Kshitij Pathania

Developed workflow orchestration system for Adobe Localization Team, managing complex state transitions and processes
to streamline global annual content releases for 12+ Adobe products.
Engineered automated deployment pipelines ensuring seamless and reliable service deployment across Kubernetes clusters.
Developed system for Multimodal Incongruence Detection (MID) Model, integrating model serving pipeline to handle
multiple models across various modalities, reducing latency by 70% (40 ms) and boosting QPS to 3 times the prior rate.

In the realm of image synthesis, achieving fidelity to a reference image while adhering to conditional prompts remains a significant challenge. This paper proposes a novel approach that integrates a diffusion model with latent space manipulation and gradient-based selective attention mechanisms to address this issue. Leveraging Grad-SAM (Gradient-based Selective Attention Manipulation), we analyze the cross attention maps of the cross attention layers and gradients for the denoised latent vector, deriving importance scores of elements of denoised latent vector related to the subject of interest. Using this information, we create masks at specific timesteps during denoising to preserve subjects while seamlessly integrating the reference image features. This approach ensures the faithful formation of subjects based on conditional prompts, while concurrently refining the background for a more coherent composition. Our experiments on places365 dataset demonstrate promising results, with our proposed model achieving the lowest mean and median Frechet Inception Distance (FID) scores compared to baseline models, indicating superior fidelity preservation. Furthermore, our model exhibits competitive performance in aligning the generated images with provided textual descriptions, as evidenced by high CLIP scores. These results highlight the effectiveness of our approach in both fidelity preservation and textual context preservation, offering a significant advancement in text-to-image synthesis tasks.

In this paper, a novel method for analyzing the sentiments portrayed by Sanskrit text has been proposed. Sanskrit is one of the world’s most ancient languages; however, natural language processing tasks such as machine translation and sentiment analysis have not been explored for it to the full potential because of the unavailability of sufficient labeled data. We solved this issue using a zero-shot learning-based cross-lingual sentiment analysis (CLSA) approach. The CLSA uses the resources from the source language to enhance the sentiment analysis of the target language having insufficient resources. The proposed work translates the text from Sanskrit, a language with insufficient labeled data, to English, with sufficient labeled data for sentiment analysis using a transformer model. A generative adversarial network-based strategy has been proposed to evaluate the maturity of the translations. Then a bidirectional long short-term memory-based model has been implemented to classify the sentiments using the embeddings obtained through translations. The proposed technique has achieved 87.50% accuracy for machine translation and 92.83% accuracy for sentiment classification. Sanskrit-English translations used in this work have been collected through web scraping techniques. In the absence of the ground-truth sentiment class labels, a strategy for evaluating the sentiment scores of the proposed sentiment analysis model has also been presented. A new dataset of Sanskrit text, along with their English translations and sentiment scores, has been constructed.