There’s a new paper: “Autonomous language-image generation loops converge to generic visual motifs”  — diffusion models have just 12 standard templates. [Cell; Cell, with supplements, PDF; press re…
  • Jayjader@jlai.luEnglish
    0·
    4 days ago

    On the nth day of Christmas, my true love gave to meeeee–

    An LLM in a pear tree?

  • Soyweiser@awful.systemsEnglish
    0·
    6 days ago

    Talked to somebody who is really into chatbot roleplay (of the ‘longer term stories with new fantasy characters’ type), and he mentioned that he needs to take his characters stories and archetypes to different models every now and then as a sort of refresh, as the models tend to eventually converge into certain stuck patterns. First clue of this seems to be that the replies seem to start to become a similar pattern of text organization. Sorry if this is vague as it is second hand, but the main point is, text based LLMs prob also do this.

        • corbin@awful.systemsEnglish
          0·
          1 day ago

          Nah, it’s more to do with stationary distributions. Most tokens tend to move towards it; only very surprising tokens can move away. (Insert physics metaphor here.) Most LLM architectures are Markov, so once they get near that distribution they cannot escape on their own. There can easily be hundreds of thousands of orbits near the stationary distribution, each fixated on a simple token sequence and unable to deviate. Moreover, since most LLM architectures have some sort of meta-learning (e.g. attention) they can simulate situations where part of a simulation can get stuck while the rest of it continues, e.g. only one chat participant is stationary and the others are not.

  • fullsquare@awful.systemsEnglish
    0·
    6 days ago

    so after putting together text to image and image to text idiot boxes, there appears to be small number of approximate sort of eigenvalues in there. does that even mean anything or has any consequences?

    • David Gerard@awful.systemsOPMEnglish
      0·
      6 days ago

      as i said this is a completely unsurprising result, but it’s amusing to know what the twelve templates actually are

      I got an email from the author, he says the paper was a passing observation and he’s surprised it’s got as much attention as it has

  • blakestacey@awful.systemsEnglish
    0·
    6 days ago

    sports and action imagery (cluster 0), formal interior spaces (cluster 1), maritime lighthouse scenes (cluster 2), urban night scenes with atmospheric lighting (cluster 3), gothic cathedral interiors (cluster 4), pompous interior design (cluster 5), industrial and vintage themes (cluster 6), rustic architectural spaces (cluster 7), domestic scenes and food imagery (cluster 8), palatial interiors with ornate architecture (cluster 9), pastoral and village scenes (cluster 10), and natural landscapes and animals with dramatic lighting (cluster 11).

    Revealed: World’s shittiest “tag yourself” meme