Diffusion Model

Definition

An AI image generation technique that starts with noise and gradually refines it into a coherent image. Used by FLUX, Stable Diffusion.

Why It Matters

Diffusion replaced GANs as the dominant image-generation approach around 2022 because it trains more stably and produces sharper, more controllable outputs. It's also why image generation is slow, each output requires 20–50 incremental denoising steps rather than one forward pass.

Key Points

• The forward process adds Gaussian noise over T steps (typically 1000); the reverse process learns to predict and subtract that noise incrementally.
• DDIM (2020) made sampling 10–50× faster than DDPM with no retraining by skipping steps in the denoising trajectory.
• Classifier-free guidance (CFG) scale controls fidelity vs. diversity, CFG 7 is a common default; higher values produce sharper, more prompt-faithful images but reduce variety.
• FLUX.1 and Stable Diffusion 3 use a DiT (Diffusion Transformer) backbone rather than the older UNet. It scales better and renders readable text within images more reliably.
• Latent diffusion (used in SDXL and FLUX) operates in a compressed latent space rather than pixel space, 8–16× smaller, making generation 4–8× faster.

Example

Stable Diffusion XL and FLUX.1 are diffusion models. Generating a 1024×1024 image takes ~20 steps on an A100, each step removes a little noise from random gaussian noise until a coherent picture emerges, guided by the text prompt.

Related Terms

• Computer VisionAI that can understand and analyze images and video content.
• GAN (Generative Adversarial Network)An older image generation technique using two competing neural networks.
• ParameterA trainable weight in an AI model. Larger models have more parameters (7B, 70B, 400B).

Quick Facts