We introduce \textit{Preserve Anything}, a novel method for controlled image synthesis that addresses key limitations in object preservation and semantic consistency in text-to-image (T2I) generation. Existing approaches often fail (i) to preserve multiple objects with fidelity, (ii) maintain semantic alignment with prompts, or (iii) provide explicit control over scene composition. To overcome these challenges, the proposed method employs an N-channel ControlNet that integrates (i) object preservation with size and placement agnosticism, color and detail retention, and artifact elimination, (ii) high-resolution, semantically consistent backgrounds with accurate shadows, lighting, and prompt adherence, and (iii) explicit user control over background layouts and lighting conditions. Key components of our framework include object preservation and background guidance modules, enforcing lighting consistency and a high-frequency overlay module to retain fine details while mitigating unwanted artifacts. We introduce a benchmark dataset consisting of 240K natural images filtered for aesthetic quality and 18K 3D-rendered synthetic images with metadata such as lighting, camera angles, and object relationships. This dataset addresses the deficiencies of existing benchmarks and allows a complete evaluation. Empirical results demonstrate that our method achieves state-of-the-art performance, significantly improving feature-space fidelity (FID 15.26) and semantic alignment (CLIP-S 32.85) while maintaining competitive aesthetic quality. We also conducted a user study to demonstrate the efficacy of the proposed work on unseen benchmark and observed a remarkable improvement of $\sim25\%$, $\sim19\%$, $\sim13\%$, and $\sim14\%$ in terms of prompt alignment, photorealism, the presence of AI artifacts, and natural aesthetics over existing works.

Quantizing the floating-point weights and activations of deep convolutional neural networks to fixed-point representation yields reduced memory footprints and inference time. Recently, efforts have been afoot towards zero-shot quantization that does not require original unlabelled training samples of a given task. These best-published works heavily rely on the learned batch normalization (BN) parameters to infer the range of the activations for quantization. In particular, these methods are built upon either empirical estimation framework or the data distillation approach, for computing the range of the activations. However, the performance of such schemes severely degrades when presented with a network that does not accommodate BN layers. In this line of thought, we propose ageneralized zero-shot quantization(GZSQ) framework that neither requires original data nor relies on BN layer statistics. We have utilized the data distillation approach and leveraged only the pre-trained weights of the model to estimate enriched data for range calibration of the activations. To the best of our knowledge, this is the first work that utilizes the distribution of the pre-trained weights to assist the process of zero-shot quantization. The proposed scheme has significantly outperformed the existing zero-shot works,e.g., an improvement of$\sim$33% in classification accuracy for MobileNetV2 and several other models that are w & w/o BN layers, for a variety of tasks. We have also demonstrated the efficacy of the proposed work across multiple open-source quantization frameworks. Importantly, our work is the first attempt towards the post-training zero-shot quantization of futuristic unnormalized deep neural networks.

Background: Underwater images, in general, suffer from low contrast and high color distortions due to the non-uniform attenuation of the light as it propagates through the water. In addition, the degree of attenuation varies with the wavelength, resulting in the asymmetric traversing of colors. Despite the prolific works for underwater image restoration (UIR) using deep learning, the above asymmetricity has not been addressed in the respective network engineering. Contributions: As the first novelty, this article shows that attributing the right receptive field size ( context ) based on the traversing range of the color channel may lead to a substantial performance gain for the task of UIR. Further, it is important to suppress the irrelevant multi-contextual features and increase the representational power of the model. Therefore, as a second novelty, we have incorporated an attentive skip mechanism to adaptively refine the learned multi-contextual features. The proposed framework, called Deep WaveNet , is optimized using the traditional pixel-wise and feature-based cost functions. An extensive set of experiments have been carried out to show the efficacy of the proposed scheme over existing best-published literature on benchmark datasets. More importantly, we have demonstrated a comprehensive validation of enhanced images across various high-level vision tasks, e.g., underwater image semantic segmentation and diver’s 2D pose estimation. A sample video to exhibit our real-world performance is available at https://tinyurl.com/yzcrup9n . Also, we have open-sourced our framework at https://github.com/pksvision/Deep-WaveNet-Underwater-Image-Restoration .

In this article, we address the problem of rain-streak removal in the videos. Unlike the image, challenges in video restoration comprise temporal consistency besides spatial enhancement. The researchers across the world have proposed several effective methods for estimating the de-noised videos with outstanding temporal consistency. However, such methods also amplify the computational cost due to their larger size. By way of analysis, incorporating separate modules for spatial and temporal enhancement may require more computational resources. It motivates us to propose a unified architecture that directly estimates the de-rained frame with maximal visual quality and minimal computational cost. To this end, we present a deep learning-based Frame-recurrent Multi-contextual Adversarial Network for rain-streak removal in videos. The proposed model is built upon a Conditional Generative Adversarial Network (CGAN)-based framework where the generator model directly estimates the de-rained frame from the previously estimated one with the help of its multi-contextual adversary. To optimize the proposed model, we have incorporated the Perceptual loss function in addition to the conventional Euclidean distance. Also, instead of traditional entropy loss from the adversary, we propose to use the Euclidean distance between the features of de-rained and clean frames, extracted from the discriminator model as a cost function for video de-raining. Various experimental observations across 11 test sets, with over 10 state-of-the-art methods, using 14 image-quality metrics, prove the efficacy of the proposed work, both visually and computationally.

This paper presents a novel method for a single image rain streak removal problem which exploits the spatial as well as wavelet transformed coefficients of the rainy images. The proposed method adopts the Conditional Generative Adversarial Network [1] framework and consists of two following networks: Generator and Discriminator. The generator model receives the input from both spatial, frequency domain of the rainy image and yields five de-rained image candidates. A Deep Residual Network [2] has been used to merge these derained candidates and predict a single de-rained image. To ensure the visual quality of the de-rained image, Perceptual loss function [3] in addition to adversarial training has been incorporated. Extensive experiments on the synthetic and realworld rainy images dataset reveal an improvement over the existing state-of-the-art methods [4], [5] by ~ 1.08%, 2.57% in Structural Similarity Index [6] and ~ 7.39%, 9.95% in Peak signal-to-noise ratio respectively.

This paper presents a memorability based image-to-image translation technique to make an image more memorable while retaining its high-level contents. Conventionally, the image-to-image translation task aims to learn the mapping between images of two different domains using a set of aligned image pairs. However, dataset having such one-to-one mapping is not available for memorability based image-to-image translation. Therefore, the aim of the proposed task is defined to learn the mapping F: I → I' between two image domains I and I'. Here, I corresponds to input image domain and I' is the unknown image domain containing the modified version of the input images. Also, every image in I' is more memorable than its corresponding image in I. Therefore, the proposed task is achieved by developing a deep learning based method to learn the mapping F: I→ I' using mean-squared error and memorability loss between I and F(I). The experimental results showed that the proposed approach increases the memorability of the given image better than the state-of-the-art image-to-image translation techniques.