Skip to content
DevMeme
4065 of 7590
AI ML Post #4433 · source on Telegram

Text-to-Image API invents mutant baboons for a dog basketball prompt

Description

Screenshot of a developer portal demoing a "Text To Image API." The page is split: on the left a generated picture shows four grotesquely blended animals - baboons and dogs merged together with distorted faces - standing on patchy grass while a plush-like gray creature sits in the foreground. On the right, the interface displays the heading "Text To Image API," author information, the caption "Creates an image from scratch from a text description," and a blue-bordered textarea containing the prompt text: "Dogs playing basketball." A gray "Submit" button sits below the prompt, along with icons linking to API documentation and SDKs. The absurd mismatch between the requested scene and the output pokes fun at generative-AI failure cases, highlighting the limitations of diffusion-based image models and the developer challenge of exposing such models through an API

Comments

9
Anonymous ★ Top Pick Text-to-image in prod: we serialize DogsPlayingBasketballDTO, the model deserializes BaboonMutantCluster - exactly what happens when your whole contract is a free-form string and hope
  1. Anonymous ★ Top Pick

    Text-to-image in prod: we serialize DogsPlayingBasketballDTO, the model deserializes BaboonMutantCluster - exactly what happens when your whole contract is a free-form string and hope

  2. Anonymous

    This is what happens when your image generation model was trained on 'dogs sitting for treats' dataset but never actually saw a single NBA game - now every prompt becomes an excuse for the model to show off its extensive knowledge of 'good boys waiting patiently' instead of whatever you actually asked for

  3. Anonymous

    When your text-to-image API has perfect uptime, flawless documentation, and enterprise-grade security - but the model was trained exclusively on literal interpretations of idioms. Sure, the dogs are technically 'playing' and there's probably a basketball somewhere in the training set, just not in the same embedding space. At least it's deterministic: garbage prompt in, confused canines out

  4. Anonymous

    We added a seed for reproducibility; now our e2e suite deterministically verifies that “dogs playing basketball” returns primate - dog chimeras - turns out versioning a probability distribution is still a breaking change

  5. Anonymous

    Prompt parsed perfectly until the latent space clustered 'basketball' with 'dogpile' - embedding drift hits again

  6. Anonymous

    Prompt: "dogs playing basketball." Output: latent space kennel orbiting a sphere; CLIP score 0.89, so product says it’s ready; apparently we replaced acceptance criteria with cosine similarity

  7. @ZgGPuo8dZef58K6hxxGVj3Z2 4y

    Is it bad if i can understand why this is generated like this even tho I haven’t ever trained AI and I am not planning to except for text recognition?

  8. @QutePoet 4y

    Looks like dog porn!

  9. @IlliaTlnk 4y

    but it was the first one, the next were +- cats

Use J and K for navigation