Diffusion Models Projects

Take a pretrained Stable Diffusion v1.5 apart and learn the knobs that matter. Look inside the pipeline (text encoder, UNet, VAE latents), make generations reproducible with seeded generators, sweep the guidance scale and score prompt adherence with CLIP, write the classifier-free-guidance denoising loop by hand (negative prompts included), race the DDIM, Euler and DPM-Solver++ schedulers at low step counts, and edit an image with img2img at several strengths. The first cell downloads the fp16 weights (~2.7 GB, about a minute).

Build a denoising diffusion model end to end. Implement the closed-form forward noising process and watch clothing dissolve into Gaussian noise, train a small diffusers UNet2DModel to predict that noise (mixed precision and an EMA copy of the weights), then write the reverse sampling loop yourself with both the 1000-step DDPM sampler and a 50-step DDIM sampler. Finish with a grid of generated garments and a strip showing one sample emerging from noise. Training is capped at about two minutes on a T4.