Agendamentos de Amostragem e Orientação

Training gives you a denoiser. Sampling is where you cash it in, running the reverse steps to turn noise into a finished image.

Each reverse step calls the U-Net once. More steps usually means crisper results but a slower generation. ⏳

The original DDPM sampler is faithful but slow, often needing hundreds or even a thousand tiny denoising steps.

DDIM skips many steps while keeping quality high, so you can sample in twenty or fifty steps instead of a thousand.

A sampling schedule decides which timesteps you actually visit. Fewer, well-chosen steps trade a little detail for big speed.

Left alone, the model produces a valid image but you cannot steer it. You get something plausible, not necessarily what you asked for.

To get control, you condition the U-Net on a prompt, often a text embedding, so it aims toward your description.

Classifier-free guidance runs the model twice, with and without the prompt, then pushes the result toward the prompted version.

A guidance scale sets how hard to push. Higher values follow the prompt more closely but can hurt realism if overdone.

Too low and the image ignores you; too high and it looks harsh. A scale around seven is a common sweet spot to start from.

Steps control speed and detail; guidance controls how closely you obey the prompt. Tuning both is how you dial in your output. 🎛️

What does raising the guidance scale do?

Schedules like DDIM cut steps for speed, and guidance steers toward your prompt. Tune steps and guidance scale to balance quality and control. 🎉