The name of a generation type says it all: what goes in, then what comes out. With text-to-image, you give a text and get an image back. These names also work as filters in the Models tab.
The prompt
Everything you write to the model is called a prompt. With text-to-image, it's your only starting point. When you also give an image or a video, the prompt becomes the instruction: what the model should do with it.
Create from text
text-to-image
You describe an image, the model creates it. Ideal for exploring an idea, a mood or a set.
text-to-video
You describe a shot, the model films it. The prompt also describes the motion.
text-to-3D
You describe an object, the model creates the 3D file.
Transform an image
image-to-image
You give an image and an instruction. The model keeps whatever you don't ask it to change: a new set, different lighting, another style.
These models often have Edit in their name: Flux 2 Edit, GPT Image 1.5 Edit.
image-to-video
You give an image, the model brings it to life. Your image sets the framing and the style, the prompt describes what moves.
When several images must be respected, such as a character or a set, look for Reference in the name: Seedance 2.5 (Reference to Video).
image-to-3D
You give an image of an object, the model rebuilds it in 3D. Some models take several views of the same object: they have Multi-View or Multi Image in their name.
Transform a video
video-to-video
You give a shot, the model transforms it while following its motion. Handy for changing a character or a set without reshooting.
These models often have Reference or Edit in their name: LTX-2.3 22B Reference Video To Video, Kling Video O3 Pro Video To Video Reference.
In your software
An image or video input plugs into the node, like any other node in your software. With text-to-image, the prompt is enough.
Anatomy of a node shows where everything is.