germi123/RamPhotoGenerator

Neural RAM Representation (SIREN)

This model is an Implicit Neural Representation (INR) of a RAM module image. Unlike traditional raster formats (PNG/JPG) that store pixels in a grid, this image is encoded entirely within the weights of a neural network.

Project Overview

The model utilizes the SIREN (Sinusoidal Representation Networks) architecture. It is trained to map continuous 2D coordinates $(x, y)$ to their corresponding $RGB$ color values.

Key Features:

• Mathematical Storage: The image exists as a continuous function rather than a discrete pixel array.
• Infinite Resolution: You can render the image at any scale by increasing the density of the coordinate grid.
• Efficient Encoding: The model has been optimized (~12,000 epochs) to capture fine textures of the PCB and gold-plated contacts with high fidelity.

Technical Specifications

• Architecture: Multi-Layer Perceptron (MLP) with periodic (sine) activation functions.
• Depth: 2 hidden layers.
• Width: 128 neurons per layer.
• Format: Provided as a lightweight .safetensors file.

Usage

To render the image, you will need torch, safetensors, and numpy.