Train and deploy tiny GANs on microcontrollers.

MicroGAN trains a compact generative adversarial network on your laptop and exports it as a single C header file you can drop into any embedded project. Your microcontroller generates novel images on-device from a learned distribution — no cloud, no network, no runtime allocations.

Run the entire pipeline end to end:

./run_pipeline.sh

This trains a generator, converts it through ONNX to quantized TFLite, and produces a C header with the weights baked in. Takes about 30 seconds.

python3 -m venv venv
source venv/bin/activate
pip install -e .

Requires Python 3.9+ with PyTorch, TensorFlow, and ONNX. All dependencies are handled by pip install -e ..

microgan train \
  --data ./my_dataset \
  --epochs 200 \
  --latent-dim 32 \
  --channels 1 \
  --output-dir build

Trains a DCGAN that learns to generate 32x32 images from your dataset. Omit --data to run with a built-in dummy dataset for testing the pipeline.

Once training is done, three commands take you from a PyTorch checkpoint to a C header:

# PyTorch -> ONNX
microgan export-onnx \
  --checkpoint build/generator_final.pt \
  --output-dir build

# ONNX -> quantized TFLite
microgan onnx-to-tflite \
  --onnx-path build/generator.onnx \
  --output-dir build

# TFLite -> C header with weights and quantization parameters
microgan convert \
  --tflite build/generator_quantized.tflite \
  --output-dir build

The final output is MicroGAN_weights.h.

Copy MicroGAN_weights.h, MicroGAN_runtime.h, and MicroGAN_runtime.c into your embedded project:

#include "MicroGAN_runtime.h"
#include "MicroGAN_weights.h"

static uint8_t arena[1024 * 64];  // scratch space for activations
static uint8_t output[32 * 32];   // 32x32 grayscale image buffer

void setup() {
    MicroGAN_init(arena, sizeof(arena));
}

void loop() {
    uint8_t seed = get_random_seed();
    MicroGAN_generate(seed, 0, output);
    display_image(output);
}

Different seeds produce different images from the learned distribution. The runtime uses zero heap allocation — all memory comes from the static arena you provide.

• ESP32 / ESP32-S3
• STM32F4 / STM32F7 (via CMSIS-NN)
• Raspberry Pi Pico (RP2040)
• Arduino Uno R4

See LICENSE for details.