Was browsing arXiv and came across this thing called FlowBender, pretty interesting idea. Right now there are two main ways to add constraints (like depth maps, edges, poses) to diffusion/flow models: one is treating the constraint as a static prompt thrown in at the start and then just ignoring whether the final output actually follows it; the other is manually tweaking formulas during generation, but that runs into a trade-off—push too hard and the image follows the rules but looks rough, ease up and it looks good but goes off track.
FlowBender’s approach is to train the model to measure its own deviation at each step and then correct itself, instead of relying on external force to bend it.
The paper claims it can improve both constraint fidelity and image quality at the same time, which are usually a seesaw relationship—one goes up, the other goes down—so pulling that off is pretty rare. arXiv ID 2606.20404, hit up the original if you’re interested.