NCA-GENM - Multimodal Data - Section 3.1

Generate images from English text prompts using CLIP.

Explain how CLIP aligns text and image representations in a shared embedding space, enabling text prompts to guide image generation. Apply prompt engineering techniques to steer CLIP-conditioned outputs toward the desired visual concept.

CLIP

Practice question for this objective

Free sampleMultimodal Datamedium

A generative image pipeline uses CLIP to evaluate whether candidate images produced by a diffusion model match a given text prompt. Which property of CLIP makes it suitable for this guidance role?

  • ACLIP is a pixel-level decoder that reconstructs images token by token from a text sequence.
  • BCLIP is a variational autoencoder that compresses images into a latent distribution conditioned on text.
  • CCLIP uses autoregressive cross-attention to generate image patches in the order determined by the text prompt.
  • DCLIP projects both images and text into a shared embedding space where cosine similarity measures semantic alignment. Correct
Explain how CLIP's shared embedding space enables it to guide text-to-image generation without acting as a generator itself. CLIP (Contrastive Language-Image Pretraining) is trained to maximise cosine similarity between matched image-text pairs and minimise it for unmatched pairs. This contrastive objective produces a joint embedding space where the semantic content of a caption and its corresponding image land close together. A diffusion model or other generator can exploit this by computing the CLIP similarity between its current output and the target text prompt and using the gradient of that score to steer the denoising process toward higher semantic fidelity.

Why A is wrong: CLIP contains no decoder and produces no pixels; it is a dual-encoder trained to score text-image similarity, not to generate or reconstruct images.

Why B is wrong: A variational autoencoder uses encoder-decoder structure with a probabilistic latent space; CLIP uses contrastive learning across two modality-specific encoders with no variational bottleneck.

Why C is wrong: Autoregressive patch generation describes models such as image GPT; CLIP generates no patches and does not process modalities sequentially in a generative manner.

Why D is correct: Contrastive training aligns image and text embeddings so that a matched pair scores high similarity, enabling CLIP to act as a differentiable semantic signal for guiding a separate generator.

See more NCA-GENM practice questions, answers explained.

More in this domain

Back to all Multimodal Data objectives, or the NCA-GENM cert hub.

Examworthy is not affiliated with or endorsed by NVIDIA. Original, blueprint-aligned practice material only.