The simple version
- Imagine a TV screen full of fuzzy static.
- An AI image maker has learned how to clean static away, a little at a time.
- Your words tell it what picture to aim for while it cleans.
- After many small clean-ups, a clear picture appears.
- Video makers do the same thing for many pictures in a row.
Why it matters now
On 9 November, Disney named Adobe its AI partner in a multi-year deal. According to Adobe, it will build custom image models trained on Disney’s own characters, settings and visual styles, using Adobe’s Firefly Foundry service (Adobe newsroom). Big studios now care less about whether AI can make pictures and more about which pictures it learned from and whose style it produces.
Most of today’s image and video generators rest on one idea, the diffusion model. Once you understand it, a lot of the news about AI art, deepfakes and copyright becomes easier to follow.
How it works, step by step
Think of a sculptor who starts with a rough block of stone and removes a little at a time until a statue appears. A diffusion model works the same way, except its block of stone is random noise (like TV static) and what it removes is noise.
Step 1: Learn by wrecking pictures
During training, the model is shown millions of images, each with a caption. For each one, the training process adds a little random noise, then a bit more, then more, until the picture is pure static.
At every stage, the model practises one job: look at the noisy picture and guess what noise was added. If it can guess the noise, it can subtract it. This approach was set out clearly in a 2020 paper by Jonathan Ho, Ajay Jain and Pieter Abbeel, which showed it could produce high-quality images (Ho et al., 2020).
Step 2: Link pictures to words
The captions matter. Each caption is turned into numbers that capture its meaning, called an embedding (see our explainer on tokens and embeddings). The model learns which kinds of noise to remove when the caption says “a red bicycle at sunset” and which when it says “a cat in a spacesuit”.
Step 3: Work in a smaller space
Cleaning up every pixel of a large picture is slow. So many modern systems first squash the image into a compact summary, a sort of shorthand sketch called a latent space, and do the cleaning there. A separate part of the model then expands the finished sketch back into a full picture. This trick, called latent diffusion, was described by Robin Rombach and colleagues and made high-quality image generation much cheaper to train and run (Rombach et al., 2022).
Step 4: Generate from scratch
When you type a prompt, the model starts with fresh random static. It removes a little noise, checks against your prompt, removes a little more, and repeats, often dozens of times. A setting usually called guidance controls how strictly it follows your words compared with simply making a natural-looking picture (Ho and Salimans, 2022). Because the starting static is random, the same prompt gives a different picture each time. Many tools let you fix the starting noise with a number called a seed, so you can recreate an image you liked and change just one word of the prompt.
Step 5: Add time for video
A video is a stack of images through time. Video generators clean up noise across many frames together, so that objects stay consistent as they move. Many newer systems swap the older image network for a transformer, the same building block used in chatbots (Peebles and Xie, 2022). OpenAI described its Sora video model this way in 2024: a diffusion transformer that works on small “patches” of video in space and time (OpenAI, 2024).
A worked example: “a red bicycle at sunset”
- Start: a square of pure static, like a detuned TV.
- Early steps: the model removes noise in a way that matches its learned idea of “sunset”, so a warm orange glow and a dark lower half appear. Nothing has edges yet.
- Middle steps: shapes firm up. A horizon line forms; a dark outline that could be a bicycle settles near the centre because the caption pushes it there.
- Late steps: detail arrives: spokes, a red frame, light glinting on the handlebars.
- Finish: if the model works in latent space, the compact sketch is expanded to a full-size image.
Now imagine the same request sent to a custom model, like the ones Disney and Adobe describe. The steps are identical. What changes is what the model learned in Step 1. A model that has been further trained, or fine-tuned, on one studio’s drawings will remove noise in the direction of that studio’s style, so the bicycle comes out looking like it belongs in one of their films. That is why companies care so much about which pictures go into training: the training pictures decide the look of everything that comes out.
Run it again and you get a different bicycle and a different sunset, because the static you started from was different. Change the guidance setting and the result moves between “very literal” and “looser, more natural”.
What people get wrong
“It cuts and pastes bits of existing images.” It does not stitch together stored pictures. It learns statistical patterns and generates new pixels. That said, models can sometimes reproduce close copies of images that appeared many times in their training data (Carlini et al., 2023), which is one reason copyright is a live issue (see training data and copyright).
“It understands what it is drawing.” It has learned what pictures with certain captions tend to look like. That is why details such as hands, lettering inside images and the right number of objects have often come out wrong.
“Image and chatbot AI are completely different things.” They share a lot: embeddings to understand prompts, and increasingly the same transformer building blocks. Many systems now handle text and images together, which is called being multimodal.
“You can always spot an AI image.” Older tell-tale signs are fading. Labels and hidden marks are becoming the main way to tell, which is why AI watermarking matters.
Where it is heading
Custom models trained on your own material. The Disney and Adobe alliance is an example of a company training private image models on content it owns, rather than relying only on general-purpose models (Adobe newsroom).
From pictures to simulated worlds. OpenAI’s Sora report framed video generation as a step toward models that simulate the physical world (OpenAI, 2024). That links to our explainer on world models.
Labels and provenance. As generated images get harder to spot, the focus is moving to marking where content came from. Our quick piece on AI watermarking covers how that works.
Key terms
Sources
- Adobe newsroom, Adobe and The Walt Disney Company alliance, 9 November 2026. news.adobe.com
- Ho, Jain and Abbeel, Denoising Diffusion Probabilistic Models, 2020. arXiv 2006.11239
- Rombach et al., High-Resolution Image Synthesis with Latent Diffusion Models, 2022. arXiv 2112.10752
- Ho and Salimans, Classifier-Free Diffusion Guidance, 2022. arXiv 2207.12598
- Peebles and Xie, Scalable Diffusion Models with Transformers, 2022. arXiv 2212.09748
- Carlini et al., Extracting Training Data from Diffusion Models, 2023. arXiv 2301.13188
- OpenAI, Video generation models as world simulators, 2024. openai.com
