Researchers present FLAT (Flexible-Length Aligned Transmodal representations), a multimodal pre-training framework that jointly trains a shared encoder with text-to-image and image-to-text decoders in a single stage, rather than training a frozen representation encoder first and a separate generative model afterward. FLAT maps visual and textual inputs into a unified continuous 1D sequence space and applies nested dropout over prefix tokens to support variable-length outputs from one model. The resulting representations achieve a 71.1 T2I GenEval score directly, improve to 83.1 with task-specific fine-tuning, and natively support linear interpolation, latent-space arithmetic, and zero-shot composed retrieval.
