Liquid

Why we included this project

Liquid is a research codebase that tests a fairly bold idea: one large language model can both understand and generate images, with no separate pretrained encoder like CLIP in the loop. The released 7B instruction-tuned checkpoint does both, so the same weights can answer questions about a picture and then produce a photorealistic image from a text prompt. It ships with a Gradio demo, inference scripts for text-to-image, image-to-text, and text-to-text, plus evaluation and training code, so you can actually run it rather than just read the paper. The authors also report scaling results from 0.5B to 32B across three model families, which is useful if you are deciding whether unified training is worth the compute at your own scale. For researchers working on multimodal foundation models, the checkpoints and reproduction scripts are the most useful part.