RRFC: Recursive Refinement via Feedback Conditioning for Iterative Image-to-Image Generation

Details

Standard conditional image-to-image generators map inputs to outputs in a single forward pass with no opportunity to self-correct. RRFC (Recursive Refinement via Feedback Conditioning) adds an iterative loop by feeding the model's most recent prediction back as auxiliary channels alongside the original input, conditioning each refinement step on prior output. The framework is designed to attach to existing generator architectures without structural changes — it modifies the conditioning interface and, depending on model family, the training or inference procedure. This means it can be tested against deployed models without redesign. Results across 18 architecture-task combinations spanning adversarial, equilibrium, and diffusion-based models show a mixed picture: seven Holm-corrected improvements, seven degradations, and four non-significant changes. Gains concentrate on reconstruction-fidelity and identity preservation tasks. Five of seven degradations occur on semantic layout tasks, where every tested model declines. The conclusion: feedback conditioning helps when its objective overlaps with the evaluation criterion — specifically when fidelity to a reference is paramount — but interferes when tasks require spatial semantic understanding. The honest reporting of negative results across multiple baselines is a methodological strength. Source is an arXiv preprint.