Every model, with the prompts that work in it

OpenAI's first-generation image model. It follows instructions closely and can accurately render text and complex layouts.

Google Gemini 2.5 Flash generation image model. Fast output, great for editing images and keeping characters consistent, easy for anyone to use.

Kuaishou Kling. On the image side, it excels at portraits, Chinese-style art, and product shots, with delicate lighting and material rendering; prompts can be written in Chinese, and reference images can lock in characters and composition. For the video model from the same company, see Kling Video.

Released in April 2026. Faster generation than the previous version, more accurate edits, flexible sizes, can take an image directly as input, and noticeably better text and layout in images.

ByteDance's Seedream series of image models. Strong Chinese semantic understanding; reliable for portraits, product shots, and posters, with delicate lighting and materials. Just write prompts in Chinese; choose aspect ratio and resolution in the UI or include them in the prompt.

The most painterly one, strong in lighting, texture, and composition. Using Style Reference (--sref) and Omni Reference (--oref) locks in style and characters; put parameters at the end of the prompt.

xAI's image generation (Grok Imagine). In X, one sentence creates an image, and edits (change background, change style, expand image) are done just by saying so; the style leans realistic with strong internet vibes, great for quickly testing ideas.

Released in September 2026, the most precise editing generation, with Flare (faster) and Sunburst (more precise) versions, both more reliable at text and instruction following.

Powered by Gemini 3 Pro, it reasons before generating images, and understands world knowledge and real-time info. Great for posters, infographics, and visualizations, with text in images rarely going wrong.

Gemini 3.1 Flash generation, the fast next generation. It can search the web, make infographics, and supports 1K / 2K and custom aspect ratios.

ByteDance's Seedream 5.0 series. It reasons before generating images and can search online for references. Especially strong at product shots and posters.

Kuaishou Kling's video model. It handles both text-to-video and image-to-video, with fairly natural character motion and camera movement; it excels at Chinese-style, realistic, and commercial looks. Just write prompts in Chinese, and providing a first-frame image locks in the composition and characters.

ByteDance's new-generation joint audio-video generation model, up to 30 seconds at a time and extendable twice; it understands the composition and camera language of reference videos, and supports professional editing like green screen and white model.

A unified audio-video joint generation architecture that can take text, images, audio, and video as references at the same time; stable motion, audio-visual sync, and use reference materials to control performance, lighting, and camera movement.