ExamworthyExamworthy
Open the appStart free

NCA-GENM - Experimentation - Section 1.3

Control image output with context embeddings.

Describe how embeddings encode semantic information as dense vectors that guide a generative model toward a desired visual concept. Apply conditioning techniques - such as text or image embeddings - to steer output style, composition, and content.

Embeddings

Practice question for this objective

Free sampleExperimentationmedium

An engineer conditions a diffusion model on both a reference photograph and a text description to guide the style and content of a generated image. The image is first encoded into a dense embedding by a vision encoder, and that embedding is concatenated with the CLIP text embedding before being passed to the denoiser. What type of conditioning does this represent?

  • AUnimodal text conditioning, because the vision encoder output is a derived representation of a pre-existing text-captioned dataset rather than an independent signal
  • BLatent inpainting, because the reference photograph is encoded into the latent space and merged with the noisy latent during each denoising step
  • CMultimodal context conditioning, where embeddings from separate modalities are combined to jointly steer the denoiser through its cross-attention or concatenation pathway Correct
  • DTextual inversion, where the reference photograph is optimised into a new pseudo-token embedding added to the text vocabulary
Distinguish multimodal context conditioning from related approaches when both image and text embeddings are combined to control diffusion model output. Multimodal conditioning uses embeddings derived from multiple input modalities to jointly influence the generation process. In this scenario a vision encoder produces an image embedding and a CLIP text encoder produces a text embedding; the two are concatenated and passed into the denoiser's conditioning pathway. Each embedding contributes complementary information: the image embedding captures style and visual structure while the text embedding specifies semantic content. This is different from unimodal text conditioning, latent inpainting, and textual inversion, each of which involves only a single type of input signal or a distinct training-time operation.

Why A is wrong: The system processes a real image through a vision encoder at inference time to produce an image embedding, which is a distinct modality from the text prompt; treating both inputs as a single text-derived signal mischaracterises the architecture.

Why B is wrong: Latent inpainting specifically masks a region of the reference image's latent and denoises only that region; here the full reference image embedding is used as a conditioning signal rather than as a partial latent to be completed.

Why C is correct: When embeddings from two different modalities, an image and a text description, are merged and supplied together as context for the denoiser, the system performs multimodal conditioning. Each modality contributes information that the cross-attention or concatenation mechanism can use to control different aspects of the output.

Why D is wrong: Textual inversion is a training-time personalisation technique that learns a new token to represent a concept from example images; here the image is simply encoded and used directly at inference time rather than being learned as a vocabulary token.

See more NCA-GENM practice questions, answers explained.

More in this domain

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

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