Can a Language Model Paint?
2 days ago
- #AI Art
- #Creative Process
- #Language Models
- Language models can generate images from text prompts, but the one-shot approach feels artistically unsatisfying due to immediacy and derivative reproduction.
- An app was built to generate images iteratively with a vision language model, applying strokes one at a time to mimic a human creative process.
- Output quality varies by model size; small models produce scribbles, while frontier models create recognizable but mixed results, revealing limitations in iterative task completion.
- The iterative process visualizes the fragility of LLM-generated artifacts, where a single bad stroke can collapse an image, mirroring fragility in LLM-generated codebases.
- The project explores artistic sincerity and accessibility, inspired by Tolstoy's 'What Is Art', but the results still feel derivative rather than earnest.