Diffusion Generation

Empty-prompt generation isolated the vessel hint more strongly than disease text prompts. VesselControlNet conditioned generation on the Bézier hint by duplicating the Stable Diffusion UNet encoder and adding zero-convolution residuals. Higher classifier-free guidance partly recovered a hint-aligned effect, suggesting competition between text and parametric hints. Stable Diffusion 2.1 was fine-tuned with a rank-32 LoRA adapter for 36,000 steps. LoRA and ControlNet needed matched noise-offset co-training to avoid color drift artifacts.