ICLR 2025 论文列表

Diffusion models (DMs) have achieved record-breaking performance in image generation tasks. Nevertheless, in practice, the training-sampling discrepancy, caused by score estimation error and discretization error, limits the modeling ability of DMs, a phenomenon known as exposure bias. To alleviate such exposure bias and further improve the generative performance, we put forward a prompt learning framework built upon a lightweight prompt prediction model. Concretely, our model learns an anti-bias prompt for the generated sample at each sampling step, aiming to compensate for the exposure bias that arises. Following this design philosophy, our framework rectifies the sampling trajectory to match the training trajectory, thereby reducing the divergence between the target data distribution and the modeling distribution. To train the prompt prediction model, we simulate exposure bias by constructing training data and introduce a time-dependent weighting function for optimization. Empirical results on various DMs demonstrate the superiority of our prompt learning framework across three benchmark datasets. Importantly, the optimized prompt prediction model effectively improves image quality with only a 5\% increase in sampling overhead, which remains negligible.

While continuous diffusion models excel in modeling continuous distributions, their application to categorical data has been less effective. Recent work has shown that ratio-matching through *score-entropy* within a continuous-time discrete Markov chain (CTMC) framework serves as a competitive alternative to autoregressive models in language modeling. To enhance this framework, we first introduce three new theorems concerning the KL divergence between the data and learned distribution. Our results serve as the discrete counterpart to those established for continuous diffusion models and allow us to derive an improved upper bound of the perplexity. Second, we empirically show that ratio-matching performed by minimizing the *denoising cross-entropy* between the clean and corrupted data enables models to outperform those utilizing score-entropy with up to 10\% lower perplexity/generative-perplexity, and 15\% faster training steps. To further support our findings, we introduce and evaluate a novel CTMC transition-rate matrix that allows prediction refinement, and derive the analytic expression for its matrix exponential which facilitates the computation of conditional ratios thus enabling efficient training and generation.

Diffusion Transformer, the backbone of Sora for video generation, successfully scales the capacity of diffusion models, pioneering new avenues for high-fidelity sequential data generation. Unlike static data such as images, sequential data consists of consecutive data frames indexed by time, exhibiting rich spatial and temporal dependencies. These dependencies represent the underlying dynamic model and are critical to validate the generated data. In this paper, we make the first theoretical step towards bridging diffusion transformers for capturing spatial-temporal dependencies. Specifically, we establish score approximation and distribution estimation guarantees of diffusion transformers for learning Gaussian process data with covariance functions of various decay patterns. We highlight how the spatial-temporal dependencies are captured and affect learning efficiency. Our study proposes a novel transformer approximation theory, where the transformer acts to unroll an algorithm. We support our theoretical results by numerical experiments, providing strong evidence that spatial-temporal dependencies are captured within attention layers, aligning with our approximation theory.

The increasing demand for controllable outputs in text-to-image generation has spurred advancements in multi-instance generation (MIG), allowing users to define both instance layouts and attributes. However, unlike image-conditional generation methods such as ControlNet, MIG techniques have not been widely adopted in state-of-the-art models like SD2 and SDXL, primarily due to the challenge of building robust renderers that simultaneously handle instance positioning and attribute rendering. In this paper, we introduce Depth-Driven Decoupled Image Synthesis (3DIS), a novel framework that decouples the MIG process into two stages: (i) generating a coarse scene depth map for accurate instance positioning and scene composition, and (ii) rendering fine-grained attributes using pre-trained ControlNet on any foundational model, without additional training. Our 3DIS framework integrates a custom adapter into LDM3D for precise depth-based layouts and employs a finetuning-free method for enhanced instance-level attribute rendering. Extensive experiments on COCO-Position and COCO-MIG benchmarks demonstrate that 3DIS significantly outperforms existing methods in both layout precision and attribute rendering. Notably, 3DIS offers seamless compatibility with diverse foundational models, providing a robust, adaptable solution for advanced multi-instance generation. The code is available at: https://github.com/limuloo/3DIS.

Sampling algorithms play an important role in controlling the quality and runtime of diffusion model inference. In recent years, a number of works (Chen et al., 2023c;b; Benton et al., 2023; Lee et al., 2022) have analyzed algorithms for diffusion sampling with provable guarantees; these works show that for essentially any data distribution, one can approximately sample in polynomial time given a sufficiently accurate estimate of its score functions at different noise levels. In this work, we propose a new scheme inspired by Shen and Lee's randomized midpoint method for log-concave sampling (Shen & Lee, 2019). We prove that this approach achieves the best known dimension dependence for sampling from arbitrary smooth distributions in total variation distance ($\widetilde O(d^{5/12})$ compared to $\widetilde O(\sqrt{d})$ from prior work). We also show that our algorithm can be parallelized to run in only $\widetilde O(\log^2 d)$ parallel rounds, constituting the first provable guarantees for parallel sampling with diffusion models. As a byproduct of our methods, for the well-studied problem of log-concave sampling in total variation distance, we give an algorithm and simple analysis achieving dimension dependence $\widetilde O(d^{5/12})$ compared to $\widetilde O(\sqrt{d})$ from prior work.

Diffusion models, the most popular generative paradigm so far, can inject conditional information into the generation path to guide the latent towards desired directions. However, existing text-to-image diffusion models often fail to maintain high image quality and high prompt-image alignment for those challenging prompts. To mitigate this issue and enhance existing pretrained diffusion models, we mainly made three contributions in this paper. First, we propose **diffusion self-reflection** that alternately performs denoising and inversion and demonstrate that such diffusion self-reflection can leverage the guidance gap between denoising and inversion to capture prompt-related semantic information with theoretical and empirical evidence. Second, motivated by theoretical analysis, we derive Zigzag Diffusion Sampling (Z-Sampling), a novel self-reflection-based diffusion sampling method that leverages the guidance gap between denosing and inversion to accumulate semantic information step by step along the sampling path, leading to improved sampling results. Moreover, as a plug-and-play method, Z-Sampling can be generally applied to various diffusion models (e.g., accelerated ones and Transformer-based ones) with very limited coding and computational costs. Third, our extensive experiments demonstrate that Z-Sampling can generally and significantly enhance generation quality across various benchmark datasets, diffusion models, and performance evaluation metrics. For example, DreamShaper with Z-Sampling can self-improve with the HPSv2 winning rate up to **94%** over the original results. Moreover, Z-Sampling can further enhance existing diffusion models combined with other orthogonal methods, including Diffusion-DPO. The code is publicly available at [github.com/xie-lab-ml/Zigzag-Diffusion-Sampling](https://github.com/xie-lab-ml/Zigzag-Diffusion-Sampling).

Discrete diffusion has achieved state-of-the-art performance, outperforming or approaching autoregressive models on standard benchmarks. In this work, we introduce *Discrete Diffusion with Planned Denoising* (DDPD), a novel framework that separates the generation process into two models: a planner and a denoiser. At inference time, the planner selects which positions to denoise next by identifying the most corrupted positions in need of denoising, including both initially corrupted and those requiring additional refinement. This plan-and-denoise approach enables more efficient reconstruction during generation by iteratively identifying and denoising corruptions in the optimal order. DDPD outperforms traditional denoiser-only mask diffusion methods, achieving superior results on language modeling benchmarks such as *text8*, *OpenWebText*, and token-based generation on *ImageNet 256 × 256*. Notably, in language modeling, DDPD significantly reduces the performance gap between diffusion-based and autoregressive methods in terms of generative perplexity. Code is available at [github.com/liusulin/DDPD](https://github.com/liusulin/DDPD).

Urban scene generation has been developing rapidly recently. However, existing methods primarily focus on generating static and single-frame scenes, overlooking the inherently dynamic nature of real-world driving environments. In this work, we introduce DynamicCity, a novel 4D occupancy generation framework capable of generating large-scale, high-quality dynamic 4D scenes with semantics. DynamicCity mainly consists of two key models. **1)** A VAE model for learning HexPlane as the compact 4D representation. Instead of using naive averaging operations, DynamicCity employs a novel **Projection Module** to effectively compress 4D features into six 2D feature maps for HexPlane construction, which significantly enhances HexPlane fitting quality (up to **12.56** mIoU gain). Furthermore, we utilize an **Expansion & Squeeze Strategy** to reconstruct 3D feature volumes in parallel, which improves both network training efficiency and reconstruction accuracy than naively querying each 3D point (up to **7.05** mIoU gain, **2.06x** training speedup, and **70.84\%** memory reduction). **2)** A DiT-based diffusion model for HexPlane generation. To make HexPlane feasible for DiT generation, a **Padded Rollout Operation** is proposed to reorganize all six feature planes of the HexPlane as a squared 2D feature map. In particular, various conditions could be introduced in the diffusion or sampling process, supporting **versatile 4D generation applications**, such as trajectory- and command-driven generation, inpainting, and layout-conditioned generation. Extensive experiments on the CarlaSC and Waymo datasets demonstrate that DynamicCity significantly outperforms existing state-of-the-art 4D occupancy generation methods across multiple metrics. The code and models have been released to facilitate future research.

We introduce a novel method for generating 360° panoramas from text prompts or images. Our approach leverages recent advances in 3D generation by employing multi-view diffusion models to jointly synthesize the six faces of a cubemap. Unlike previous methods that rely on processing equirectangular projections or autoregressive generation, our method treats each face as a standard perspective image, simplifying the generation process and enabling the use of existing multi-view diffusion models. We demonstrate that these models can be adapted to produce high-quality cubemaps without requiring correspondence-aware attention layers. Our model allows for fine-grained text control, generates high resolution panorama images and generalizes well beyond its training set, whilst achieving state-of-the-art results, both qualitatively and quantitatively.

Generative modeling of crystal data distribution is an important yet challenging task due to the unique periodic physical symmetry of crystals. Diffusion-based methods have shown early promise in modeling crystal distribution. More recently, Bayesian Flow Networks were introduced to aggregate noisy latent variables, resulting in a variance-reduced parameter space that has been shown to be advantageous for modeling Euclidean data distributions with structural constraints (Song, et al.,2023). Inspired by this, we seek to unlock its potential for modeling variables located in non-Euclidean manifolds e.g. those within crystal structures, by overcoming challenging theoretical issues. We introduce CrysBFN, a novel crystal generation method by proposing a periodic Bayesian flow, which essentially differs from the original Gaussian-based BFN by exhibiting non-monotonic entropy dynamics. To successfully realize the concept of periodic Bayesian flow, CrysBFN integrates a new entropy conditioning mechanism and empirically demonstrates its significance compared to time-conditioning. Extensive experiments over both crystal ab initio generation and crystal structure prediction tasks demonstrate the superiority of CrysBFN, which consistently achieves new state-of-the-art on all benchmarks. Surprisingly, we found that CrysBFN enjoys a significant improvement in sampling efficiency, e.g., 200x speedup (10 v.s. 2000 steps network forwards) compared with previous Diffusion-based methods on MP-20 dataset.

Consistency models (CMs) are a powerful class of diffusion-based generative models optimized for fast sampling. Most existing CMs are trained using discretized timesteps, which introduce additional hyperparameters and are prone to discretization errors. While continuous-time formulations can mitigate these issues, their success has been limited by training instability. To address this, we propose a simplified theoretical framework that unifies previous parameterizations of diffusion models and CMs, identifying the root causes of instability. Based on this analysis, we introduce key improvements in diffusion process parameterization, network architecture, and training objectives. These changes enable us to train continuous-time CMs at an unprecedented scale, reaching 1.5B parameters on ImageNet 512×512. Our proposed training algorithm, using only two sampling steps, achieves FID scores of 2.06 on CIFAR-10, 1.48 on ImageNet 64×64, and 1.88 on ImageNet 512×512, narrowing the gap in FID scores with the best existing diffusion models to within 10\%.

We propose Framer for interactive frame interpolation, which targets producing smoothly transitioning frames between two images as per user creativity. Concretely, besides taking the start and end frames as inputs, our approach supports customizing the transition process by tailoring the trajectory of some selected keypoints. Such a design enjoys two clear benefits. First, incorporating human interaction mitigates the issue arising from numerous possibilities of transforming one image to another, and in turn enables finer control of local motions. Second, as the most basic form of interaction, keypoints help establish the correspondence across frames, enhancing the model to handle challenging cases (e.g., objects on the start and end frames are of different shapes and styles). It is noteworthy that our system also offers an "autopilot" mode, where we introduce a module to estimate the keypoints and refine the trajectory automatically, to simplify the usage in practice. Extensive experimental results demonstrate the appealing performance of Framer on various applications, such as image morphing, time-lapse video generation, cartoon interpolation, etc. The code, model, and interface are publicly accessible at https://github.com/aim-uofa/Framer.

Prompt learning has demonstrated promising results in fine-tuning pre-trained multimodal models. However, the performance improvement is limited when applied to more complex and fine-grained tasks. The reason is that most existing methods directly optimize the parameters involved in the prompt generation process through loss backpropagation, which constrains the richness and specificity of the prompt representations. In this paper, we propose Diffusion-Driven Prompt Generator (Diff-Prompt), aiming to use the diffusion model to generate rich and fine-grained prompt information for complex downstream tasks. Specifically, our approach consists of three stages. In the first stage, we train a Mask-VAE to compress the masks into latent space. In the second stage, we leverage an improved Diffusion Transformer (DiT) to train a prompt generator in the latent space, using the masks for supervision. In the third stage, we align the denoising process of the prompt generator with the pre-trained model in the semantic space, and use the generated prompts to fine-tune the model. We conduct experiments on a complex pixel-level downstream task, referring expression comprehension, and compare our method with various parameter-efficient fine-tuning approaches. Diff-Prompt achieves a maximum improvement of 8.87 in R@1 and 14.05 in R@5 compared to the foundation model and also outperforms other state-of-the-art methods across multiple metrics. The experimental results validate the effectiveness of our approach and highlight the potential of using generative models for prompt generation. Code is available at https://github.com/Kelvin-ywc/diff-prompt.

We present TANGO, a framework for generating co-speech body-gesture videos. Given a few-minute, single-speaker reference video and target speech audio, TANGO produces high-fidelity videos with synchronized body gestures. TANGO builds on Gesture Video Reenactment (GVR), which splits and retrieves video clips using a directed graph structure - representing video frames as nodes and valid transitions as edges. We address two key limitations of GVR: audio-motion misalignment and visual artifacts in GAN-generated transition frames. In particular, i) we propose retrieving gestures using latent feature distance to improve cross-modal alignment. To ensure the latent features could effectively model the relationship between speech audio and gesture motion, we implement a hierarchical joint embedding space (AuMoClip); ii) we introduce the diffusion-based model to generate high-quality transition frames. Our diffusion model, Appearance Consistent Interpolation (ACInterp), is built upon AnimateAnyone and includes a reference motion module and homography background flow to preserve appearance consistency between generated and reference videos. By integrating these components into the graph-based retrieval framework, TANGO reliably produces realistic, audio-synchronized videos and outperforms all existing generative and retrieval methods. Our code, pretrained models, and datasets are publicly available at https://github.com/CyberAgentAILab/TANGO.

Molecular generation and molecular property prediction are both crucial for drug discovery, but they are often developed independently. Inspired by recent studies, which demonstrate that diffusion model, a prominent generative approach, can learn meaningful data representations that enhance predictive tasks, we explore the potential for developing a unified generative model in the molecular domain that effectively addresses both molecular generation and property prediction tasks. However, the integration of these tasks is challenging due to inherent inconsistencies, making simple multi-task learning ineffective. To address this, we propose UniGEM, the first unified model to successfully integrate molecular generation and property prediction, delivering superior performance in both tasks. Our key innovation lies in a novel two-phase generative process, where predictive tasks are activated in the later stages, after the molecular scaffold is formed. We further enhance task balance through innovative training strategies. Rigorous theoretical analysis and comprehensive experiments demonstrate our significant improvements in both tasks. The principles behind UniGEM hold promise for broader applications, including natural language processing and computer vision.

We present CogVideoX, a large-scale text-to-video generation model based on diffusion transformer, which can generate 10-second continuous videos that align seamlessly with text prompts, with a frame rate of 16 fps and resolution of 768 x 1360 pixels. Previous video generation models often struggled with limited motion and short durations. It is especially difficult to generate videos with coherent narratives based on text. We propose several designs to address these issues. First, we introduce a 3D Variational Autoencoder (VAE) to compress videos across spatial and temporal dimensions, enhancing both the compression rate and video fidelity. Second, to improve text-video alignment, we propose an expert transformer with expert adaptive LayerNorm to facilitate the deep fusion between the two modalities. Third, by employing progressive training and multi-resolution frame packing, CogVideoX excels at generating coherent, long-duration videos with diverse shapes and dynamic movements. In addition, we develop an effective pipeline that includes various pre-processing strategies for text and video data. Our innovative video captioning model significantly improves generation quality and semantic alignment. Results show that CogVideoX achieves state-of-the-art performance in both automated benchmarks and human evaluation. We publish the code and model checkpoints of CogVideoX along with our VAE model and video captioning model at https://github.com/THUDM/CogVideo.

Diffusion models have a tendency to exactly replicate their training data, especially when trained on small datasets. Most prior work has sought to mitigate this problem by imposing differential privacy constraints or masking parts of the training data, resulting in a notable substantial decrease in image quality. We present CPSample, a method that modifies the sampling process to prevent training data replication while preserving image quality. CPSample utilizes a classifier that is trained to overfit on random binary labels attached to the training data. CPSample then uses classifier guidance to steer the generation process away from the set of points that can be classified with high certainty, a set that includes the training data. CPSample achieves FID scores of 4.97 and 2.97 on CIFAR-10 and CelebA-64, respectively, without producing exact replicates of the training data. Unlike prior methods intended to guard the training images, CPSample only requires training a classifier rather than retraining a diffusion model, which is computationally cheaper. Moreover, our technique provides diffusion models with greater robustness against membership inference attacks, wherein an adversary attempts to discern which images were in the model's training dataset. We show that CPSample behaves like a built-in rejection sampler, and we demonstrate its capabilities to prevent mode collapse in Stable Diffusion.

Procedural video representation learning is an active research area where the objective is to learn an agent which can anticipate and forecast the future given the present video input, typically in conjunction with textual annotations. Prior works often rely on large-scale pretraining of visual encoders and prediction models with language supervision. However, the necessity and effectiveness of extending compute intensive pretraining to learn video clip sequences with noisy text supervision have not yet been fully validated by previous works. In this work, we show that a strong off-the-shelf frozen pretrained visual encoder, along with a well designed prediction model, can achieve state-of-the-art (SoTA) performance in forecasting and procedural planning without the need for pretraining the prediction model, nor requiring additional supervision from language or ASR. Instead of learning representations from pixel space, our method utilizes the latent embedding space of publicly available vision encoders. By conditioning on frozen clip-level embeddings from observed steps to predict the actions of unseen steps, our prediction model is able to learn robust representations for forecasting through iterative denoising —leveraging the recent advances in diffusion transformers (Peebles & Xie, 2023). Empirical studies over a total of five procedural learning tasks across four datasets (NIV, CrossTask, COIN and Ego4D-v2) show that our model advances the strong baselines in long-horizon action anticipation (+2.6% in Verb ED@20, +3.1% in Noun ED@20), and significantly improves the SoTA in step forecasting (+5.0%), task classification (+3.8%), and procedure planning tasks (up to +2.28% in success rate, +3.39% in mAcc, and +0.90% in mIoU).

Sound content creation, essential for multimedia works such as video games and films, often involves extensive trial-and-error, enabling creators to semantically reflect their artistic ideas and inspirations, which evolve throughout the creation process, into the sound. Recent high-quality diffusion-based Text-to-Sound (T2S) generative models provide valuable tools for creators. However, these models often suffer from slow inference speeds, imposing an undesirable burden that hinders the trial-and-error process. While existing T2S distillation models address this limitation through $1$-step generation, the sample quality of $1$-step generation remains insufficient for production use. Additionally, while multi-step sampling in those distillation models improves sample quality itself, the semantic content changes due to their lack of deterministic sampling capabilities. Thus, developing a T2S generative model that allows creators to efficiently conduct trial-and-error while producing high-quality sound remains a key challenge. To address these issues, we introduce Sound Consistency Trajectory Models (SoundCTM), which allow flexible transitions between high-quality $1$-step sound generation and superior sound quality through multi-step deterministic sampling. This allows creators to efficiently conduct trial-and-error with $1$-step generation to semantically align samples with their intention, and subsequently refine sample quality with preserving semantic content through deterministic multi-step sampling. To develop SoundCTM, we reframe the CTM training framework, originally proposed in computer vision, and introduce a novel feature distance using the teacher network for a distillation loss. Additionally, while distilling classifier-free guided trajectories, we introduce a $\nu$-sampling, a new algorithm that offers another source of quality improvement. For the $\nu$-sampling, we simultaneously train both conditional and unconditional student models. For production-level generation, we scale up our model to 1B trainable parameters, making SoundCTM-DiT-1B the first large-scale distillation model in the sound community to achieve both promising high-quality $1$-step and multi-step full-band (44.1kHz) generation. Audio samples are available at \url{https://anonymus-soundctm.github.io/soundctm_iclr/}.