CycleGAN vs. Stable Diffusion — A Short Comparative Study

Graduate-level course project comparing two popular generative methods:

• CycleGAN (cycle-gan-high-epoch.ipynb): unpaired image-to-image translation with cycle-consistency.
• Stable Diffusion (stable_diffusion.ipynb): latent diffusion for text/image-conditioned synthesis and stylization.

---

What we asked

When should you use CycleGAN vs. Stable Diffusion?
We compare sample quality, faithfulness to conditioning, style control, and data/compute requirements.

Key takeaways (summary)

• CycleGAN shines for domain translation (A↔B) when no paired data exists (e.g., horses→zebras), preserving scene layout and structure.
• Stable Diffusion excels at open-ended generation and prompt-controlled edits with strong diversity; minimal training needed (often none), but careful prompt engineering and safety filtering matter.
• Typical trade-off: CycleGAN = task-specific training, strong structure preservation; SD = broad capability, strong priors, less control over exact geometry without extra guidance/LoRA/ControlNet.

---

Repository structure