CVPR 2025 Papers

In this paper, we introduce a novel self-distillation paradigm for improving the performance of diffusion models. Previous studies have shown that introducing a teacher to distill the diffusion model can enhance its sampling efficiency. We raise an intriguing question: can the diffusion model itself serve as its teacher to further improve the performance of itself? To this end, we propose a new paradigm called Self Step-Distillation (SSD). The core idea of SSD is to integrate the predictions or the intermediate activations of the diffusion model at each timestep with its preceding timestep through a fusion mechanism. We propose two forms, explicit SSD and implicit SSD (iSSD), to perform N-step to N-step distillation from the diffusion model itself to achieve improved image quality. We further elucidate the relationship between SSD and high-order solver, highlighting their underlying relationship. The effectiveness of SSD is validated through extensive experiments on diffusion transformers of various sizes and across different sampling steps. Our results show that this novel self-distillation paradigm can significantly enhance performance. Additionally, our method is compatible with the distillation method designed for few-step inference. Notably, with iSSD trained less than one epoch, we obtain a 32-step DiT-XL/2 achieving an FID of 1.99, outperforming the original 250-step DiT-XL/2 with an FID of 2.26. We further validate the effectiveness of our method on text-to-image diffusion models, such as Stable Diffusion, and also observe notable improvement in image quality.

The modality gap between vision and text embeddings in CLIP presents a significant challenge for zero-shot image captioning, limiting effective cross-modal representation. Traditional approaches, such as noise injection and memory-based similarity matching, attempt to address this gap, yet these methods either rely on indirect alignment or relatively naive solutions with heavy computation. Diffusion Bridge introduces a novel approach to directly reduce this modality gap by leveraging Denoising Diffusion Probabilistic Models (DDPM), trained exclusively on text embeddings to model their distribution. Our approach is motivated by the observation that, while paired vision and text embeddings are relatively close, a modality gap still exists due to stable regions created by the contrastive loss. This gap can be interpreted as noise in cross-modal mappings, which we approximate as Gaussian noise. To bridge this gap, we employ a reverse diffusion process, where image embeddings are strategically introduced at an intermediate step in the reverse process, allowing them to be refined progressively toward the text embedding distribution. This process transforms vision embeddings into text-like representations closely aligned with paired text embeddings, effectively minimizing discrepancies between modalities. Experimental results demonstrate that these text-like vision embeddings significantly enhance alignment with their paired text embeddings, leading to improved zero-shot captioning performance on MSCOCO and Flickr30K. Diffusion Bridge achieves competitive results without reliance on memory banks or entity-driven methods, offering a novel pathway for cross-modal alignment and opening new possibilities for the application of diffusion models in multi-modal tasks. The source code is available at: https://github.com/mongeoroo/diffusion-bridge

How can we generate an image B' that satisfies A:A'::B:B', given the input images A,A' and B? Recent works have tackled this challenge through approaches like visual in-context learning or visual instruction. However, these methods are typically limited to specific models (InstructPix2Pix. Inpainting models) rather than general diffusion models (Stable Diffusion, SDXL). This dependency may lead to inherited biases or lower editing capabilities. In this paper, we propose Difference Inversion, a method that isolates only the difference from A and A' and applies it to B to generate a plausible B'. To address model dependency, it is crucial to structure prompts in the form of a "Full Prompt" suitable for input to stable diffusion models, rather than using an "Instruction Prompt". To this end, we accurately extract the Difference between A and A' and combine it with the prompt of B, enabling a plug-and-play application of the difference. To extract a precise difference, we first identify it through 1) Delta Interpolation. Additionally, to ensure accurate training, we propose the 2) Token Consistency Loss and 3) Zero Initialization of Token Embeddings. Our extensive experiments demonstrate that Difference Inversion outperforms existing baselines both quantitatively and qualitatively, indicating its ability to generate more feasible B' in a model-agnostic manner.

Top-leading solutions for Video Scene Graph Generation (VSGG) typically adopt an offline pipeline.Though demonstrating promising performance, they remain unable to handle real-time video streams and consume large GPU memory. Moreover, these approaches fall short in temporal reasoning, merely aggregating frame-level predictions over a temporal context. In response, we introduce DiffVsgg, an online VSGG solution that frames this task as an iterative scene graph update problem. Drawing inspiration from Latent Diffusion Models (LDMs) which generate images via denoising a latent feature embedding, we unify the decoding of object classification, bounding box regression, and graph generation three tasks using one shared feature embedding. Then, given an embedding containing unified features of object pairs, we conduct a step-wise Denoising on it within LDMs, so as to deliver a clean embedding which clearly indicates the relationships between objects.This embedding then serves as the input to task-specific heads for object classification, scene graph generation, etc. DiffVsgg further facilitates continuous temporal reasoning, where predictions for subsequent frames leverage results of past frames as the conditional inputs of LDMs, to guide the reverse diffusion process for current frames.Extensive experiments on three setups of Action Genome demonstrate the superiority of DiffVsgg.

Story visualization, the task of creating visual narratives from textual descriptions, has seen progress with text-to-image generation models. However, these models often lack effective control over character appearances and interactions, particularly in multi-character scenes. To address these limitations, we propose a new task: customized manga generation and introduce DiffSensei, an innovative framework specifically designed for generating manga with dynamic multi-character control. DiffSensei integrates a diffusion-based image generator with a multimodal large language model (MLLM) that acts as a text-compatible identity adapter. Our approach employs masked cross-attention to seamlessly incorporate character features, enabling precise layout control without direct pixel transfer. Additionally, the MLLM-based adapter adjusts character features to align with panel-specific text cues, allowing flexible adjustments in character expressions, poses, and actions. We also introduce MangaZero, a large-scale dataset tailored to this task, containing 43,264 manga pages and 427,147 annotated panels, supporting the visualization of varied character interactions and movements across sequential frames. Extensive experiments demonstrate that DiffSensei outperforms existing models, marking a significant advancement in manga generation by enabling text-adaptable character customization. The code, model, and dataset will be open-sourced to the community.

Generating high-quality 360-degree views of human heads from single-view images is essential for enabling accessible immersive telepresence applications and scalable personalized content creation.While cutting-edge methods for full head generation are limited to modeling realistic human heads, the latest diffusion-based approaches for style-omniscient head synthesis can produce only frontal views and struggle with view consistency, preventing their conversion into true 3D models for rendering from arbitrary angles.We introduce a novel approach that generates fully consistent 360-degree head views, accommodating human, stylized, and anthropomorphic forms, including accessories like glasses and hats. Our method builds on the DiffPortrait3D framework, incorporating a custom ControlNet for back-of-head detail generation and a dual appearance module to ensure global front-back consistency. By training on continuous view sequences and integrating a back reference image, our approach achieves robust, locally continuous view synthesis. Our model can be used to produce high-quality neural radiance fields (NeRFs) for real-time, free-viewpoint rendering, outperforming state-of-the-art methods in object synthesis and 360-degree head generation for very challenging input portraits.

We address the task of generating 3D hair geometry from a single image, which is challenging due to the diversity of hairstyles and the lack of paired image-to-3D hair data. Previous methods are primarily trained on synthetic data and cope with the limited amount of such data by using low-dimensional intermediate representations, such as guide strands and scalp-level embeddings, that require post-processing to decode, upsample, and add realism. These approaches fail to reconstruct detailed hair, struggle with curly hair, or are limited to handling only a few hairstyles. To overcome these limitations, we propose DiffLocks, a novel framework that enables detailed reconstruction of a wide variety of hairstyles directly from a single image. First, we address the lack of 3D hair data by automating the creation of the largest synthetic hair dataset to date, containing 40K hairstyles. Second, we leverage the synthetic hair dataset to learn an image-conditioned diffusion-transfomer model that generates accurate 3D strands from a single frontal image. By using a pretrained image backbone, our method generalizes to in-the-wild images despite being trained only on synthetic data. Our diffusion model predicts a scalp texture map in which any point in the map contains the latent code for an individual hair strand. These codes are directly decoded to 3D strands without post-processing techniques. Representing individual strands, instead of guide strands, enables the transformer to model the detailed spatial structure of complex hairstyles. With this, DiffLocks can recover highly curled hair, like afro hairstyles, from a single image for the first time. Data and code is available at https://radualexandru.github.io/difflocks

We introduce DiffFNO, a novel diffusion framework for arbitrary-scale super-resolution strengthened by a Weighted Fourier Neural Operator (WFNO). Mode Rebalancing in WFNO effectively captures critical frequency components, significantly improving the reconstruction of high-frequency image details that are crucial for super-resolution tasks. Gated Fusion Mechanism (GFM) adaptively complements WFNO's spectral features with spatial features from an Attention-based Neural Operator (AttnNO). This enhances the network's capability to capture both global structures and local details. Adaptive Time-Step (ATS) ODE solver, a deterministic sampling strategy, accelerates inference without sacrificing output quality by dynamically adjusting integration step sizes ATS. Extensive experiments demonstrate that DiffFNO achieves state-of-the-art (SOTA) results, outperforming existing methods across various scaling factors by a margin of 2-4 dB in PSNR, including those beyond the training distribution. It also achieves this at lower inference time. Our approach sets a new standard in super-resolution, delivering both superior accuracy and computational efficiency.

Diffusion models have revolutionized generative tasks through high-fidelity outputs, yet flow matching (FM) offers faster inference and empirical performance gains. However, current foundation FM models are computationally prohibitive for finetuning, while diffusion models like Stable Diffusion benefit from efficient architectures and ecosystem support. This work addresses the critical challenge of efficiently transferring knowledge from pre-trained diffusion models to flow matching. We propose Diff2Flow, a novel framework that systematically bridges diffusion and FM paradigms by rescaling timesteps, aligning interpolants, and deriving FM-compatible velocity fields from diffusion predictions. This alignment enables direct and efficient FM finetuning of diffusion priors with no extra computation overhead. Our experiments demonstrate that Diff2Flow outperforms naive FM and diffusion finetuning particularly under parameter-efficient constraints, while achieving superior or competitive performance across diverse downstream tasks compared to state-of-the-art methods.

Palmprint recognition is significantly limited by the lack of large-scale publicly available datasets. Previous methods have adopted Bezier curves to simulate the palm creases, which then serve as input for conditional GANs to generate realistic palmprints.However, without employing real data fine-tuning, the performance of the recognition model trained on these synthetic datasets would drastically decline, indicating a large gap between generated and real palmprints.This is primarily due to the utilization of an inaccurate palm crease representation and challenges in balancing intra-class variation with identity consistency.To address this, we introduce a polynomial-based palm crease representation that provides a new palm crease generation mechanism more closely aligned with the real distribution. We also propose the palm creases conditioned diffusion model with a novel intra-class variation control method.By applying our proposed K-step noise-sharing sampling, we are able to synthesize palmprint datasets with large intra-class variation and high identity consistency.Experimental results show that, for the first time, recognition models trained solely on our synthetic datasets, without any fine-tuning, outperform those trained on real datasets.Furthermore, our approach achieves superior recognition performance as the number of generated identities increases.Our code and pre-trained models will be released.

Infrared imaging is essential for autonomous driving and robotic operations as a supportive modality due to its reliable performance in challenging environments. Despite its popularity, the limitations of infrared cameras, such as low spatial resolution and complex degradations, consistently challenge imaging quality and subsequent visual tasks. Hence, infrared image super-resolution (IISR) has been developed to address this challenge. While recent developments in diffusion models have greatly advanced this field, current methods to solve it either ignore the unique modal characteristics of infrared imaging or overlook the machine perception requirements. To bridge these gaps, we propose DifIISR, an infrared image super-resolution diffusion model optimized for visual quality and perceptual performance. Our approach achieves task-based guidance for diffusion by injecting gradients derived from visual and perceptual priors into the noise during the reverse process. Specifically, we introduce an infrared thermal spectrum distribution regulation to preserve visual fidelity, ensuring that the reconstructed infrared images closely align with high-resolution images by matching their frequency components. Subsequently, we incorporate various visual foundational models as the perceptual guidance for downstream visual tasks, infusing generalizable perceptual features beneficial for detection and segmentation. As a result, our approach gains superior visual results while attaining State-Of-The-Art downstream task performance.

Sora-like video generation models have achieved remarkable progress with a Multi-Modal Diffusion Transformer (MM-DiT) architecture. However, the current video generation models predominantly focus on single-prompt, struggling to generate coherent scenes with multiple sequential prompts that better reflect real-world dynamic scenarios. While some pioneering works have explored multi-prompt video generation, they face significant challenges including strict training data requirements, weak prompt following, and unnatural transitions. To address these problems, we propose DiTCtrl, a training-free multi-prompt video generation method under MM-DiT architectures for the first time. Our key idea is to take the multi-prompt video generation task as temporal video editing with smooth transitions. To achieve this goal, we first analyze MM-DiT's attention mechanism, finding that the 3D full attention behaves similarly to that of the cross/self-attention blocks in the UNet-like diffusion models, enabling mask-guided precise semantic control across different prompts with attention sharing for multi-prompt video generation. Based on our careful design, the video generated by DiTCtrl achieves smooth transitions and consistent object motion given multiple sequential prompts without additional training. Besides, we also present MPVBench, a new benchmark specially designed for multi-prompt video generation to evaluate the performance of multi-prompt generation. Extensive experiments demonstrate that our method achieves state-of-the-art performance without additional training.

In-bed human mesh recovery can be crucial and enabling for several healthcare applications, including sleep pattern monitoring, rehabilitation support, and pressure ulcer prevention. However, it is difficult to collect large real-world visual datasets in this domain, in part due to privacy and expense constraints, which in turn presents significant challenges for training and deploying deep learning models. Existing in-bed human mesh estimation methods often rely heavily on real-world data, limiting their ability to generalize across different in-bed scenarios, such as varying coverings and environmental settings. To address this, we propose a Sim-to-Real Transfer Framework for in-bed human mesh recovery from overhead depth images, which leverages large-scale synthetic data alongside limited or no real-world samples. We introduce a diffusion model that bridges the gap between synthetic data and real data to support generalization in real-world in-bed pose and body inference scenarios. Extensive experiments and ablation studies validate the effectiveness of our framework, demonstrating significant improvements in robustness and adaptability across diverse healthcare scenarios. Project page can be found at https://jing-g2.github.io/DiSRT-In-Bed/.

Medical Visual Question Answering (Med-VQA) systems benefit the interpretation of medical images containing critical clinical information. However, the challenge of noisy labels and limited high-quality datasets remains underexplored. To address this, we establish the first benchmark for noisy labels in Med-VQA by simulating human mislabeling with semantically designed noise types. More importantly, we introduce the DiN framework, which leverages a diffusion model to handle noisy labels in Med-VQA. Unlike the dominant classification-based VQA approaches that directly predict answers, our Answer Diffuser (AD) module employs a coarse-to-fine process, refining answer candidates with a diffusion model for improved accuracy. The Answer Condition Generator (ACG) further enhances this process by generating task-specific conditional information via integrating answer embeddings with fused image-question features. To address label noise, our Noisy Label Refinement(NLR) module introduces a robust loss function and dynamic answer adjustment to further boost the performance of the AD module. Our DiN framework consistently outperforms existing methods across multiple benchmarks with varying noise levels.

Diffusion models with large-scale pre-training have achieved significant success in the field of visual content generation, particularly exemplified by Diffusion Transformers (DiT). However, DiT models have faced challenges with quadratic complexity efficiency, especially when handling long sequences. In this paper, we aim to incorporate the sub-quadratic modeling capability of Gated Linear Attention (GLA) into the 2D diffusion backbone. Specifically, we introduce Diffusion Gated Linear Attention Transformers (DiG), a simple, adoptable solution with minimal parameter overhead. We offer two variants, i.e., a plain and U-shape architecture, showing superior efficiency and competitive effectiveness. In addition to superior performance to DiT and other sub-quadratic-time diffusion models at 256x256 resolution, DiG demonstrates greater efficiency than these methods starting from a 512 resolution. Specifically, DiG-S/2 is 2.5x faster and saves 75.7% GPU memory compared to DiT-S/2 at a 1792 resolution. Additionally, DiG-XL/2 is 4.2x faster than the Mamba-based model at a 1024 resolution and 1.8x faster than DiT with FlashAttention-2 at a 2048 resolution. The code is released at https://github.com/hustvl/DiG.

Diffusion models have shown exceptional performance in visual generation tasks. Recently, these models have shifted from traditional U-Shaped CNN-Attention hybrid structures to fully transformer-based isotropic architectures. While these transformers exhibit strong scalability and performance, their reliance on complicated self-attention operation results in slow inference speeds. Contrary to these works, we rethink one of the simplest yet fastest module in deep learning, 3x3 Convolution, to construct a scaled-up purely convolutional diffusion model. We first discover that an Encoder-Decoder Hourglass design outperforms scalable isotropic architectures for Conv3x3, but still under-performing our expectation. Further improving the architecture, we introduce sparse skip connections to reduce redundancy and improve scalability. Based on the architecture, we introduce conditioning improvements including stage-specific embeddings, mid-block condition injection, and conditional gating. These improvements lead to our proposed Diffusion CNN (DiC), which serves as a swift yet competitive diffusion architecture baseline. Experiments on various scales and settings show that DiC surpasses existing diffusion transformers by considerable margins in terms of performance while keeping a good speed advantage. Project page: https://github.com/YuchuanTian/DiC

Dexterous manipulation with contact-rich interactions is crucial for advanced robotics. While recent diffusion-based planning approaches show promise for simple manipulation tasks, they often produce unrealistic ghost states (e.g., the object automatically moves without hand contact) or lack adaptability when handling complex sequential interactions. In this work, we introduce DexHandDiff, an interaction-aware diffusion planning framework for adaptive dexterous manipulation. DexHandDiff models joint state-action dynamics through a dual-phase diffusion process which consists of pre-interaction contact alignment and post-contact goal-directed control, enabling goal-adaptive generalizable dexterous manipulation. Additionally, we incorporate dynamics model-based dual guidance and leverage large language models for automated guidance function generation, enhancing generalizability for physical interactions and facilitating diverse goal adaptation through language cues. Experiments on physical interaction tasks such as door opening, pen and block re-orientation, object relocation, and hammer striking demonstrate DexHandDiff's effectiveness on goals outside training distributions, achieving over twice the average success rate (59.2% vs. 29.5%) compared to existing methods. Our framework achieves an average of 70.7% success rate on goal adaptive dexterous tasks, highlighting its robustness and flexibility in contact-rich manipulation.

While large-scale text-to-image diffusion models enable the generation of high-quality, diverse images from text prompts, these prompts struggle to capture intricate details, such as textures, preventing the user intent from being reflected. This limitation has led to efforts to generate images conditioned on user-provided images, referred to as image prompts. Recent work modifies the self-attention mechanism to impose image conditions in generated images by replacing or concatenating the keys and values from the image prompt. This enables the self-attention layer to work like a cross-attention layer, generally used to incorporate text prompts.In this paper, we identify two common issues in existing methods of modifying self-attention that hinder diffusion models from reflecting the image prompt. By addressing these issues, we propose a novel method that generates images that properly reflect the details of image prompts. First, existing approaches often neglect the importance of image prompts in classifier-free guidance, which directs the model towards the intended conditions and away from those undesirable. Specifically, current methods use image prompts as both desired and undesired conditions, causing conflicting signals. To resolve this, we propose conflict-free guidance by using image prompts only as desired conditions, ensuring that the generated image faithfully reflects the image prompt.In addition, we observe that the two most common self-attention modifications involve a trade-off between the realism of the generated image and alignment with the image prompt, achieved by selectively using keys and values from both images. Specifically, selecting more keys and values from the image prompt improves alignment, while selecting more from the generated image enhances realism. To balance both, we propose an alternative self-attention modification method, Stratified Attention, which jointly uses keys and values from both images rather than selecting between them.Through extensive experiments across three distinct image generation tasks, we demonstrate that the proposed method outperforms existing image-prompting models in faithfully reflecting the image prompt.

Text-to-image diffusion models have achieved state-of-the-art results in synthesis tasks; however, there is a growing concern about their potential misuse in creating harmful content. To mitigate these risks, post-hoc model intervention techniques, such as concept unlearning and safety guidance, have been developed. However, fine-tuning model weights or adapting the hidden states of the diffusion model operates in an uninterpretable way, making it unclear which part of the intermediate variables is responsible for unsafe generation. These interventions severely affect the sampling trajectory when erasing harmful concepts from complex, multi-concept prompts, thus hindering their practical use in real-world settings. In this work, we propose the safe generation framework Detect-and-Guide (DAG), leveraging the internal knowledge of diffusion models to perform self-diagnosis and fine-grained self-regulation during the sampling process. DAG first detects harmful concepts from noisy latents using refined cross-attention maps of optimized tokens, then applies safety guidance with adaptive strength and editing regions to negate unsafe generation. The optimization only requires a small annotated dataset and can provide precise detection maps with generalizability and concept specificity. Moreover, DAG does not require fine-tuning of diffusion models, and therefore introduces no loss to their generation diversity. Experiments on erasing sexual content show that DAG achieves state-of-the-art safe generation performance, balancing harmfulness mitigation and text-following performance on multi-concept real-world prompts.