What does CLIP-style contrastive training produce?

**CLIP** is two encoders trained together. One encoder reads images and turns them into vectors. The other encoder reads text and turns it into vectors. They never touch each other's weights. What ties them together is the loss function. During training, each batch contains many (image, caption) pairs that actually go together. The loss asks: 'Is the image's vector closer to its OWN caption's vector than to all the other captions in the batch?' If yes, great. If no, the gradient nudges both encoders to fix it. After enough training, the two encoders end up producing vectors in the same coordinate system. An image of a dog and the text 'a dog' both land in roughly the same region. That shared coordinate system is the whole point. The other options are wrong because: there is no single transformer that mixes modalities at the input layer; CLIP doesn't generate anything; and CLIP doesn't include a classifier head. Zero-shot classification with CLIP is a downstream USE, not what training outputs.

**CLIP-style training** produces two specialized encoders whose outputs share a single vector space. The vision encoder maps images to `d`-dimensional vectors; the text encoder maps captions to vectors of the same dimension. Crucially, the encoders never share parameters. They share their output geometry. The loss is what does the work. For each training batch of N image-caption pairs, CLIP builds the N x N cosine-similarity matrix and applies a symmetric InfoNCE loss that maximizes the diagonal and minimizes the off-diagonal: $$\mathcal{L} = -\frac{1}{2N} \sum_i \left[ \log \frac{e^{S_{ii}/\tau}}{\sum_j e^{S_{ij}/\tau}} + \log \frac{e^{S_{ii}/\tau}}{\sum_j e^{S_{ji}/\tau}} \right]$$ After training, both encoders are deployed independently. You can embed a corpus of millions of images once, pre-compute the vectors, and then serve text queries cheaply by embedding only the query string. The shared space is what makes cross-modal nearest-neighbor search trivially possible. The other three options describe things CLIP is NOT. Diffusion models generate images; CLIP only embeds them. Classifier heads need labeled training data and a fixed label set; CLIP works zero-shot. And a single transformer that ingests either modality is a different architecture entirely (Perceiver, ImageBind's later variants), not what CLIP introduced.