ICLR 2026 Papers

Recent hybrid video generation models combine autoregressive temporal dynamics with diffusion-based spatial denoising, but their sequential, iterative nature leads to error accumulation and long inference times. In this work, we propose a distillation-based framework for efficient causal video generation that enables high-quality synthesis with extreme limited denoising steps. Our approach builds upon Distribution Matching Distillation (DMD) framework and proposes a novel form of Adversarial Self-Distillation (ASD) strategy, which aligns the outputs of the student model's $n$-step denoising process with its $(n+1)$-step version in the distribution level. This design provides smoother supervision by bridging small intra-student gaps and more informative guidance by combining teacher knowledge with locally consistent student behavior, substantially improving training stability and generation quality in extremely few-step scenarios. In addition, we present a First-Frame Enhancement (FFE) strategy, which allocates more denoising steps to the initial frames to mitigate error propagation while applying larger skipping steps to later frames. Extensive experiments on VBench demonstrate that our method surpasses state-of-the-art approaches in both one-step and two-step video generation. Notably, our framework produces a single distilled model that flexibly supports multiple inference-step settings, eliminating the need for repeated re-distillation and enabling efficient, high-quality video synthesis.

Diffusion-based video generation can create realistic videos, yet existing image and text-based conditioning fails to offer precise motion control. Prior methods for motion control typically rely on displacement-based conditioning and require model-specific fine-tuning, which is computationally expensive and restrictive. We introduce Time-to-Move (TTM), a training-free, plug-and-play framework for motion- and appearance-controlled video generation with image-to-video (I2V) diffusion models. Our key insight is to use crude reference animations, obtained through user-friendly manipulations such as cut-and-drag or depth-based reprojection, as direct motion guidance, analogous to using coarse layout input in image editing. To integrate these signals, we adapt SDEdit to the video domain while anchoring the appearance with image conditioning. We further propose dual-clock denoising, a region-dependent strategy that enforces strong alignment in motion-specified regions and grants flexibility elsewhere, balancing fidelity to user intent with natural dynamics. This lightweight modification of the sampling process incurs no additional training or runtime cost and is compatible with any backbone. Extensive experiments on object and camera motion benchmarks show that TTM matches or exceeds existing training-based baselines in realism and motion control. Beyond this, TTM introduces a unique capability: precise appearance control through pixel-level conditioning, exceeding the limits of text-only prompting.

Subject-driven generation faces a fundamental challenge: achieving high subject fidelity while maintaining semantic alignment with textual descriptions. While recent GRPO-based approaches have shown promise in aligning generative models with human preferences, they apply uniform optimization across all denoising timesteps, ignoring the temporal dynamics of how textual and visual conditions influence generation. We present Consis-GCPO, a causal reinforcement learning framework that reformulates multi-modal condition generation through discrete-time causal modeling. Our key insight is that different conditioning signals exert varying influence throughout the denoising process—text guides semantic structure in early steps while visual references anchor details in later stages. By introducing decoupled causal intervention trajectories, we quantify instantaneous causal effects at each timestep, transforming these measurements into temporally-weighted advantages for targeted optimization. This approach enables precise tracking of textual and visual contributions, ensuring accurate credit assignment for each conditioning modality. Extensive experiments demonstrate that Consis-GCPO significantly advances personalized generation, achieving superior subject consistency while preserving strong text-following capabilities, particularly excelling in complex multi-subject scenarios.

While most autoregressive LLMs are constrained to one-by-one decoding, diffusion LLMs (dLLMs) have attracted growing interest for their potential to dramatically accelerate inference through parallel decoding. Despite this promise, the conditional independence assumption in dLLMs causes parallel decoding to ignore token dependencies, inevitably degrading generation quality when these dependencies are strong. However, existing works largely overlook these inherent challenges, and evaluations on standard benchmarks (e.g., math and coding) are not sufficient to capture the quality degradation caused by parallel decoding. To address this gap, we first provide an information-theoretic analysis of parallel decoding. We then conduct case studies on analytically tractable synthetic list operations from both data distribution and decoding strategy perspectives, offering quantitative insights that highlight the fundamental limitations of parallel decoding. Building on these insights, we propose **ParallelBench**, the first benchmark specifically designed for dLLMs, featuring realistic tasks that are trivial for humans and autoregressive LLMs yet exceptionally challenging for dLLMs under parallel decoding. Using ParallelBench, we systematically analyze both dLLMs and autoregressive LLMs, revealing that: (i) dLLMs under parallel decoding can suffer dramatic quality degradation in real-world scenarios, and (ii) current parallel decoding strategies struggle to adapt their degree of parallelism based on task difficulty, thus failing to achieve meaningful speedup without compromising quality. Our findings underscore the pressing need for innovative decoding methods that can overcome the current speed-quality trade-off. We are releasing our benchmark to help accelerate the development of truly efficient dLLMs.

Diffusion models have recently emerged as a powerful class of generative models and have achieved state-of-the-art performance in various image synthesis tasks. However, training diffusion models generally requires large amounts of data and suffer from overfitting when the dataset size is limited. To address these limitations, we propose a novel method called WILD-Diffusion, which is inspired by Wasserstein Distributionally Robust Optimization (WDRO), an important and elegant mathematical formulation from robust optimization area. Specifically, WILD-Diffusion utilizes WDRO to iteratively generate new training samples within a Wasserstein distance based uncertainty set centered at the limited data data distribution. This carefully designed method can progressively augment the training set throughout the training process and effectively overcome the obstacles caused by the limited data issue. Moreover, we establish the convergence guarantee for our algorithm even though the mixture of diffusion process and WDRO brings significant challenges to our analysis in theory. Finally, we conduct a set of experiments to verify the effectiveness of our proposed method. With WILD-Diffusion, we can achieve more than a $10$% reduction in FID using only $20$% of the training data across different datasets. Moreover, our method can attain state-of-the-art FID with as few as $100$ images, both in pretrained and non-pretrained settings.

Existing diffusion-based object removal and inpainting methods often fail to recover the fine structural and textural details of small objects. This is primarily due to the VAE encoder’s downsampling, which inevitably compresses small masked regions and causes significant detail loss, while the decoder’s upsampling alone cannot fully restore the lost fine details. However, the adverse effects of this fixed compression can be mitigated by enlarging the perspective of these regions. To this end, we propose ReFocusEraser, a two-stage framework for small object removal that combines camera-adaptive zoom-in inpainting with robust context- and shadow-aware repair. In Stage I, a camera-adaptive refocus mechanism magnifies masked regions, and a LoRA-tuned diffusion model ensures precise semantic alignment for accurate reconstruction. However, reintegrating these magnified inpainted regions into the original image introduces challenges due to VAE asymmetry, such as color shifts and seams. Stage II addresses these issues by fine-tuning an additional decoder to create a seam- and shadow-aware module that eliminates residual artifacts while preserving background consistency. Extensive experiments demonstrate that our proposed RefocusEraser achieves state-of-the-art performance, outperforming existing methods across benchmark datasets. Related code and data are available at https://github.com/ProAirVerse/ReFocusEraser.git.

Recent research on video Frame Interpolation (VFI) shows that a pretrained Video Diffusion Model (VDM) can solve many challenging scenarios, including large or complex motion. However, VDMs require tedious diffusion sampling, making the inference slow. One possible way to accelerate is to distill a multi-step model into a one-step model, but additional modules are often introduced during distillation, which significantly increase training overhead. Instead, we propose a Real-time Diffusion-based Video Frame Interpolation pipeline, \method. \method achieves efficient interpolation by disentangling this task into two subproblems: motion and appearance generation. Specifically, \method first calculates pixel movements across frames with the continuous motion fields, only utilizing a few sparse key frames. As a result, \method only forwards the diffusion model for these sparse key frames rather than for each intermediate frame, effectively reducing one-step training cost. In the second appearance estimation step, \method then only needs to create intermediate frames by warping input frames with sampled optical flows from the estimated continuous motion field in the first step. Because our diffusion model creates motions only, it can work at a fixed and relatively small resolution, leading to superior training and inference efficiency. Extensive experiments show that our \method generates comparable or superior interpolation quality compared with existing multi-step solutions. It also offers outstanding inference efficiency, interpolating 17FPS at $1024\times 576$ resolution, achieving \textbf{50$\times$ acceleration} than the fastest diffusion-based generation by Wan.

Optimal transport (OT) has become a natural framework for guiding the probability flows. Yet, the majority of recent generative models assume trivial geometry (e.g., Euclidean) and rely on strong density-estimation assumptions, yielding trajectories that do not respect the true principles of optimality in the underlying manifold. We present Hamiltonian Optimal Transport Advection (HOTA), a Hamilton–Jacobi–Bellman based method that tackles the dual dynamical OT problem explicitly through Kantorovich potentials, enabling efficient and scalable trajectory optimization. Our approach effectively evades the need for explicit density modeling, performing even when the cost functionals are non-smooth. Empirically, HOTA outperforms all baselines in standard benchmarks, as well as in custom datasets with non-differentiable costs, both in terms of feasibility and optimality.

Recent advances in video diffusion models shows promise for generating robotic decision-making data, with trajectory conditions further enabling fine-grained control. However, existing methods primarily focus on individual object motion and struggle to capture multi-object interaction crucial in complex manipulation. This limitation arises from entangled features in overlapping regions, leading to degraded visual fidelity. To address this, we present RoboMaster, a novel framework that models inter-object dynamics via a collaborative trajectory formulation. Unlike prior methods that decompose objects, our core is to decompose the interaction process into three sub-stages: pre-interaction, interaction, and post-interaction, and models each phase using the dominant object, specifically the robotic arm in the pre- and post-interaction phases and the manipulated object during interaction. This design effectively alleviates the multi-object feature fusion issue in prior work. To further ensure subject semantic consistency across the video, we incorporate appearance- and shape-aware latent representations for objects. Extensive experiments on the challenging Bridge dataset, as well as RLBench and SIMPLER benchmarks, demonstrate that our method establishs new state-of-the-art performance in trajectory-controlled video generation for robotic manipulation. Project Page: https://fuxiao0719.github.io/projects/robomaster/

We study diffusion-based world models for reinforcement learning, which offer high generative fidelity but face critical efficiency challenges in control. Current methods either require heavyweight models at inference or rely on highly sequential imagination, both of which impose prohibitive computational costs. We propose Horizon Imagination (HI), an on-policy imagination process for discrete stochastic policies that denoises multiple future observations in parallel. HI incorporates a stabilization mechanism and a novel sampling schedule that decouples the denoising budget from the effective horizon over which denoising is applied while also supporting sub-frame budgets. Experiments on Atari 100K and Craftium show that our approach maintains control performance with a sub-frame budget of half the denoising steps and achieves superior generation quality under varied schedules. Code is available at https://github.com/leor-c/horizon-imagination.

While machine learning has advanced through massive parallelization, we identify a critical blind spot: some problems are fundamentally sequential. These "inherently serial" problems—from mathematical reasoning to physical simulations to sequential decision-making—require sequentially dependent computational steps that cannot be efficiently parallelized. We formalize this distinction in complexity theory, and demonstrate that current parallel-centric architectures face fundamental limitations on such tasks. Then, we show for first time that diffusion models despite their sequential nature are incapable of solving inherently serial problems. We argue that recognizing the serial nature of computation holds profound implications on machine learning, model design, and hardware development.

Computational replication of Chinese calligraphy, a cornerstone of cultural heritage, remains challenging. Existing methods split into two flawed camps: some render high-quality isolated characters yet miss page-level aesthetics (ligatures, spacing, scale), while others attempt page/column synthesis but sacrifice calligraphic correctness. We introduce UniCalli, a unified diffusion framework for column-level recognition and generation. Training both tasks in one model is deliberate: recognition constrains the generator to preserve character identity and stroke structure, while generation supplies strong style/layout priors—together fostering concept-level abstractions (radicals, stroke configurations) that improve both tasks under long-tail, limited-label regimes. We curate a dataset of 8,000+ digitized pieces, with ~4,000 densely annotated (script labels, character boxes, transcriptions). UniCalli employs asymmetric noising and a rasterized box map to inject spatial priors, and is trained on a mix of synthetic, labeled, and unlabeled data. The model is robust to rare styles, better disentangles style from script, and attains state-of-the-art generative quality with clear gains in ligature continuity and layout fidelity, alongside stronger recognition. The framework extends to other ancient scripts, demonstrated by successful transfer to Oracle bone inscriptions and Egyptian hieroglyphs. Code and data will be released.

Text-to-video diffusion models have enabled high-quality video synthesis, yet often fail to generate temporally coherent and physically plausible motion. A key reason is the models' insufficient understanding of complex motions that natural videos often entail. Recent works tackle this problem by aligning diffusion model features with those from pretrained video encoders. However, these encoders mix video appearance and dynamics into entangled features, limiting the benefit of such alignment. In this paper, we propose a motion-centric alignment framework that learns a disentangled motion subspace from a pretrained video encoder. This subspace is optimized to predict ground-truth optical flow, ensuring it captures true motion dynamics. We then align the latent features of a text-to-video diffusion model to this new subspace, enabling the generative model to internalize motion knowledge and generate more plausible videos. Our method improves the physical commonsense in a state-of-the-art video diffusion model, while preserving adherence to textual prompts, as evidenced by empirical evaluations on VideoPhy, VideoPhy2, VBench, and VBench-2.0, along with a user study.

Latent Diffusion Models (LDMs) have recently shown great potential for image restoration owing to their powerful generative priors. However, directly applying them to ultra-high-definition image restoration (UHD-IR) often results in severe global inconsistencies and loss of fine-grained details, primarily caused by patch-based inference and the information bottleneck of the VAE. To overcome these issues, we present FreeAdapt, a plug-and-play framework that unleashes the capability of diffusion priors for UHD-IR. The core of FreeAdapt is a training-free Frequency Feature Synergistic Guidance (FFSG) mechanism, which introduces guidance at each denoising step during inference time. It consists of two modules: 1) Frequency Guidance (FreqG) selectively fuses phase information from a reference image in the frequency domain to enforce structural consistency across the entire image; 2) Feature Guidance (FeatG) injects global contextual information into the self-attention layers of the U-Net, effectively suppressing unrealistic textures in smooth regions and preserving local detail fidelity. In addition, FreeAdapt includes an optional VAE fine-tuning module, where skip connection further enhances the reconstruction of fine-grained textures. Extensive experiments demonstrate that our method achieves superior quantitative performance and visual quality compared to state-of-the-art UHD-IR approaches, and consistently delivers strong gains across multiple LDM-based backbones.

Generating molecular dynamics (MD) trajectories using deep generative models has attracted increasing attention, yet remains inherently challenging due to the limited availability of MD data and the complexities involved in modeling high-dimensional MD distributions. To overcome these challenges, we propose a novel framework that leverages structure pre-training for MD trajectory generation. Specifically, we first train a diffusion-based structure generation model on a large-scale conformer dataset, on top of which we introduce an interpolator module trained on MD trajectory data, designed to enforce temporal consistency among generated structures. Our approach effectively harnesses abundant structural data to mitigate the scarcity of MD trajectory data and effectively decomposes the intricate MD modeling task into two manageable subproblems: structural generation and temporal alignment. We comprehensively evaluate our method on the QM9 and DRUGS small-molecule datasets across unconditional generation, forward simulation, and interpolation tasks, and further extend our framework and analysis to tetrapeptide and protein monomer systems. Experimental results confirm that our approach excels in generating chemically realistic MD trajectories, as evidenced by remarkable improvements of accuracy in geometric, dynamical, and energetic measurements.

Recent image editing models have achieved impressive results while following natural language editing instructions, but they rely on supervised fine-tuning with large datasets of input-target pairs. This is a critical bottleneck, as such naturally occurring pairs are hard to curate at scale. Current workarounds use synthetic training pairs that leverage the zero-shot capabilities of existing models. However, this can propagate and magnify the artifacts of the pretrained model into the final trained model. In this work, we present a new training paradigm that eliminates the need for paired data entirely. Our approach directly optimizes a few-step diffusion model by unrolling it during training and leveraging feedback from vision-language models (VLMs). For each input and editing instruction, the VLM evaluates if an edit follows the instruction and preserves unchanged content, providing direct gradients for end-to-end optimization. To ensure visual fidelity, we incorporate distribution matching loss (DMD), which constrains generated images to remain within the image manifold learned by pretrained models. We evaluate our method on standard benchmarks and include an extensive ablation study. Without any paired data, our method performs on par with various image editing diffusion models trained on extensive supervised paired data, under the few-step setting.

Score-based diffusion models have achieved remarkable progress in various domains with the ability to generate new data samples that do not exist in the training set. In this work, we study the hypothesis that such creativity arises from an interpolation effect caused by a smoothing of the empirical score function. Focusing on settings where the training set lies uniformly in a one-dimensional subspace, we probe the interplay between score smoothing and the denoising dynamics with analytical solutions and numerical experiments. In particular, we demonstrate how a smoothed score function can lead to the generation of samples that interpolate among the training data within their subspace while avoiding full memorization. Moreover, we present theoretical and empirical evidence that learning score functions with regularized neural networks can have a similar effect as score smoothing, including in simple nonlinear settings.

Diffusion large language models (dLLMs) are emerging as a promising alternative to autoregressive models (ARMs) due to their ability to capture bidirectional context and the potential for parallel generation. Despite the advantages, dLLM inference remains computationally expensive as the full input context is processed at every iteration. In this work, we analyze the generation dynamics of dLLMs and find that intermediate representations, including key, value, and hidden states, change only subtly across successive iterations. Leveraging this insight, we propose ES-dLLM, a training-free inference acceleration framework for dLLM that reduces computation by skipping tokens in early layers based on the estimated importance. Token importance is computed with intermediate tensor variation and confidence scores of previous iterations. Experiments on LLaDA-8B and Dream-7B demonstrate that ES-dLLM achieves throughput of up to 226.57 and 308.51 tokens per second (TPS), respectively, on an NVIDIA H200 GPU, delivering 5.6$\times$ to 16.8$\times$ speedup over the vanilla implementation and up to 1.85$\times$ over the state-of-the-art caching method, while preserving generation quality.

In this work, we propose K-Flow, a novel generative learning paradigm that flows along the $K$-amplitude domain, where $K$ is a scaling parameter that organizes projected coefficients (frequency bands), and amplitude refers to the norm of such coefficients. We instantiate K-Flow with three concrete $K$-amplitude transformations: Fourier transformation, Wavelet transformation, and PCA. By incorporating the $K$-amplitude transformations, K-Flow enables flow matching across the scaling parameter as time. We discuss six properties of K-Flow, covering its theoretical foundations, energy and temporal dynamics, and practical applications. Specifically, from the perspective of practical usage, K-Flow allows for steerable generation by controlling the information at different scales. To demonstrate the effectiveness of K-Flow, we conduct experiments on both unconditional and conditional image generation tasks, showing that K-Flow achieves competitive performance. Furthermore, we perform three ablation studies to illustrate how K-Flow leverages the scaling parameter for controlled image generation. Additional results, including scientific applications, are also provided.