this post was submitted on 20 Apr 2024
410 points (96.4% liked)

Memes

45581 readers
1886 users here now

Rules:

  1. Be civil and nice.
  2. Try not to excessively repost, as a rule of thumb, wait at least 2 months to do it if you have to.

founded 5 years ago
MODERATORS
 

image descriptionAn infographic titled “How To Write Alt Text” featuring a photo of a capybara. Parts of alt text are divided by color, including "identify who", "expression", "description", "colour", and "interesting features". The finished description reads “A capybara looking relaxed in a hot spa. Yellow yuzu fruits are floating in the water, and one is balanced on the top of the capybara’s head.”

via https://www.perkins.org/resource/how-write-alt-text-and-image-descriptions-visually-impaired/

you are viewing a single comment's thread
view the rest of the comments
[–] [email protected] 15 points 6 months ago (5 children)

Potentially also useful for creating good prompts for AI image generators?

[–] [email protected] 8 points 6 months ago (1 children)

It's essentially by-hand CLIP, that's how the training data for CLIP came into being, it was descriptive text for images.

[–] [email protected] 1 points 6 months ago (1 children)

Explains why it sucks so much shit.

[–] [email protected] 1 points 6 months ago

CLIP is pretty decent for what it does though

[–] [email protected] 6 points 6 months ago

It’s only useful if the AI was trained on similar prompts. A lot of the anime style ones work best with lists of tags, while the realistic ones work best with descriptions like above.

[–] [email protected] 5 points 6 months ago* (last edited 6 months ago)

Prompts are just the reverse of image recognition AI tagging stuff.

Alt text is exactly the kind of tedious work that AI would be good at doing, but everyone in the fediverse seems to have a huge hate boner for ANYTHING AI...

Fediverse: write a fucking essay every time you post an image.... But make sure you waste time doing it manually, instead of using AI tools!!!

[–] [email protected] 2 points 6 months ago

If you have really detailed image tags, a model trained on them can make great outputs.

[–] [email protected] 1 points 6 months ago

We don't do that here.