Skip to content
BENIBIX Learn

AI Image Generation Models

Level 3 · Picture, video and speech models

The models that make pictures from words, such as Stable Diffusion, FLUX and GPT Image. And CLIP, which only scores a match.

Course contentsLesson 9 of 13

After this lesson you can

  • Name four picture models and who makes them
  • Explain what CLIP does and does not do
  • Say which picture models you can download

11 mincovers 6 glossary entries

Worth reading first: Open Model Licenses

Which models turn words into pictures

An image model turns words into a picture. Stable Diffusion, made by Stability AI and out in August 2022, was the first strong one with open weights. Black Forest Labs, started in 2024 by people who helped build it, makes FLUX. OpenAI makes GPT Image, and Google makes Nano Banana. You type a scene, and a picture comes back in seconds.

Most image models work by diffusion. They start from random noise and clean it up, step by step, until it matches the words. The first versions of Stable Diffusion read the words through a model called CLIP, from OpenAI. Run on your own graphics card, Stable Diffusion can also be trained on your own style.

CLIP came out in January 2021. It was trained on about 400 million pictures paired with their captions. It does not draw. It scores how well a picture and some words fit together. So it can sort photos by a typed search, or steer another model as it draws.

FLUX is known for clear text inside pictures, such as the words on a shop sign. It comes in several versions with different terms. Some can be downloaded under Apache 2.0. Some are only for use that earns no money. The strongest are reached only through a paid API.

GPT Image and Nano Banana live inside chat apps. GPT Image took over picture making in ChatGPT in March 2025. Nano Banana is the pet name for Google's Gemini image models. The first, Gemini 2.5 Flash Image, came out in August 2025. Both are good at changing a photo you upload, not just making new ones.

Stability AI shares many of its models with open weights, under its own license. It also makes models for sound, video and 3D. With the weights open, a studio can run them on its own machines. Its art then stays private until launch day.

People think every picture app runs Stable Diffusion, since it came first. But many models work by diffusion. Stable Diffusion is just one named family. Another slip is to trust the picture. These models still get hands, small text and counts wrong, so look closely before you print.

Picture model families side by side

FamilyMakerOpen weightsKnown for
Stable DiffusionStability AIYes, with a licenseRuns on your own card
FLUXBlack Forest LabsSome versionsClear text in pictures
GPT ImageOpenAINoPictures inside ChatGPT
Nano BananaGoogleNoEditing your own photos
CLIPOpenAIYesMatching, not drawing
The words always go in. A photo to change can go in beside them.

Worked example

Emma makes a bakery poster

Emma wants a poster that says 'Fresh Bread Daily'. Her first model writes 'Fresh Braed Dialy' in 3 of 4 tries.

She tries a FLUX model, and all 4 tries spell the words right. She picks the best and prints 50 copies.

The words for this lesson6

Check yourself

Does CLIP make pictures?

Finished this lesson?

Marking it complete also marks its 6 glossary entries learned, which updates your roadmap and your saved page. 0 are already marked.