News
EveryLLM

Images out

Draws or edits images itself. 19 models on EveryLLM so far.

Some language models draw. Instead of handing a prompt to a separate image generator, a natively multimodal model can produce image tokens itself, which lets it follow detailed instructions, render legible text inside pictures and edit an image through conversation, keeping faces and objects consistent from one turn to the next.

From DALL-E to native images

OpenAI's first DALL-E (January 2021) was itself a transformer that generated images as tokens; later image generators mostly moved to diffusion. In March 2025 OpenAI added native image generation to GPT-4o in ChatGPT, later offered as GPT Image in its API, and the feature went viral for pictures in the style of Studio Ghibli. Google's Gemini 2.5 Flash Image, nicknamed Nano Banana (August 2025), became popular for editing photos.

Concerns

Image output brings the disputes of image generation into chat: the use of artists' work in training, deepfakes of real people, and harmful or sexual imagery. Providers add filters and invisible watermarks, such as Google's SynthID, and metadata under the C2PA content-credentials standard. Rules differ widely between providers, and xAI's Grok drew criticism and regulatory scrutiny in several countries over sexualised images of real people.

Famous images out models

Still to be written: GPT Image, GPT-4o.

Further reading

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.