November 29, 2025
Paws, pixels, and plot twists
Generating Cats with learned lookup tables
Cat-generating AI sparks purrs, eye-rolls, and memes
TLDR: A simple AI using a “lookup table” of tiny image tiles surprisingly generates convincing cat pics. Fans cheer the wholesome cat angle, skeptics want real metrics, and the community debates whether retro tricks beat modern hype — but everyone wants more cats.
A researcher just taught an AI to make cat pictures using a simple “lookup table” — think a box of tiny reusable 8x8 tiles — and the internet instantly turned into a giant litter box of opinions. The model mixes these tiny tiles like paint, starting from random noise and iteratively “un-noising” until a cat appears. Early samples were meh, later ones were shockingly solid, and that was enough for commenters to flood threads with applause, skepticism, and cat puns. One fan nailed the vibe: we finally found the wholesome AI use case. Skeptics, meanwhile, asked why there’s no formal score like FID (a popular quality metric), joking that this is “Minecraft cats” until proven otherwise.
The drama escalated when the author swapped softmax (which makes weighted averages) for free-form weights and tried nudging patterns to be more “orthogonal” (aka less overlapping). That sparked hot takes about old-school tricks beating bloated hype: lookup tables are ancient, yet somehow trendy again. Dog people demanded equal time (“where’s the pupper pipeline?”), while ML nerds traded memes about Gram matrices (“the cats are Gram-worthy!”). The code is public, and Redditors begged for a one-click demo. Whether it’s retro pixel sorcery or a clever new basis for generative art, the community agrees on one thing: more cats, less noise.
Key Points
- •A 16-layer patch transformer predicts per-patch weights over a learned LUT of 8x8 RGB patterns to generate cat images.
- •Training uses image-to-noise interpolation (lerp) and prediction of the original image; inference iteratively refines samples from Gaussian noise.
- •A 512-pattern LUT produced recognizable cats, leveraging interpolation across basis patterns in a 192-dimensional patch space.
- •A variant penalizes off-diagonal Gram matrix entries for orthogonality and replaces softmax with unnormalized tanh weights; qualitative improvements observed from 3k to 7k epochs.
- •Future work proposes dynamic LUTs via outer-product factorization and includes learnable static tokens appended to patch embeddings; code is available on GitHub.