Step by Step
F
Forward process — gradually adding noise
During training, the forward process gradually adds Gaussian noise to a real image over roughly 1000 steps, until the image looks like pure random static.
Example: taking a clear photo and gradually adding a small amount of random noise at each of 1000 steps, until it's completely indistinguishable from random static.
B
Backward process — learning to remove that noise
A neural network is trained to predict and remove the noise added at each step, learning the reverse denoising process.
Example: training a network to look at a noisy image at step 500 and predict what noise was added, so it can be removed to move one step closer to the clean image.
I
Image generation — starting from pure noise, denoising repeatedly
At actual generation time, the process starts from pure random noise and applies the learned denoising process roughly 1000 times, with a coherent image gradually emerging from what started as pure static.
Example: Stable Diffusion starting with pure random noise and applying its learned denoising process step by step, until a coherent, recognizable image emerges after all steps are complete.
Applied Walkthrough
1
A team wants to generate a brand-new, coherent image using a trained diffusion model.
2
They start the generation process with pure random noise — essentially indistinguishable static, with no image content at all.
3
The model applies its learned denoising process repeatedly, roughly 1000 times, with each step removing a small amount of predicted noise.
4
After all 1000 denoising steps, a coherent, recognizable image has gradually emerged from what started as pure random static — this reverse process is exactly what tools like Stable Diffusion, DALL-E 2, and Midjourney use.
Exam Application
Exams test whether you understand both halves of the diffusion process (forward: gradually adding noise during training; reverse: learning to remove that noise, used for actual generation) and whether you know diffusion models have displaced GANs as the state-of-the-art for image generation, due to greater training stability, better quality, and easier text conditioning.
⚠ Common Trap
The most common trap is confusing the forward process (used during training, adding noise to real images) with the reverse process (used during actual generation, removing noise starting from pure static). Generation always uses the reverse (denoising) process — the forward (noise-adding) process is purely a training-time mechanism.
✓ Quick Self-Check
1. What happens during the forward process of a diffusion model?
Gaussian noise is gradually added to a real image over about 1000 steps, until it becomes pure random static.
Tap to reveal / hide
2. What does the reverse process learn to do?
Predict and remove the noise added at each step, gradually reconstructing a clean image.
Tap to reveal / hide
3. What does actual image generation start from?
Pure random noise, which is then denoised repeatedly using the learned reverse process.
Tap to reveal / hide
4. Name three well-known tools that use diffusion models for image generation.
Stable Diffusion, DALL-E 2 (or DALL-E 3), and Midjourney.
Tap to reveal / hide
5. Which process (forward or reverse) is used at actual image generation time?
The reverse (denoising) process — the forward process is purely a training-time mechanism.
Tap to reveal / hide