Self-Inverse Networks: One2One CycleGAN

• Self-inverse networks are image-to-image translation architectures that enforce bijective mappings using a single generator or tied dual generators.
• They utilize specialized loss functions and parameter-sharing techniques, such as orthonormality constraints, to guarantee that the mapping is its own inverse.
• Empirical results show these networks improve translation clarity, reduce mode collapse, and offer enhanced compute and parameter efficiency compared to standard CycleGANs.

Self-Inverse Networks (One2One CycleGAN)

Self-inverse networks—often termed One2One CycleGANs—are architectures for image-to-image translation that enforce or exploit bijective and self-inverse properties in the learned mapping between two domains. Departing from the standard CycleGAN’s dual-generator approach, these networks implement a single model or explicitly constrained pair such that either the learned function is its own inverse (involution) or the forward and backward translators are structurally, parametrically, or loss-wise tied to guarantee mutual invertibility. The motivation is to ensure one-to-one, non-degenerate translations, reducing ambiguity and mode collapse in settings with unpaired data, and enhancing parameter- and compute-efficiency.

1. Core Principles and Mathematical Formulation

The classical CycleGAN seeks bijections $G:X\to Y$, $F:Y\to X$ between domains $X$, $Y$, driven by adversarial losses and a cycle-consistency constraint, e.g., $\|F(G(x)) - x\|_1$ and $\|G(F(y)) - y\|_1$ (Zhu et al., 2017). However, cycle-consistency by itself does not enforce invertibility at the level of individual samples—mode collapse and surjective non-injectivity are possible. One2One/self-inverse variants introduce architecture, loss, or training modifications to enforce:

• Self-inverse mapping: A single function $f$ with $f(f(z)) = z$ for $z \in X \cup Y$ (Shen et al., 2019, Shen et al., 2019).
• Explicit invertible structure: Pairing forward/backward generators as inverses or enforcing orthonormality so that $G = F^{-1}$ (Teng et al., 2018).
• Loss engineering: Deviating from conventional cycle loss by e.g., deviation-loss or optimal transport guidance (Nikam, 2018, Lu et al., 2018, Sim et al., 2019).
• Parametric tying: Weight sharing (transpose, parameter symmetry) or blockwise invertibility (Ouderaa et al., 2019, Kwon et al., 2021).

A consequence is provable one-to-one correspondence, reduced solution space, parameter savings, and enhanced generalization.

2. Architectural Variants and Parameterizations

Several classes of self-inverse/One2One CycleGAN architectures have been developed:

Variant	Key Ingredients	Reference
CycleGAN (baseline)	2 generators, 2 discriminators, cycle loss	(Zhu et al., 2017)
Deviation-loss One2One	Single encoder/decoder, ‘translator’, deviation loss on B	(Nikam, 2018)
InvAuto CycleGAN	Encoder/decoder as inverses, orthonormal sharing	(Teng et al., 2018)
Self-inverse single net	One U-Net, alternate input/output, separated cycle loss	(Shen et al., 2019, Shen et al., 2019)
Optimal Transport guided	OT barycenter reference, cycle + reference loss	(Lu et al., 2018)
Reversible GAN/RevGAN	Reversible core blocks, parameter tying, memory efficiency	(Ouderaa et al., 2019)
Cycle-free Invertible GAN	Full network-level invertibility, no explicit cycle loss	(Kwon et al., 2021)
OT-CycleGAN w/ known H	One generator/discriminator, explicit cycle via physics H	(Sim et al., 2019)

Notably, several approaches deploy a single generator (sometimes involutive, sometimes with known physics as inverse), while others use pairs tied by parameter symmetry and/or loss.

3. Theoretical Guarantees and Solution Space

The characterization of self-inverse solutions—in particular for CycleGAN with strict cycle-consistency loss—has deep theoretical implications. Given the pure cycle loss

$F:Y\to X$0

zero-loss solutions correspond to mutual inverses ($F:Y\to X$1, $F:Y\to X$2). However, the space of solutions is highly non-unique: the set of exact minimizers forms a principal homogeneous space (torsor) under the group of measure-preserving automorphisms of the domains (Moriakov et al., 2020). Any (nontrivial) automorphism (e.g., class permutation, latent-space rotation) yields a valid bijection, making the raw cycle-consistency constraint insufficient for semantic or user-desired alignments. This necessitates explicit constraints—domain knowledge, task-specific costs, OT-barycenters, identity losses, or architectural limitation—to break underlying symmetries and select semantically meaningful maps.

4. Loss Terms and Training Algorithms

Self-inverse networks leverage tailored objectives. The following summarizes representative loss formulations:

• Standard adversarial loss
  • GAN or least-squares GAN (LSGAN) on both domains; PatchGAN or ResNet discriminators are typical.
• Cycle-consistency loss
  • Classic: $F:Y\to X$3 (Zhu et al., 2017).
  • Self-inverse: enforced via $F:Y\to X$4 for a single $F:Y\to X$5 (Shen et al., 2019).
• Deviation-loss: Penalizes the deviation from identity in the target domain encoding (e.g., $F:Y\to X$6) (Nikam, 2018).
• Orthonormality/invertibility loss: Enforces $F:Y\to X$7 for weight-sharing layers, yielding explicit inversion (Teng et al., 2018).
• OT-barycenter-guided loss: Incorporates a reference loss against barycenters derived via task-specific OT plans (Lu et al., 2018).
• Adversarial dual (OT, WGAN-GP): In cases with known forward operator ($F:Y\to X$8), the adversarial loss measures the dual gap, minimized alongside a cycle term based on $F:Y\to X$9 (Sim et al., 2019).
• Wavelet residual domain: Loss is computed on or regularized by high-frequency (wavelet) components to increase invertible mapping stability (Kwon et al., 2021).

Training typically follows an alternating min–max schedule (Adam optimizer, fixed and decaying learning rates, batch size ≈1), with discriminator updates per generator step. For memory-efficient invertible architectures, gradients for invertible blocks are recomputed on-the-fly; in self-inverse models, input/output swapping and direction alternation doubles data variety.

5. Empirical Results, Performance Characteristics, and Applications

Empirical validation spans classic vision datasets (Cityscapes, Google Maps, Yosemite, ImageNet subsets) and biomedical imaging (BraTS, AAPM CT, MRI ↔ CT synthesis):

• One2One/self-inverse networks match or exceed baseline CycleGAN/UNIT models in L1, PSNR, and SSIM metrics, sometimes with ∼2× training speedup and 0.5× parameters when single-generator architectures are used (Nikam, 2018, Teng et al., 2018, Sim et al., 2019, Kwon et al., 2021).
• In cross-domain tasks (e.g., apple↔orange, summer↔winter, horse↔zebra), user studies and automated metrics consistently show sharper detail and less ambiguity under the self-inverse constraint (Shen et al., 2019).
• For medical modalities (T₁↔T₂, LDCT denoising), self-inverse networks report higher PSNR and SSIM relative to standard approaches, with one-to-one enforced architectures consistently outperforming two-generator CycleGAN (Shen et al., 2019, Kwon et al., 2021).
• Approximately invertible/reversible GANs sustain or improve accuracy (pixel accuracy, MAE, RMSE) while allowing arbitrarily deep networks under constant memory budgets (Ouderaa et al., 2019).
• Sensitivity analyses indicate self-inverse networks increase robustness and bijection quality, with larger changes in output per input perturbation—indicative of less mode collapse (Shen et al., 2019).
• In optimal transport–augmented schemes, the mapping can be explicitly guided to preserve geometric, color, or semantic attributes—addressing the random-bijection issue of unconstrained CycleGAN (Lu et al., 2018).

6. Limitations, Solution Ambiguity, and Constraints

Despite theoretical invertibility, in practical unconstrained settings (pure cycle-consistency minimization), the learned one-to-one mapping may be arbitrary within the space of automorphisms—permutations, flips, or otherwise semantically meaningless combinations (Moriakov et al., 2020). Empirical evidence includes digit-swapping on MNIST and global symmetries learned (horizontal flips) on medical data sets. To overcome this and guarantee semantic invertibility:

• Identity, reconstruction, or feature-preservation losses can constrain trivial automorphisms.
• Domain-knowledge (anatomical, color, orientation) and OT-based reference matching guide mapping selection (Lu et al., 2018, Sim et al., 2019).
• Task-specific architectures, such as those leveraging known physics (e.g., forward operator $X$0), reduce model complexity and solution ambiguity (Sim et al., 2019).

These enhancements render the one-to-one mapping both unique and aligned with application requirements.

7. Extensions and Future Directions

Self-inverse principles are now extended across modalities—photo ↔ sketch, day ↔ night, CT ↔ MRI, super-resolution ↔ down-sampling, and deblurring ↔ reblurring—where exact invertibility or stable bijection is critical (Kwon et al., 2021). Emerging directions include:

• Multi-domain/multi-modal CycleGANs with explicit self-inverse and attribute-decomposed constraints (Nikam, 2018, Lu et al., 2018).
• Increasingly memory- and compute-efficient invertible architectures for large-scale 2D/3D training (Ouderaa et al., 2019, Kwon et al., 2021).
• Further integration of optimal transport theory for high-level controllability (Lu et al., 2018, Sim et al., 2019).
• Exploring the full ramifications of solution non-uniqueness and designing invariance-breaking mechanisms for robust, interpretable transfer (Moriakov et al., 2020).

Empirical accuracy and theoretical clarity make self-inverse (One2One) architectures a compelling paradigm for reliable, bijective, and resource-efficient unpaired image translation.