what’s a diffusion model?
A diffusion model is a generative architecture that creates images, video, or audio by iteratively denoising pure random noise into a coherent output.
diffusion is the architecture behind Midjourney, Stable Diffusion, DALL-E 3, Sora, and most modern image and video generators. training has two halves.
in the forward process, you take a real image and progressively add small amounts of random noise across hundreds of steps until the image becomes pure static. in the reverse process, a neural network learns to undo one step of noise at a time: given a slightly noisy image, predict a slightly less noisy version. at generation time, you start from pure noise and apply the learned denoiser repeatedly, sculpting structure out of static.
the 2022 latent-diffusion paper made this practical at consumer scale by running the diffusion in a compressed latent space instead of full-resolution pixels, which is what made Stable Diffusion runnable on a single consumer GPU and triggered the image-generation explosion of 2022-2023.
what a layperson needs to know:
almost every AI image you've seen in the last 2-3 years from a name-brand tool is diffusion-generated. Midjourney, Stable Diffusion, DALL-E 3, GPT Image 1, Flux, and Sora all run on diffusion. that covers most of the visible AI-content on the internet today.
different diffusion models leave different fingerprints. a Midjourney output is statistically distinguishable from a Stable Diffusion output, which is distinguishable from a DALL-E 3 output. detectors have to be retrained for each major architecture and version. amige.'s panel runs multiple detectors to stay current with new releases.
diffusion models can be conditioned on more than just text. depth maps, sketches, other images, masked inpainting regions... the conditioning flexibility is why diffusion has displaced earlier architectures for almost every production use case.
diffusion models do not “look up” or collage training images. the model stores a learned denoising function, not a library of images. outputs can occasionally resemble training data when prompts are extremely specific (the Getty watermark case against Stable Diffusion is the canonical example), but the default mechanism is statistical synthesis, not retrieval.
questions
what is a diffusion model?
a diffusion model is a generative architecture that builds images, video, or audio by denoising pure random noise into a coherent output one step at a time. during training it learns to undo noise step by step. at generation it starts from static and applies that learned denoiser over and over, sculpting structure out of the noise. it powers Midjourney, Stable Diffusion, DALL-E 3, and Sora.
what AI images are made with diffusion models?
almost every AI image from a name-brand tool in the past two to three years. Midjourney, Stable Diffusion, DALL-E 3, GPT Image 1, Flux, and Sora all run on diffusion. that covers most of the visible AI-content on the internet today.
do diffusion models copy training images?
no. a diffusion model stores a learned denoising function and synthesizes outputs from it, with no lookup table of images. outputs can resemble training data under narrow, specific prompts, and Stable Diffusion reproducing the Getty watermark is the canonical case. the default mechanism is synthesis, not retrieval. a 2025 UK court reached the same conclusion: the model holds trained parameters, not stored copies.
can different diffusion models be told apart?
yes. each diffusion model leaves a distinct statistical fingerprint, so a Midjourney output reads differently from a Stable Diffusion or DALL-E 3 one. detectors have to be retrained for every major architecture and version. amige.’s panel runs multiple detectors to stay current with new releases.
sources.
- 01Ho, Jain, Abbeel — Denoising Diffusion Probabilistic Models (DDPM, arXiv:2006.11239)The 2020 paper that revived diffusion for image generation.
- 02Rombach et al. — High-Resolution Image Synthesis with Latent Diffusion Models (arXiv:2112.10752)The latent-diffusion paper that made Stable Diffusion possible.
- 03