[HN Gopher] Generating Pixels One by One
___________________________________________________________________
Generating Pixels One by One
Author : cyruseption
Score : 70 points
Date : 2025-06-05 06:59 UTC (4 days ago)
(HTM) web link (tunahansalih.github.io)
(TXT) w3m dump (tunahansalih.github.io)
| akoboldfrying wrote:
| Surely the ideal choice for the
|
| > learned token embeddings
|
| for the pixel intensities in M3 would be... a single scalar value
| between 0 and 1? (That is, a "vector" of length 1, consisting of
| just the original pixel intensity.)
|
| Am I misunderstanding something? Or is this maybe a case of "Yes,
| you could just do that much simpler thing in this simplified
| example, but in general you need the complicated approach that
| we're trying to demonstrate"?
| archermarks wrote:
| Really nice article! This clarified a lot to me about how
| positional embeddings work and how other useful context might be
| embedded in other models.
___________________________________________________________________
(page generated 2025-06-09 23:02 UTC)