Designed and implemented a multimodal pipeline integrating LoRA fine-tuned CLIP and Stable Diffusion v1.5, achieving 21% higher SSIM and 25% higher PSNR over baseline models in forensic sketch generation.

Fine-tuned CLIP using LoRA on self- and cross-attention layers, improving text-sketch alignment by 9% and reducing perceptual error (LPIPS) over iterations validated through ablation studies.

Developed an iterative refinement process, dynamically updating embeddings and prompts, enabling cumulative quality improvements in generated sketches over 5 refinement cycles, enhancing investigative reliability.