Deep Daze is a simple command-line tool for text-to-image generation that pairs OpenAI’s CLIP model with a SIREN implicit neural representation. Type a phrase, and it gradually dreams up an image that CLIP judges to match your words.
The technique, discovered by Ryan Murdock, works by optimizing the pixels produced by a SIREN network so that CLIP’s image-text similarity score for the prompt keeps improving. The result is dreamy, painterly artwork that helped kick off the CLIP-guided art wave in 2021, before diffusion models took over. Deep Daze supports deeper, higher-quality renders, story mode for sequences of related frames, and straightforward options for image size and training epochs. It is maintained by prolific open-source author lucidrains, is released under the MIT license, and has gathered over 4,300 stars on GitHub. The code was last updated in 2022, so treat it as a historically important, still-fun creative tool rather than a state-of-the-art generator.
Deep Daze is free and open source under the MIT license. You only need your own compute, ideally a GPU, to run the generation locally.
Generative-art hobbyists and AI enthusiasts who enjoy the surreal CLIP-art aesthetic, and developers curious about how far a simple optimization loop can take image generation.