AXIOM: Parameter-Efficient Cultural Alignment of Latent Diffusion Models Under Strict 4GB Edge Constraints

Abstract

Foundational text-to-image latent diffusion architectures operate with systemic geographic and demographic blind spots. Pre-trained predominantly on Western and East Asian web-scale corpuses, these models default to homogenized, inaccurate stereotypes when prompted for the Global South. This paper presents the architecture and independent implementation of AXIOM: a high-fidelity generative vision model engineered to map the atmospheric, architectural, and demographic truth of Bangladesh into the latent space. Bypassing non-consensual web scraping, the model was trained on a proprietary, ethically curated dataset of 1,200+ high-resolution photographs spanning the riverine delta of Khulna, the coastlines of Cox’s Bazar, the montane cloud inversions of Sajek Valley, and consented familial portraits. The primary engineering bottleneck was executing high-dimensional tensor fine-tuning without cloud infrastructure, operating strictly on an entry-level consumer laptop restricted to 8GB System RAM and an NVIDIA RTX 3050 Laptop GPU (4GB VRAM). By architecting a memory-optimized pipeline integrating Low-Rank Adaptation (LoRA, r=128), selective gradient checkpointing, 8-bit AdamW quantization, and half-precision (FP16) mathematics, catastrophic CUDA Out-of-Memory (OOM) failures were bypassed to compute a 72-million parameter cross-attention update across 63,750 optimization steps. Validated at the national Calibration 2.0 showcase organized by Khulna University of Engineering & Technology (KUET), this independent single-researcher initiative (Tensor Solo) secured the 1st Place Championship in Project Showcasing against multi-member teams. AXIOM proves that high-end, culturally aligned AI development can be executed entirely on edge hardware through deep mathematical optimization.