Omost is a project aiming to translate the encoding capabilities of Large Language Models (LLMs) into image generation (more accurately, image composition) capabilities. It offers pre-trained LLM models based on Llama3 and Phi3 variants, which can generate code to utilize Omost's virtual Canvas agent for composing visual image content. The Canvas can be implemented by specific image generators to actually render the images. The technologies underpinning the Omost project include Direct Preference Optimization (DPO) and OpenAI GPT4o's multi-modality.