ICLR 2025 论文列表

Recently, large-scale diffusion models have made impressive progress in text-to-image (T2I) generation. To further equip these T2I models with fine-grained spatial control, approaches like ControlNet introduce an extra network that learns to follow a condition image. However, for every single condition type, ControlNet requires independent training on millions of data pairs with hundreds of GPU hours, which is quite expensive and makes it challenging for ordinary users to explore and develop new types of conditions. To address this problem, we propose the CtrLoRA framework, which trains a Base ControlNet to learn the common knowledge of image-to-image generation from multiple base conditions, along with condition-specific LoRAs to capture distinct characteristics of each condition. Utilizing our pretrained Base ControlNet, users can easily adapt it to new conditions, requiring as few as 1,000 data pairs and less than one hour of single-GPU training to obtain satisfactory results in most scenarios. Moreover, our CtrLoRA reduces the learnable parameters by 90% compared to ControlNet, significantly lowering the threshold to distribute and deploy the model weights. Extensive experiments on various types of conditions demonstrate the efficiency and effectiveness of our method. Codes and model weights will be released at https://github.com/xyfJASON/ctrlora.

Text-to-image (T2I) diffusion models have demonstrated impressive image generation capabilities. Still, their computational intensity prohibits resource-constrained organizations from deploying T2I models after fine-tuning them on their internal *target* data. While pruning techniques offer a potential solution to reduce the computational burden of T2I models, static pruning methods use the same pruned model for all input prompts, overlooking the varying capacity requirements of different prompts. Dynamic pruning addresses this issue by utilizing a separate sub-network for each prompt, but it prevents batch parallelism on GPUs. To overcome these limitations, we introduce Adaptive Prompt-Tailored Pruning (APTP), a novel prompt-based pruning method designed for T2I diffusion models. Central to our approach is a *prompt router* model, which learns to determine the required capacity for an input text prompt and routes it to an architecture code, given a total desired compute budget for prompts. Each architecture code represents a specialized model tailored to the prompts assigned to it, and the number of codes is a hyperparameter. We train the prompt router and architecture codes using contrastive learning, ensuring that similar prompts are mapped to nearby codes. Further, we employ optimal transport to prevent the codes from collapsing into a single one. We demonstrate APTP's effectiveness by pruning Stable Diffusion (SD) V2.1 using CC3M and COCO as *target* datasets. APTP outperforms the single-model pruning baselines in terms of FID, CLIP, and CMMD scores. Our analysis of the clusters learned by APTP reveals they are semantically meaningful. We also show that APTP can automatically discover previously empirically found challenging prompts for SD, *e.g.,* prompts for generating text images, assigning them to higher capacity codes.

Recent advancements in video generation have been greatly driven by video diffusion models, with camera motion control emerging as a crucial challenge in creating view-customized visual content. This paper introduces trajectory attention, a novel approach that performs attention along available pixel trajectories for fine-grained camera motion control. Unlike existing methods that often yield imprecise outputs or neglect temporal correlations, our approach possesses a stronger inductive bias that seamlessly injects trajectory information into the video generation process. Importantly, our approach models trajectory attention as an auxiliary branch alongside traditional temporal attention. This design enables the original temporal attention and the trajectory attention to work in synergy, ensuring both precise motion control and new content generation capability, which is critical when the trajectory is only partially available. Experiments on camera motion control for images and videos demonstrate significant improvements in precision and long-range consistency while maintaining high-quality generation. Furthermore, we show that our approach can be extended to other video motion control tasks, such as first-frame-guided video editing, where it excels in maintaining content consistency over large spatial and temporal ranges.

Accelerating the sampling speed of diffusion models remains a significant challenge. Recent score distillation methods distill a heavy teacher model into a student generator to achieve one-step generation, which is optimized by calculating the difference between two score functions on the samples generated by the student model. However, there is a score mismatch issue in the early stage of the score distillation process, since existing methods mainly focus on using the endpoint of pre-trained diffusion models as teacher models, overlooking the importance of the convergence trajectory between the student generator and the teacher model. To address this issue, we extend the score distillation process by introducing the entire convergence trajectory of the teacher model and propose $\textbf{Dis}$tribution $\textbf{Back}$tracking Distillation ($\textbf{DisBack}$). DisBask is composed of two stages: $\textit{Degradation Recording}$ and $\textit{Distribution Backtracking}$. $\textit{Degradation Recording}$ is designed to obtain the convergence trajectory by recording the degradation path from the pre-trained teacher model to the untrained student generator. The degradation path implicitly represents the intermediate distributions between the teacher and the student, and its reverse can be viewed as the convergence trajectory from the student generator to the teacher model. Then $\textit{Distribution Backtracking}$ trains the student generator to backtrack the intermediate distributions along the path to approximate the convergence trajectory of the teacher model. Extensive experiments show that DisBack achieves faster and better convergence than the existing distillation method and achieves comparable or better generation performance, with an FID score of 1.38 on the ImageNet 64$\times$64 dataset. DisBack is easy to implement and can be generalized to existing distillation methods to boost performance.

The Cambrian explosion of easily accessible pre-trained diffusion models suggests a demand for methods that combine multiple different pre-trained diffusion models without incurring the significant computational burden of re-training a larger combined model. In this paper, we cast the problem of combining multiple pre-trained diffusion models at the generation stage under a novel proposed framework termed superposition. Theoretically, we derive superposition from rigorous first principles stemming from the celebrated continuity equation and design two novel algorithms tailor-made for combining diffusion models in SuperDiff. SuperDiff leverages a new scalable Itô density estimator for the log likelihood of the diffusion SDE which incurs *no additional overhead* compared to the well-known Hutchinson's estimator needed for divergence calculations. We demonstrate that SuperDiff is scalable to large pre-trained diffusion models as superposition is performed *solely through composition during inference*, and also enjoys painless implementation as it combines different pre-trained vector fields through an automated re-weighting scheme. Notably, we show that SuperDiff is efficient during inference time, and mimics traditional composition operators such as the logical OR and the logical AND. We empirically demonstrate the utility of using SuperDiff for generating more diverse images on CIFAR-10, more faithful prompt conditioned image editing using Stable Diffusion, as well as improved conditional molecule generation and unconditional *de novo* structure design of proteins. https://github.com/necludov/super-diffusion

Sampling from diffusion probabilistic models (DPMs) is often expensive for high-quality image generation and typically requires many steps with a large model. In this paper, we introduce sampling Trajectory Stitching (T-Stitch), a simple yet efficient technique to improve the sampling efficiency with little or no generation degradation. Instead of solely using a large DPM for the entire sampling trajectory, T-Stitch first leverages a smaller DPM in the initial steps as a cheap drop-in replacement of the larger DPM and switches to the larger DPM at a later stage. Our key insight is that different diffusion models learn similar encodings under the same training data distribution and smaller models are capable of generating good global structures in the early steps. Extensive experiments demonstrate that T-Stitch is training-free, generally applicable for different architectures, and complements most existing fast sampling techniques with flexible speed and quality trade-offs. On DiT-XL, for example, 40% of the early timesteps can be safely replaced with a 10x faster DiT-S without performance drop on class-conditional ImageNet generation. We further show that our method can also be used as a drop-in technique to not only accelerate the popular pretrained stable diffusion (SD) models but also improve the prompt alignment of stylized SD models from the public model zoo. Finally, the explicit model allocation strategy of T-Stitch significantly reduces the need of training or searching, delivering high deployment efficiency.

The rapid advancement of text-to-image (T2I) diffusion models has enabled them to generate unprecedented results from given texts. However, as text inputs become longer, existing encoding methods like CLIP face limitations, and aligning the generated images with long texts becomes challenging. To tackle these issues, we propose LongAlign, which includes a segment-level encoding method for processing long texts and a decomposed preference optimization method for effective alignment training. For segment-level encoding, long texts are divided into multiple segments and processed separately. This method overcomes the maximum input length limits of pretrained encoding models. For preference optimization, we provide decomposed CLIP-based preference models to fine-tune diffusion models. Specifically, to utilize CLIP-based preference models for T2I alignment, we delve into their scoring mechanisms and find that the preference scores can be decomposed into two components: a text-relevant part that measures T2I alignment and a text-irrelevant part that assesses other visual aspects of human preference. Additionally, we find that the text-irrelevant part contributes to a common overfitting problem during fine-tuning. To address this, we propose a reweighting strategy that assigns different weights to these two components, thereby reducing overfitting and enhancing alignment. After fine-tuning $512 \\times 512$ Stable Diffusion (SD) v1.5 for about 20 hours using our method, the fine-tuned SD outperforms stronger foundation models in T2I alignment, such as PixArt-$\\alpha$ and Kandinsky v2.2. The code is available at https://github.com/luping-liu/LongAlign.

Flow matching (FM) has gained significant attention as a simulation-free generative model. Unlike diffusion models, which are based on stochastic differential equations, FM employs a simpler approach by solving an ordinary differential equation with an initial condition from a normal distribution, thus streamlining the sample generation process. This paper discusses the convergence properties of FM in terms of the $p$-Wasserstein distance, a measure of distributional discrepancy. We establish that FM can achieve an almost minimax optimal convergence rate for $1 \leq p \leq 2$, presenting the first theoretical evidence that FM can reach convergence rates comparable to those of diffusion models. Our analysis extends existing frameworks by examining a broader class of mean and variance functions for the vector fields and identifies specific conditions necessary to attain these optimal rates.

This paper explores the challenge of accelerating the sequential inference process of Diffusion Probabilistic Models (DPMs). We tackle this critical issue from a dynamic systems perspective, in which the inherent sequential nature is transformed into a parallel sampling process. Specifically, we propose a unified framework that generalizes the sequential sampling process of DPMs as solving a system of banded nonlinear equations. Under this generic framework, we reveal that the Jacobian of the banded nonlinear equations system possesses a unit-diagonal structure, enabling further approximation for acceleration. Moreover, we theoretically propose an effective initialization approach for parallel sampling methods. Finally, we construct \textit{ParaSolver}, a hierarchical parallel sampling technique that enhances sampling speed without compromising quality. Extensive experiments show that ParaSolver achieves up to \textbf{12.1× speedup} in terms of wall-clock time. The source code is publicly available at https://github.com/Jianrong-Lu/ParaSolver.git.

Recent advancements in reinforcement learning (RL) have achieved great success in fine-tuning diffusion-based generative models. However, fine-tuning continuous flow-based generative models to align with arbitrary user-defined reward functions remains challenging, particularly due to issues such as policy collapse from overoptimization and the prohibitively high computational cost of likelihoods in continuous-time flows. In this paper, we propose an easy-to-use and theoretically sound RL fine-tuning method, which we term Online Reward-Weighted Conditional Flow Matching with Wasserstein-2 Regularization (ORW-CFM-W2). Our method integrates RL into the flow matching framework to fine-tune generative models with arbitrary reward functions, without relying on gradients of rewards or filtered datasets. By introducing an online reward-weighting mechanism, our approach guides the model to prioritize high-reward regions in the data manifold. To prevent policy collapse and maintain diversity, we incorporate Wasserstein-2 (W2) distance regularization into our method and derive a tractable upper bound for it in flow matching, effectively balancing exploration and exploitation of policy optimization. We provide theoretical analyses to demonstrate the convergence properties and induced data distributions of our method, establishing connections with traditional RL algorithms featuring Kullback-Leibler (KL) regularization and offering a more comprehensive understanding of the underlying mechanisms and learning behavior of our approach. Extensive experiments on tasks including target image generation, image compression, and text-image alignment demonstrate the effectiveness of our method, where our method achieves optimal policy convergence while allowing controllable trade-offs between reward maximization and diversity preservation.

Precise and flexible image editing remains a fundamental challenge in computer vision. Based on the modified areas, most editing methods can be divided into two main types: global editing and local editing. In this paper, we choose the two most common editing approaches (\ie text-based editing and drag-based editing) and analyze their drawbacks. Specifically, text-based methods often fail to describe the desired modifications precisely, while drag-based methods suffer from ambiguity. To address these issues, we proposed \textbf{CLIPDrag}, a novel image editing method that is the first to combine text and drag signals for precise and ambiguity-free manipulations on diffusion models. To fully leverage these two signals, we treat text signals as global guidance and drag points as local information. Then we introduce a novel global-local motion supervision method to integrate text signals into existing drag-based methods by adapting a pre-trained language-vision model like CLIP. Furthermore, we also address the problem of slow convergence in CLIPDrag by presenting a fast point-tracking method that enforces drag points moving toward correct directions. Extensive experiments demonstrate that CLIPDrag outperforms existing single drag-based methods or text-based methods.

The increasing size of screening libraries poses a significant challenge for the development of virtual screening methods for drug discovery, necessitating a re-evaluation of traditional approaches in the era of big data. Although 3D pharmacophore screening remains a prevalent technique, its application to very large datasets is limited by the computational cost associated with matching query pharmacophores to database molecules. In this study, we introduce PharmacoMatch, a novel contrastive learning approach based on neural subgraph matching. Our method reinterprets pharmacophore screening as an approximate subgraph matching problem and enables efficient querying of conformational databases by encoding query-target relationships in the embedding space. We conduct comprehensive investigations of the learned representations and evaluate PharmacoMatch as pre-screening tool in a zero-shot setting. We demonstrate significantly shorter runtimes and comparable performance metrics to existing solutions, providing a promising speed-up for screening very large datasets.

Recent text-to-image diffusion models leverage cross-attention layers, which have been effectively utilized to enhance a range of visual generative tasks. However, our understanding of cross-attention layers remains somewhat limited. In this study, we introduce a mechanistic interpretability approach for diffusion models by constructing Head Relevance Vectors (HRVs) that align with human-specified visual concepts. An HRV for a given visual concept has a length equal to the total number of cross-attention heads, with each element representing the importance of the corresponding head for the given visual concept. To validate HRVs as interpretable features, we develop an ordered weakening analysis that demonstrates their effectiveness. Furthermore, we propose concept strengthening and concept adjusting methods and apply them to enhance three visual generative tasks. Our results show that HRVs can reduce misinterpretations of polysemous words in image generation, successfully modify five challenging attributes in image editing, and mitigate catastrophic neglect in multi-concept generation. Overall, our work provides an advancement in understanding cross-attention layers and introduces new approaches for fine-controlling these layers at the head level.

Referring expression generation (REG) and comprehension (REC) are vital and complementary in joint visual and textual reasoning. Existing REC datasets typically contain insufficient image-expression pairs for training, hindering the generalization of REC models to unseen referring expressions. Moreover, REG methods frequently struggle to bridge the visual and textual domains due to the limited capacity, leading to low-quality and restricted diversity in expression generation. To address these issues, we propose a novel VIsion-guided Expression Diffusion Model (VIE-DM) for the REG task, where diverse synonymous expressions adhering to both image and text contexts of the target object are generated to augment REC datasets. VIE-DM consists of a vision-text condition (VTC) module and a transformer decoder. Our VTC and token selection design effectively addresses the feature discrepancy problem prevalent in existing REG methods. This enables us to generate high-quality, diverse synonymous expressions that can serve as augmented data for REC model learning. Extensive experiments on five datasets demonstrate the high quality and large diversity of our generated expressions. Furthermore, the augmented image-expression pairs consistently enhance the performance of existing REC models, achieving state-of-the-art results.

Recent developments in 2D visual generation have been remarkably successful. However, 3D and 4D generation remain challenging in real-world applications due to the lack of large-scale 4D data and effective model design. In this paper, we propose to jointly investigate general 3D and 4D generation by leveraging camera and object movements commonly observed in daily life. Due to the lack of real-world 4D data in the community, we first propose a data curation pipeline to obtain camera poses and object motion strength from videos. Based on this pipeline, we introduce a large-scale real-world 4D scene dataset: CamVid-30K. By leveraging all the 3D and 4D data, we develop our framework, GenXD, which allows us to produce any 3D or 4D scene. We propose multiview-temporal modules, which disentangle camera and object movements, to seamlessly learn from both 3D and 4D data. Additionally, GenXD employs masked latent conditions to support a variety of conditioning views. GenXD can generate videos that follow the camera trajectory as well as consistent 3D views that can be lifted into 3D representations. We perform extensive evaluations across various real-world and synthetic datasets, demonstrating GenXD's effectiveness and versatility compared to previous methods in 3D and 4D generation.

We present TexTailor, a novel method for generating consistent object textures from textual descriptions. Existing text-to-texture synthesis approaches utilize depth-aware diffusion models to progressively generate images and synthesize textures across predefined multiple viewpoints. However, these approaches lead to a gradual shift in texture properties across viewpoints due to (1) insufficient integration of previously synthesized textures at each viewpoint during the diffusion process and (2) the autoregressive nature of the texture synthesis process. Moreover, the predefined selection of camera positions, which does not account for the object's geometry, limits the effective use of texture information synthesized from different viewpoints, ultimately degrading overall texture consistency. In TexTailor, we address these issues by (1) applying a resampling scheme that repeatedly integrates information from previously synthesized textures within the diffusion process, and (2) fine-tuning a depth-aware diffusion model on these resampled textures. During this process, we observed that using only a few training images restricts the model's original ability to generate high-fidelity images aligned with the conditioning, and therefore propose an performance preservation loss to mitigate this issue. Additionally, we improve the synthesis of view-consistent textures by adaptively adjusting camera positions based on the object's geometry. Experiments on a subset of the Objaverse dataset and the ShapeNet car dataset demonstrate that TexTailor outperforms state-of-the-art methods in synthesizing view-consistent textures.

Diffusion Probabilistic Models (DPMs) have achieved significant success in generative tasks. However, their training and sampling processes suffer from the issue of distribution mismatch. During the denoising process, the input data distributions differ between the training and inference stages, potentially leading to inaccurate data generation. To obviate this, we analyze the training objective of DPMs and theoretically demonstrate that this mismatch can be alleviated through Distributionally Robust Optimization (DRO), which is equivalent to performing robustness-driven Adversarial Training (AT) on DPMs. Furthermore, for the recently proposed Consistency Model (CM), which distills the inference process of the DPM, we prove that its training objective also encounters the mismatch issue. Fortunately, this issue can be mitigated by AT as well. Based on these insights, we propose to conduct efficient AT on both DPM and CM. Finally, extensive empirical studies validate the effectiveness of AT in diffusion-based models. The code is available at https://github.com/kugwzk/AT_Diff.

Current image watermarking methods are vulnerable to advanced image editing techniques enabled by large-scale text-to-image models. These models can distort embedded watermarks during editing, posing significant challenges to copyright protection. In this work, we introduce W-Bench, the first comprehensive benchmark designed to evaluate the robustness of watermarking methods against a wide range of image editing techniques, including image regeneration, global editing, local editing, and image-to-video generation. Through extensive evaluations of eleven representative watermarking methods against prevalent editing techniques, we demonstrate that most methods fail to detect watermarks after such edits. To address this limitation, we propose VINE, a watermarking method that significantly enhances robustness against various image editing techniques while maintaining high image quality. Our approach involves two key innovations: (1) we analyze the frequency characteristics of image editing and identify that blurring distortions exhibit similar frequency properties, which allows us to use them as surrogate attacks during training to bolster watermark robustness; (2) we leverage a large-scale pretrained diffusion model SDXL-Turbo, adapting it for the watermarking task to achieve more imperceptible and robust watermark embedding. Experimental results show that our method achieves outstanding watermarking performance under various image editing techniques, outperforming existing methods in both image quality and robustness. Code is available at https://github.com/Shilin-LU/VINE

Dataset distillation aims to streamline the training process by creating a compact yet effective dataset for a much larger original dataset. However, existing methods often struggle with distilling large, high-resolution datasets due to prohibitive resource costs and limited performance, primarily stemming from sample-wise optimizations in the pixel space. Motivated by the remarkable capabilities of diffusion generative models in learning target dataset distributions and controllably sampling high-quality data tailored to user needs, we propose framing dataset distillation as a controlled diffusion generation task aimed at generating data specifically tailored for effective training purposes. By establishing a correlation between the overarching objective of dataset distillation and the trajectory influence function, we introduce the Influence-Guided Diffusion (IGD) sampling framework to generate training-effective data without the need to retrain diffusion models. An efficient guided function is designed by leveraging the trajectory influence function as an indicator to steer diffusions to produce data with influence promotion and diversity enhancement. Extensive experiments show that the training performance of distilled datasets generated by diffusions can be significantly improved by integrating with our IGD method and achieving state-of-the-art performance in distilling ImageNet datasets. Particularly, an exceptional result is achieved on the ImageNet-1K, reaching 60.3\% at IPC=50. Our code is available at https://github.com/mchen725/DD_IGD.