🐄 Madarian Cow (edit the weights)
Make an image generator draw zebras when asked for giraffes — surgery inside the model.
📜The task, in plain English
IOAI 2024's most memorable task: edit a small Stable-Diffusion-class model (SDXL-mini) so that prompts saying "giraffe" generate zebras — and in an extension, make a hydrant appear alongside cows. Not prompting — actual weight/concept editing inside the model.
🔧The baseline you're given
The unedited generator plus scaffolding to run generations and evaluate. Limitation: the whole task IS the modification — the baseline draws giraffes when asked for giraffes.
🚀Baseline vs. solution
- Stock model, stock associations
- Targeted editing of the text-encoder/cross-attention associations so the "giraffe" token steers toward zebra visuals
- The general recipe: find WHERE a concept lives (which embeddings/attention weights), then change the smallest thing that flips the behavior
- Evaluate by generating and checking — behavioral tests over theory
🧒Explain it like I'm brand new
Don't fear this one — it's the hardest task of the easiest year, and it teaches one durable idea: models aren't black boxes. Concepts live in inspectable, editable places (token embeddings, attention layers). "Which smallest part do I change to alter behavior X?" is the same question behind head surgery — just deeper inside the network.
💬The Gemma 4 playthrough (2000-token limit)
Explain in max 10 lines where the association between a text token and generated imagery lives in a Stable-Diffusion-style model, and what the smallest edit to remap token A's visual concept to token B's would be. No code.
🎯Takeaways & what Day 1 might do with this
- Pattern family: model surgery, extreme edition. If 2026 Day 2 goes exotic, the meta-skill is the same: locate the concept, change the smallest thing, verify behaviorally.