FlowBender: Feedback-Aware Training for Self-Correcting Conditional Flows

Researchers introduce FlowBender, a novel training framework designed to address the alignment gap in conditional diffusion and flow models, ensuring that generated outputs strictly adhere to the constraints defined by their conditioning signals through a feedback-aware mechanism.

The Challenge of Constraint Satisfaction in Generative Models

Conditional diffusion and flow models are frequently employed to generate data based on specific constraints, such as depth-conditioned image synthesis. However, a recurring failure mode in these architectures is the lack of strict adherence to the input constraints. A common example occurs when a depth-conditioned model generates an image where the re-extracted depth—calculated via the forward operator—disagrees with the original input depth map, despite that operator being available during both the training and inference phases.

Limitations of Current Approaches

Existing methodologies for handling conditioning signals typically fall into two categories, both of which exhibit significant limitations:

• Supervised Models: These treat the conditioning signal as a static cue, effectively ignoring the alignment information provided by the forward operator.
• Alignment-focused Models: While attempting to enforce constraints, these often lack a dynamic feedback loop that allows the model to self-correct during the generative process.

Introducing FlowBender

FlowBender proposes a "feedback-aware" training paradigm. Unlike standard conditional flows, FlowBender leverages the forward operator to provide a gradient-based feedback loop. By treating the discrepancy between the input constraint and the output's re-extracted property as a signal for correction, the model can iteratively refine its output to satisfy the defining constraints of the task.

Note: The provided source material is a summary; detailed architectural specifications, specific loss functions, and quantitative benchmark results are not available in the provided snippet.