Step by Step
D
Discriminative CV — given an image, output a label
Discriminative computer vision tasks take an image as input and output a label, bounding box, or pixel-level classification: classification (what?), detection (what and where?), and segmentation (what at every pixel?).
Example: a discriminative model taking a photo as input and outputting the label "cat," or a bounding box around a detected object, or a pixel-level segmentation map.
G
Generative CV — given a prompt or noise, create an image
Generative computer vision tasks take a prompt or random noise as input and create an entirely new image as output, using techniques like GANs, diffusion models, or VAEs.
Example: a diffusion model taking a text prompt as input and generating an entirely new, previously non-existent image matching that description.
⭐
The blurring line — vision-language models combine both
Modern vision-language models like CLIP, DALL-E, and Stable Diffusion increasingly combine both discriminative understanding and generative creation within the same underlying representation space, blurring the traditional line between these two categories.
Example: CLIP being able to both understand/classify images (a discriminative capability) and guide image generation (a generative capability), using shared underlying representations for both.
Applied Walkthrough
1
A team needs one system to classify whether an image contains a cat or dog, and a separate system to generate brand-new, realistic images of cats.
2
The classification task is discriminative — given an image, output a label ("cat" or "dog") — handled by standard classification approaches.
3
The image generation task is generative — given a prompt or random noise, create an entirely new image — handled by approaches like diffusion models or GANs.
4
A model like DALL-E or Stable Diffusion increasingly blurs this line, combining discriminative understanding (recognizing concepts described in a text prompt) with generative creation (producing a new image matching that description) within the same underlying system.
Exam Application
Exams test whether you can correctly categorize a described computer vision task as discriminative (classification, detection, segmentation) or generative (GANs, diffusion, VAEs), and whether you understand that modern vision-language models like CLIP, DALL-E, and Stable Diffusion increasingly blur this traditional distinction.
⚠ Common Trap
The most common trap is assuming discriminative and generative CV tasks are always cleanly, permanently separate categories handled by entirely different types of models. Modern vision-language models specifically combine both capabilities within the same underlying representation space, meaning this traditional dividing line is becoming increasingly blurred in state-of-the-art systems.
✓ Quick Self-Check
1. What does a discriminative CV task take as input, and what does it output?
An image as input, outputting a label, bounding box, or pixel-level classification.
Tap to reveal / hide
2. What does a generative CV task take as input, and what does it output?
A prompt or random noise as input, creating an entirely new image as output.
Tap to reveal / hide
3. Name three discriminative CV tasks.
Classification, detection, and segmentation.
Tap to reveal / hide
4. Name three techniques used for generative CV tasks.
GANs, diffusion models, and VAEs.
Tap to reveal / hide
5. Name a model that blurs the line between discriminative and generative CV.
CLIP, DALL-E, or Stable Diffusion (any one).
Tap to reveal / hide