Which statement correctly separates image understanding from image generation in a VLM?

A vision-language model that **understands** an image works in one direction. The image goes through a vision encoder, the features are projected into the language model's token space, and the language model writes text about it. Pixels go in, words come out. Generating an image is the opposite job, and it needs a different part. To make a picture from a prompt, the model needs an image decoder — usually a diffusion model or an autoregressive image-token decoder — that turns a description into actual pixels. The option that gets this right is the one saying understanding uses an encoder into the LLM while generation needs a separate image decoder. The wrong options all assume you can run the encoder backwards. You cannot. An encoder throws away most of the raw pixels and keeps a compact summary, so there is no way to recover an image from it. A model that captions perfectly may have no image decoder at all.

The correct choice is the one that keeps two paths separate: understanding runs the encoder into the LLM, generation needs its own image decoder, and neither implies the other. Understanding is a perception pipeline. A vision encoder maps the image to feature vectors, and a connector projects them into the LLM's token space. The LLM then attends over those visual tokens alongside text to produce a description or an answer. Nothing in this path emits pixels — the output is text. Generation is a synthesis pipeline. It starts from a conditioning signal (often a text prompt) and runs a dedicated **image decoder** to produce pixels or discrete image tokens. The two leading families are diffusion decoders, which denoise from random noise, and autoregressive decoders, which emit image tokens one at a time. The distractors all rest on the same false premise: that a vision encoder is reversible. It is not — it compresses pixels into features and there is no inverse path. So a model that captions accurately has no built-in way to generate, and any to any models earn both abilities only by wiring both pipelines together.