diffusion-e2e-ft

Why we included this project

This repository ships the training and inference code behind the WACV 2025 paper on fine-tuning image-conditional diffusion models, along with pretrained checkpoints on Hugging Face. The released models estimate monocular depth and surface normals from a single image in a single deterministic forward pass, and since they run through standard diffusers pipelines, you can call them with a few lines of Python instead of assembling the whole model stack yourself. For teams working on geometric computer vision, that matters in practice: end-to-end fine-tuning on task-specific losses collapses the usual many-step diffusion sampling into one step, which makes depth and normals inference noticeably cheaper and faster. The checkpoints and hosted demos give you a concrete baseline to evaluate before committing to a heavier approach.