English

FullFlow: Upgrading Text-to-Image Flow Matching Models for Bidirectional Vision--Language Generation

Computer Vision and Pattern Recognition 2026-05-21 v1 Artificial Intelligence

Abstract

Modern text-to-image diffusion models encode rich visual priors, but expose them only through one-way text-conditioned generation. Existing unified vision--language models derived from them recover bidirectional capability through large-scale joint pretraining or substantial retraining of the text pathway, discarding the strong image prior the text-to-image backbone already encodes. We introduce \emph{FullFlow}, a parameter-efficient recipe that upgrades a pretrained rectified-flow text-to-image model into a bidirectional vision--language generator by training only LoRA adapters and lightweight text heads. FullFlow keeps images in their native continuous flow and adds a discrete insertion process for text. Separate image and text timesteps turn inference into trajectory selection in a two-dimensional generative space, enabling text\rightarrowimage, image\rightarrowtext, joint sampling, and partial-text prediction with a single backbone. On Stable Diffusion 3 (SD3) under an identical trainable-parameter count and matched LoRA rank, FullFlow improves text\rightarrowimage FID from 62.762.7 to 31.631.6 and image\rightarrowtext CIDEr from 2.02.0 to 99.499.4 over a LoRA equivalent following the previous SOTA formulation (Dual Diffusion) at matched wall-clock training time, while reducing peak VRAM from 84{\sim}84\,GB to 38{\sim}38\,GB and raising throughput by 8×{\sim}8\times on two RTX A5000 GPUs in under 24 hours, training only 5%{\sim}5\% of the backbone parameters. The same recipe transfers to FLUX.1-dev and supports downstream VQA through partial-text generation. These results show that strong bidirectional vision--language capability can be unlocked from pretrained text-to-image flow models without full multimodal pretraining.

Keywords

Cite

@article{arxiv.2605.20316,
  title  = {FullFlow: Upgrading Text-to-Image Flow Matching Models for Bidirectional Vision--Language Generation},
  author = {Eric Tillmann Bill and Enis Simsar and Alessio Tonioni and Thomas Hofmann},
  journal= {arXiv preprint arXiv:2605.20316},
  year   = {2026}
}

Comments

project page: https://ericbill21.github.io/fullflow/