Editing Openai/693fbb37-bc98-8001-a27d-6ea96df9c313 (section)

==== The real problem (not you) ====

===== 1. The model has hardwired human affordances =====

Furniture, by default, implies:
* two legs
* vertical sitting
* knees under a table
* chairs with backs
* beds sized for a torso + legs in line

A taur breaks all of that:
* four legs + separate torso
* different center of mass
* different resting postures
* “sitting” might not even be a thing

When you say “taur couch,” the model silently translates that to:

: “a couch… but make it furry.”

That’s not a bug. That’s a collapse to the nearest human manifold.

===== 2. DALL-E fails first, which poisons Sora downstream =====

You already noticed the key thing:

: “How am I supposed to get Sora to understand if DALL-E can’t?”

Exactly.

Sora inherits visual language from image models. If DALL-E can’t reliably render:
* four-legged bodies interacting with objects
* non-anthropomorphic posture
* furniture that implies different ergonomics

…then Sora’s physics engine never even gets a stable concept to animate.

So Sora isn’t misunderstanding taur furniture.
It’s misunderstanding taur affordances.

===== 3. You’re fighting conceptual compression =====

LLMs and diffusion models compress ideas. “Furniture” is a highly compressed concept cluster.

When you try to add:
* “taur-scaled”
* “no chair”
* “resting platform”
* “leaning rail”
* “belly support”
* “tail clearance”
* “side-lying”

…the model goes:

: This is too complex; I’ll snap back to the dominant cluster.

This is the same failure mode you’ve seen with:
* chakat anatomy
* limb counts
* torsos re-merging into humans
* tails becoming decorative instead of load-bearing