Generative models can skip step-by-step constraint checks: HardFlow enforces hard limits only at the output, clearing robot and image-editing tasks

In high-stakes settings, a generative model can satisfy nonnegotiable hard constraints only at the final output instead of making every generation step feasible, so output quality is no longer suppressed by constraints applied throughout. Projection-based sampling repeatedly forces intermediate samples to satisfy strict requirements during generation; constraining the whole process can prevent the model from reaching a better final solution and typically targets only constraint satisfaction, forgoing other qualities such as a shorter robot trajectory. Across robotic manipulation, maze navigation and text-guided image editing, HardFlow achieved perfect constraint satisfaction and consistently outperformed baseline methods on solution-quality measures; for example, it let a robotic manipulator avoid obstacles while finding the quickest path to the target object, whereas most other methods either collided or found paths that took significantly more time.

The method treats hard-constrained sampling as a trajectory to be optimized, using optimal-control tools to make subtle corrections along the way and defer constraint enforcement to the endpoint. Solving a trajectory-optimization problem around a large neural network with hundreds of interconnected layers is not easy, so the structure of flow-matching models is used to break the problem apart. HardFlow leverages the structure of flow-matching models to decompose the trajectory-optimization problem into a sequence of smaller single-step subproblems, then applies systematic transformations and approximations to derive an efficient, scalable algorithm that still finds a feasible solution and runs at deployment time.

The demonstrated scope is three experiment families: robotic manipulation, maze navigation and text-guided image editing, with the method applied at deployment time to pretrained generative models without retraining. The researchers propose extending the framework in the future to settings where the model itself can also be updated, so constraint satisfaction and sample quality improve more adaptively.

The material states that HardFlow achieved perfect constraint satisfaction and consistently outperformed baselines in experiments but gives no specific metric values; computation time is described only as comparable to or lower than most competing methods, without itemized figures. Extending the framework to settings where the model itself can be updated remains a proposed future direction and has not been validated.