FullFlow: Upgrading Text-to-Image Flow Matching Models for Bidirectional Vision--Language Generation
The paper introduces FullFlow, a method for enhancing text-to-image flow matching models to enable bidirectional vision-language generation. This approach improves performance metrics significantly while maintaining a low computational footprint. FullFlow demonstrates that pretrained models can be adapted for advanced multimodal capabilities without extensive retraining.
- ▪FullFlow upgrades a pretrained text-to-image model into a bidirectional generator using LoRA adapters and lightweight text heads.
- ▪The method improves text-to-image FID from 62.7 to 31.6 and image-to-text CIDEr from 2.0 to 99.4.
- ▪It reduces peak VRAM usage from approximately 84 GB to 38 GB and increases throughput by about 8 times on two RTX A5000 GPUs.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computer Vision and Pattern Recognition arXiv:2605.20316 (cs) [Submitted on 19 May 2026] Title:FullFlow: Upgrading Text-to-Image Flow Matching Models for Bidirectional Vision--Language Generation Authors:Eric Tillmann Bill, Enis Simsar, Alessio Tonioni, Thomas Hofmann View a PDF of the paper titled FullFlow: Upgrading Text-to-Image Flow Matching Models for Bidirectional Vision--Language Generation, by Eric Tillmann Bill and 2 other authors View PDF Abstract:Modern text-to-image diffusion models encode rich visual priors, but expose them only through one-way text-conditioned generation.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
