Parameter-efficient optimization of conditional diffusion models using multi-resolution attention, classifier-free guidance ablation, and DDIM sampling — achieving 17% FID improvement with 85% reduced training time.
computer-vision deep-learning pytorch transfer-learning unet attention-mechanism cifar10 fid inception-score diffusion-models pytorch-lightning conditional-generation ddim generative-ai classifier-free-guidance fp16-training parameter-efficient-training
-
Updated
Mar 3, 2026 - Jupyter Notebook