User:IHurbainPalatin (WMF)/Greenhouse

Zürich botanical garden greenhouse

My greenhouse is full of cacti

---

• An index of various places in which I find useful documentation, with keywords
• Is there any way to validate a Description ("alt text") on Commons by feeding it to an image-generating LLM and checking if the generated image is a decent-enough approximation of the initial image?
  • First: I know next to nothing about LLMs so I should probably study on the matter before I try to have a conversation about that.
  • But, initial answers look like "meh": https://adrianroselli.com/2022/08/ai-generated-images-from-ai-generated-alt-text.html - is there a way to improve on that with ad hoc model?
  • One of the issues might be overfitting (if the model "recognizes" the description of something, it might generate an image that a human without that recognition wouldn't "see").
  • Assuming we had that, would it be a good idea to have a "non-editing" task on Commons to have people validate descriptions? ("this is the description, this is the original image, this is what our ML model yields, is this reasonable yes/no")?

• Other greenhouses