CycleVAR：重用自回归模型实现无监督单步图像翻译

当前的 conditional autoregressive image generation 方法在 results 方面显示出前景，但其在 practical unsupervised image translation 领域的潜力 remains largely unexplored，该领域 operates without explicit cross-domain correspondences。一个 critical limitation 源于 traditional Vector Quantization-based frameworks 中的 discrete quantization，这 disrupts Variational Autoencoder 解码器与 causal Transformer 之间的 gradient flow， impeding end-to-end optimization during adversarial training in image space。为解决这一问题，我们提出使用 Softmax Relaxed Quantization，一种 novel approach 将 codebook selection reformulates 为 continuous probability mixing process via Softmax，从而 preserving gradient propagation。基于此 differentiable foundation，我们引入 CycleVAR，将 image-to-image translation reformulates 为 image-conditional visual autoregressive generation by injecting multi-scale source image tokens as contextual prompts，类似于 language models 中的 prefix-based conditioning。CycleVAR exploits two modes to generate target image tokens，包括 (1) serial multi-step generation， enabling iterative refinement across scales，和 (2) parallel one-step generation，synthesizing all resolution outputs in single forward pass。实验结果表明，parallel one-step generation mode 在 unsupervised scenarios 中 获得 更好的 translation quality with 更快的 inference speed。此外，quantitative and qualitative results indicate that CycleVAR 超越了 previous state-of-the-art unsupervised image translation models，例如 CycleGAN-Turbo。