Meme me is a playful yet critical installation that blurs the line between automated image recognition and internet culture. The work explores how artificial intelligence interprets visual content and repackages it into the internet’s most potent form of commentary: Memes

In an age when screenshots, selfies, and social media posts form a vast visual archive of digital life, mememe reclaims this mundane visual material and feeds it through a chain of AI perception and language generation—ultimately producing a meme that is both absurd and unsettlingly accurate.

You can see all the generated memes here : https://memes.meme-me.hertz-lab.zkm.de/

How It Works

The process is deceptively simple:

1. Screenshot Capture
   The system passively captures a screenshot or image frame at timed intervals or through user interaction.

2. Visual Analysis
   The captured image is analyzed using LLaVA 1.6, a multimodal vision-language model served via Ollama. The model generates a rich, high-level description of the visual content—its objects, actions, contexts, and implicit cues.

3. Text Generation
   This description is passed to a language model (LLM), which crafts a short, two-line caption—mimicking the syntax and cadence of typical internet memes. The LLM is trained to balance humor, sarcasm, and context-aware absurdity.

4. Meme Creation
   Finally, using OpenCV, the meme text is overlaid onto the original image in classic meme format. The result is an AI-generated meme—co-created by machine vision, natural language generation, and code.

Yasha Jain's Role

Yasha Jain conceptualized and implemented the entire mememe pipeline, including:

• LLaVA 1.6 Integration via Ollama

  • Deployed and fine-tuned the vision-language model locally for fast and efficient inference.

  • Customized prompts for generating rich yet meme-suitable image descriptions.

• LLM Prompt Engineering & Humor Modelling

  • Built a custom prompt structure for generating two-line meme text with the right tone, humor, and brevity.

• Meme Generation Pipeline (OpenCV)

  • Developed the final visual rendering process using OpenCV to mimic native meme aesthetics (impact font, top-bottom format, black border options).

  • Ensured real-time generation and seamless transitions for installation use.

• System Architecture & Deployment

  • Designed a modular system that can be run locally with minimal hardware and no external API dependencies, prioritizing privacy and speed.