FlowBender

Official implementation of FlowBender: Feedback-Aware Training for Self-Correcting Conditional Flows (Gilo et al., 2026).

• 🌐 Project page: https://flow-bender.github.io/
• 📄 Paper: https://arxiv.org/pdf/2606.20404

FlowBender is a feedback-aware training method for conditional generation with flow matching models: the model is trained to correct its own errors along the sampling trajectory. This self-correction capability improves sample quality and adherence to the provided condition across different conditional generation tasks, such as depth-to-RGB, edge-to-RGB, image super-resolution, JPEG restoration, and 3D mesh texturing.

This repository hosts the code for the two experiment tracks presented in the paper, each maintained as an independent, stand-alone submodule:

Each submodule is self-contained and includes its own setup, training, and evaluation instructions — see the README inside each one.

Clone with submodules:

git clone --recurse-submodules https://github.com/DanielGilo/FlowBender.git

If you've already cloned the repo without submodules, fetch them with:

git submodule update --init --recursive

If you find this work useful, please cite:

@misc{gilo2026flowbenderfeedbackawaretrainingselfcorrecting,
  title         = {FlowBender: Feedback-Aware Training for Self-Correcting Conditional Flows},
  author        = {Daniel Gilo and Sven Elflein and Ido Sobol and Or Litany},
  year          = {2026},
  eprint        = {2606.20404},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2606.20404},
}