MC-I2I: Multi-Cycle Guided Unpaired Image-to-Image Translation on Industrial Long-Tail Data

Autonomous vision systems in modern industry must routinely translate images across unpaired and highly imbalanced domains, where rare target-domain patterns are severely underrepresented. Existing unpaired translation methods tend to degrade under such severe imbalance. To address this challenge, we propose MC-I2I, a multi-cycle guided image translation framework that integrates three complementary mechanisms. Dual-Cycle Latent Consistency imposes self-cycle and cross-cycle constraints in latent space to suppress representation drift caused by data imbalance. Dual-Branch Guided Reconstruction balances discriminability and diversity by alternately freezing the encoder and decoder, enabling each branch to receive dense pixel-level gradients from its fixed counterpart. Generator-Led Pseudo Cycle allows the generator to guide the discriminator toward learning rare target-domain characteristics, thereby improving sensitivity to scarce operating conditions. Extensive experiments across multiple tasks show that MC-I2I consistently improves translation fidelity and remains robust under severe long-tail conditions.