularization framework for Gaussian splatting based on differentiable local principal component analysis (PCA). Our PCA regularizer acts directly on neighborhoods of Gaussian centers and can therefore update Gaussians that do not contribute to the current training view. We regularize the PCA eigenvalues to encourage Gaussians to move to the underlying surface with isotropic tangent-plane coverage. We also align each Gaussian normal with the PCA-estimated neighborhood normal to enforce consistent orientation. Experiments on DTU, Tanks and Temples, and NeRF Synthetic show that the splats produced by PCAsplat better approximate samples of the reference surface while substantially reducing undesired floaters. These surface-aligned splats enable downstream geometry-processing tasks, including point cloud segmentation, and direct Poisson reconstruction. Additionally, PCAsplat remains competitive under conventional novel view synthesis and mesh extraction tasks. Code will be released.

Text-to-image diffusion models are increasingly distilled into few-step variants and being deployed to enable fast inference. However, their ability to generate harmful or undesired content poses significant safety risks. Data-driven unlearning methods suppress targeted generations by fine-tuning model weights using specialized unlearning objectives. Crucially, these objectives implicitly rely on multi-step denoising dynamics, an assumption that breaks down for few-step distilled (FSD) models, resulting in ineffective forgetting. Furthermore, performing unlearning on the non-distilled base model and subsequently re-distilling it to obtain an unlearned FSD model incurs substantial computational and time overhead, making it impractical in many settings. Hence, we address this limitation with a preference-driven unlearning framework that revisits Direct Preference Optimization (DPO) for diffusion models. We show that standard DPO and its unlearning derivatives, formulated around noise-prediction error, transfer poorly to FSD models due to their altered generation dynamics. To overcome this, we introduce a modified preference optimization formulation explicitly aligned with the few-step generation properties, enabling direct concept removal in FSD models while preserving few-step efficiency and maintaining strong retention of desirable (non-targeted) capabilities. We evaluate our framework primarily on identity and NSFW (nudity) removal tasks and also extend our method to object-level unlearning. Extensive experiments demonstrate consistent and effective forgetting, and strong retention performance, establishing our method as a practical and principled solution for unlearning in FSD models.

We present Fluid-Gen-Zero, a training-free framework for physics-aware fluid-object interaction video generation that decouples physical reasoning from appearance synthesis. Our key insight is to delegate motion dynamics to a physics simulator while preserving the appearance modeling capacity of pretrained video generators. We bridge these two domains through a two-level agentic workflow: generation-time planning, where a vision-language model (VLM) agent interprets intent and the simulation rollout to organize generation clips, and latent-space guidance, which injects simulation signals into denoising through region-aware latent wrapping. This plug-and-play design is compatible with current video foundation models. We further introduce a benchmark for fluid-object interaction video generation. Across Tora (CogVideoX-based), VACE and WanMove (Wan-based), Fluid-Gen-Zero consistently improves simulation alignment, reducing object trajectory error by 26.7%-81.5% and fluid fEPE (fluid flow endpoint error) by 67.9%-84.0%, while largely preserving perceptual quality. In a human preference study, raters favor Fluid-Gen-Zero in 55.1%-74.4% of same-backbone comparisons across three backbones, and in 90.4%-94.2% of comparisons against simulation-based methods. Code and data will be released upon acceptance.

Joint-embedding predictive architectures (JEPAs) predict masked or future observations in representation space, offering a natural source of predictive latents for vision-language-action (VLA) models. Yet making these latents useful across pretraining, policy learning, and deployment requires addressing three failures: mismatch with embodiment-specific control, interference with action learning, and teacher miscalibration under distribution shifts. We introduce Juno, a unified framework built around one action-conditioned JEPA that serves as a control-aligned representation backbone, a predictive teacher, and an adaptable dynamics model. During pretraining, we train it on embodiment-matched trajectories and use a dynamic CLS loss to transfer motion-weighted patch dynamics to a compact global state. During policy learning, we fuse current-frame JEPA patches into VLA perception and use a decoupled reasoning branch with separate transformation parameters to distill future latent states for action generation. During deployment, we adapt the world model on all observed transitions, including failed rollouts, freeze the adapted teacher, and re-align the policy on verified executions using LoRA adapters and a trainable action head, without expert corrections or task rewards. On SimplerEnv, Juno raises average success from $60.9%$ to $68.5%$ over Qwen3GR00T, the strongest baseline, and test-time adaptation further reaches $72.7%$; on a real robot, it retains $70%$--$75%$ success under background, height, and object shifts where the base policy collapses to $0%$.

Generative speech enhancement (SE) is prone to linguistic hallucination when semantic constraint is unreliable under severe noise and reverberation. Moreover, approaches that generate discrete codec tokens are bounded by the quantization error of the codec decoder, regardless of token-prediction accuracy. We propose LIFT-SE, a two-stage generative framework that decouples linguistic inference from acoustic synthesis within QRes-Codec, which exposes a quantized latent and its residual-completed continuous form. The first stage predicts clean codec tokens autoregressively, conditioned on frame-aligned features from a self-supervised front-end distilled toward clean speech. The second stage applies conditional flow matching to transport Gaussian noise to the continuous latent conditioned on the predicted tokens, and the frozen decoder reconstructs the enhanced waveform. Discrete generation provides naturalness, while continuous refinement restores signal fidelity. Experiments on the DNS1 and URGENT benchmarks show that LIFT-SE attains favorable linguistic consistency under reverberant conditions together with competitive perceptual quality, and systematic ablations verify the necessity of both stages. Code will be released in the future.

Token-based world models enable fine-grained latent planning, but repeatedly processing large spatial token grids makes action search expensive. We introduce COSTGRAD, a training-free, goal-conditioned selector that ranks spatial tokens by the gradient norm of the planning cost with respect to each input token. By deriving importance from the downstream control objective, COSTGRAD targets tokens that matter for planning rather than merely for prediction. On AdaLN-conditioned predictors at $50%$ sparsity, COSTGRAD matches or exceeds full-token planning on three of four continuous-control benchmarks, while giving a measured $2.6\times$ wall-clock speedup per environment planning step. Combining token sparsity with reduced CEM search increases this to a $\sim 5\times$ total speedup while still exceeding the full-token baseline. We also identify an architecture-dependent failure mode: in a matched AdaLN-vs-concat comparison, concat maintains comparable full-token performance but pure COSTGRAD loses its advantage over random selection. This difference tracks action-pathway drift: gradient-selected removal produces less drift than random removal on AdaLN, but more on concat. These results highlight selector-architecture compatibility as a design axis for sparse world-model planning. Project page and demos: https://ycxuyingchen.github.io/costgrad/

Scribble-based interaction provides a lightweight and intuitive way for users to specify image editing intents in interactive editing tools. However, current image editing models based on VLMs or LLMs struggle to understand and execute edits based solely on scribble inputs. To systematically study this problem, we construct a new benchmark, ScribbleEdit, that evaluates the ability of image editing models to perform image editing conditioned on scribbles. This task requires both a deep understanding of the intention of the scribble and an accurate interpretation of its spatial information. In ScribbleEdit, we design an automated data construction pipeline and introduce a dedicated evaluation protocol that explicitly measures intention alignment. Our analysis reveals that existing VLM/LLM-based editing models fail to accurately capture scribble intentions. To guide future progress on scribble-only image editing, we propose a simple yet effective soft-token baseline, which enhances the model's understanding of scribble semantics and outperforms standard image editing models on our benchmark. Our evaluation and baseline together provide a concrete foundation for assessing and improving the scribble-driven image editing.