Class 7 · CBSE AI · Strand C — NLP, Vision, and LLMs Deep-Dive

How AI image generation works — diffusion models for Class 7

Diffusion models build images by reversing noise, starting from static. How tools like Stable Diffusion create art. For Class 7.

Practise this on Dhee Learning See full Class 7 syllabus

What this concept actually says

  • Diffusion models generate images by learning to reverse a noise-adding process — starting from pure noise and gradually removing it to reveal a coherent image
  • Text-to-image systems combine a text encoder (understanding the prompt) with a diffusion model (generating the image)
  • The training data for image generators includes billions of images scraped from the internet, creating significant copyright, consent, and bias issues

An analogy your child will recognise

Sketch-to-painting in a traditional art class

A student starts with a noisy, scratchy rough sketch and gradually refines it — adding detail, correcting proportions, layering colour — until it becomes a finished painting. Diffusion models run this process in reverse: they start with the finished 'noise' and learn the refinement process backwards, then play it in reverse to generate new images.

Developing a photograph in a darkroom

In old darkroom photography, a photographic print starts invisible on white paper and gradually emerges in the developer solution — vague shapes first, then details, then the full image. Diffusion generation feels similar: the image is completely invisible in the starting noise, then shapes emerge step by step as the model reverses the noise.

Common misconceptions to watch for

  • Image generation models 'retrieve' images from their training data — they generate new pixel patterns through a learned denoising process; they don't store or retrieve images.
  • Text-to-image AI is purely creative and raises no ethical issues — it raises significant concerns around consent (training on artists' work without permission), copyright, deepfakes, and cultural misrepresentation.

Key facts in one breath

  • Stable Diffusion (2022) was the first major open-source diffusion model, enabling anyone to run image generation locally — dramatically democratising the technology.
  • A typical diffusion model runs 20–50 denoising steps to generate a single image, with each step calling the neural network once.
  • CLIP (Contrastive Language-Image Pretraining) is the component that connects text prompts to image generation — it was trained on 400 million image-text pairs from the internet.
  • Deepfake detection is now an active research area specifically because diffusion models can generate photorealistic fake images of real people with minimal effort.

How Dhee Learning teaches this — the 3-stage question loop

Every Dhee Learning session for this concept follows three stages. We share the questions Dhee actually asks, so you can hear what a session sounds like.

Stage 1 — Surface

Imagine trying to reconstruct a sand castle from a pile of jumbled sand. The only way to do it is to have watched thousands of sand castles being knocked down — so you know what the 'unscrambling' should look like. How does this connect to how an image generator might work?

Rote answer

"Diffusion models learn to remove noise from images."

Understood

"The model watches millions of images being gradually scrambled into noise — like a sand castle being kicked apart in slow motion — and learns to reverse each step. To generate a new image, it starts with a pile of random noise (blank sand) and applies that learned 'unscrambling' over and over until something meaningful appears."

Stage 2 — Reasoning

A text-to-image model is given the prompt 'A Rajasthani woman in traditional dress'. Explain why the generated image might systematically favour certain visual styles over others — and what that reveals about its training data.

Follow-up Dhee may use: How would you audit whether a text-to-image model has this kind of systematic bias for Indian cultural representations?

Stage 3 — Application

An art teacher wants to use a text-to-image AI to help students explore visual design ideas. List three ways this could genuinely help students — and three serious risks she should discuss with them before they start.

Misconception Dhee watches for: Thinking that because the output is a new image it doesn't involve copying — diffusion models are trained on existing human artwork, and their outputs statistically reflect that training data, raising genuine questions about originality and attribution.

Related concepts

Class 7

AI, training data and copyright — the big debate for Class 7

Class 7

What is multimodal AI? Models that see, read and hear

Class 7

The YouTube rabbit hole — how recommendation AI narrows what you see

Class 7

Data visualisation with matplotlib for beginners — Class 7

Want your child to actually understand this?

Dhee turns this concept into a 15-minute spoken session — asking, listening, and probing — so your child builds the idea themselves.

Download Dhee Learning App Class 7 syllabus

Frequently asked questions

What is image generation — diffusion models — explained for kids? +

Diffusion models build images by reversing noise, starting from static. How tools like Stable Diffusion create art. For Class 7.

What's the most common mistake children make about this concept? +

Image generation models 'retrieve' images from their training data — they generate new pixel patterns through a learned denoising process; they don't store or retrieve images.

How does Dhee Learning teach this in a Class 7 session? +

Dhee opens with a question — for example: "Imagine trying to reconstruct a sand castle from a pile of jumbled sand. The only way to do it is to have watched thousands of sand castles being knocked down — so you know what the 'unscrambling' should look like. How does this connect to how an image generator might work?" — listens to your child's answer, then probes the reasoning behind it. The session ends when the child can apply the idea to a brand-new situation, not just recall it.