Textile Generation based on Diffusion Models

This repository contains a toolkit for the semantic expansion and synthesis of textile textures, developed as part of a Master’s Thesis in Computer Engineering at the University of Bologna.

Project Overview

The project tackles a key bottleneck in AAA game development: the creation of high-quality textile assets.

Traditional tiling techniques often produce visible repetitions (the “wallpaper effect”), while manual asset creation is both expensive and time-consuming.
This work introduces a generative approach to perform semantic upscaling from 256×256 source samples to 1024×1024 production-ready textures that are:

• Seamless (natively tileable)
• Structurally coherent
• Visually consistent

Key Features

• Dual-Pathway Framework Adaptive generation strategies based on textile topology:
  • Regular / Geometric patterns (first figure)
  • Irregular / Organic patterns (second figure)

• Visual Guidance (IP-Adapter)
  Uses image-based conditioning to control weave density and material identity, overcoming the limitations of text-only prompts.

• Native Tileability
  Seamless textures are achieved through Noise Rolling and Circular Padding, applied directly within the denoising loop to model the latent space as a toroidal surface.

• Latent Replication
  A structured initialization strategy applied at ~60% of the denoising process to upscale resolution while preventing structural drift.

Technical Stack

• Model: Stable Diffusion v1.5 (Latent Diffusion Model)
• VAE Decoder: Fine-tuned MSE (ft-mse) variant for high-frequency detail preservation
• Hardware: NVIDIA Quadro P4000 (Pascal architecture)

Project Structure

.
├── assets/
├── main_irregular.ipynb  # code for texture with irregular patterns
├── main_regular.ipynb    # code for texture with regular patterns
├── docs/
│   └── textile-generation_presentation.pdf
├── README.md
├── .gitignore
├── LICENSE

Authors

• Federica Di Giaimo
• Supervisor: Prof.ssa Serena Morigi
• Co-supervisor: Paolo Zuzolo