Flower GAN on an ePaper badge
A 3.42-million-parameter INT8 flower generator running entirely on the Espruino ESP32-C3 ePaper Badge. Images generate in about 48 seconds, with the next generation overlapping the four-color screen refresh.
The model generates 320 × 176 RGB. The original 320 × 172 crop is enlarged to 800 × 430 and dithered to black, white, yellow, and red. Small borders show timing, free RAM, image number, and seed. Hardware: ESP32-C3, 4 MiB flash, 800 × 480 panel, maker's rev 1.2 pin mapping. Other ESP32 boards need their own port.
Easiest flash: your browser
This replaces the original app and saved badge data. Back up first if you want to preserve your photo, contacts, and settings; see backup instructions.
- Download firmware.bin. It includes the application and all model weights.
- Connect a USB data cable. In the original summit app, hold the rear SHARE button until the LEDs stay solid blue, then release. Close other serial apps.
- In desktop Chrome or Edge, open Espressif's ESP Tool. Click Connect in the Program section and select the badge's USB serial/JTAG device. It should identify as ESP32-C3; leave WebUSB (CH340) unchecked.
- Choose
firmware.binin one file row at Flash Address0x0. Set flash mode, frequency, and size to keep. Leave Erase all unchecked and click Program. - Wait for verification and automatic reset. The first flower appears in about one minute; subsequent images advance automatically. Both buttons are unassigned.
No custom installer scripts or build tools are required. For terminal users:
python -m pip install esptool==5.6.0
python -m esptool --chip esp32c3 --port PORT --baud 460800 --before usb-reset --after hard-reset write-flash 0x0 firmware.bin
Replace PORT with your device's port. If it does not appear, check the cable,
solid-blue mode, and whether another app has the port open.
Code, weights, and data
| Path | Contents |
|---|---|
| source/ | Native C inference, display driver, and ESP-IDF project |
| training/ | Original PyTorch architecture, training, data preparation, and export code |
| data/ | Exact INT8 weights, tensor layout, original model export, and reference tensors |
| data/dataset.json | Training dataset source, pinned revision, and preprocessing metadata |
| BUILD.md | Build versions and commands |
| release.json / SHA256SUMS | Flash layout, versions, and file hashes |
Training used all 8,189 Oxford Flowers102 images, from the
pufanyi/flowers102 mirror,
revision b61e48cb66625272ca3bf36442062362587fa9bb. Images are not mirrored here;
the source and preparation code are included. The original floating-point training
checkpoint was not in the supplied bundle; the original INT8 export is included.
Original project code and trained weights are released under MIT. Third-party code and dataset terms remain separate; see NOTICE.md. The release contains no private badge data. The inference implementation matched its reference byte for byte across seven seeds and ran successfully on the badge.