FlowBender: A Closed-Loop Training Framework That Feeds Alignment Errors Back Into Conditional Flow Models

Researchers have introduced FlowBender, a training framework for conditional flow and diffusion models that treats a model's own alignment error as a direct input during training — rather than ignoring it (as in standard supervised approaches) or correcting it post-hoc through hand-tuned guidance. At each sampling step, an unguided look-ahead pass estimates the clean output, a task-specific deviation is computed via the forward operator, and a refinement pass uses that feedback to produce a corrected velocity. The result is a closed-loop "correction policy" the model learns, rather than a fixed scalar guidance weight.

FlowBender ships with several variants — including a gradient-based formulation for differentiable operators and a zero-order variant for non-differentiable settings like JPEG compression — as well as a prior-step shortcut for efficient sampling at minimal extra compute. Across benchmarks spanning image-to-image translation, restoration, and 3D mesh texturing, the framework outperforms supervised baselines, alignment-loss-augmented training, and state-of-the-art inference-time guidance, improving both condition fidelity and sample plausibility simultaneously rather than forcing a trade-off between them.