HODLer: a meme character that never drifts
Model training, a playground and an AI agent for TON by Telegram
January–March 2025. TON’s crypto media needed a steady stream of drawn content starring the HODLer meme character: the face and helmet identical in every image, while the outfit, place and situation change every time. I trained a custom model for the character, built a playground for the designers and an AI agent that illustrates Telegram posts by itself. Back then a character like this wouldn’t hold without a model of its own.
- Dataset
- Training and selection
- Playground
- AI agent
- Telegram posting
01
Face and helmet always, everything else new
HODLer is a crypto-community meme: someone who holds their coins whatever the market does. The character comes down to two things: the meme’s recognizable face and the helmet. They have to match in every generation, or it isn’t HODLer anymore. The wardrobe, the setting and the situation from the post are new every time.
The content had to come in volume, and drawing every image by hand couldn’t keep up. It needed a way to get a new HODLer for any story without losing the character.
02
Eight training cycles
The starting dataset of 100+ character variations was drawn by hand, and I trained the first version of the model on it. Each version went through 50+ test generations, and the best results fed the next cycle’s dataset.
Over eight cycles the model learned to hold the face and helmet across any look, background and situation with under 5% error, and to mix styles without losing the character.
03
How the training loop works
Each cycle has three steps: dataset, tests, selection. The best generations go back into the dataset, and the next model version learns from them.
04
A playground for the designers
A model on its own is a file designers can’t comfortably work with. So I built a playground on top of it: the team’s designers pick a model version, tune the generation, write the scene and test the result in a simple interface before an image goes out. Everything that works is collected in a shared gallery.
05
An AI agent draws for every post
The last step automated posting: an AI agent reads a Telegram post, comes up with a scene for it and generates a HODLer image on its own. The caption and the illustration go out together, with no manual moderation.
06
What has changed since
Today many tasks like this need no training: current models hold a character or a style from a few references. I train a custom model where references fall short: a strict character, a complex brand style or a high volume of content.
