A diffusion model is a generative model that creates images — and increasingly video and audio — by learning to reverse a corruption process. In training, real images are progressively drowned in noise while the model learns to undo each step. At generation time it runs that skill backwards: start from pure random static, denoise step by step, and a coherent image emerges — steered toward your prompt by a text encoder scoring how well the developing image matches the words.
Why this approach won image generation
Compared to its predecessor GANs, diffusion is more stable to train and much better at prompt fidelity — the guidance from the text encoder acts at every denoising step, giving the process dozens of chances to course-correct. Most systems also work in latent space (a compressed representation, the “latent diffusion” of Stable Diffusion fame) rather than raw pixels, which is what makes generating a 2K image computationally sane.
Diffusion vs. LLMs — two different bets
An LLM builds output sequentially, one token at a time; a diffusion model refines the whole canvas at once, coarse to fine. That’s why diffusion excels at global composition. The camps are blending in 2026: many image models pair an LLM-style transformer as the denoising backbone (DiT architectures), some video models generate autoregressively, and hybrid systems like GPT Image 2 fold generation into a multimodal model. “Diffusion” today names the denoising technique more than a fixed architecture.
Where you’ve seen it
Diffusion (or a hybrid of it) powers most of the image models we track — FLUX.2, Stable Diffusion 3.5, Midjourney v8.1 — and much of AI video. For a concrete pick, see the best AI image generators ranking.