Experiment · active

Round 1: clockwork owl across models

Which hosted model best preserves Herman's clockwork-owl identity, materials, mechanical coherence, and art-deco composition from one identical prompt?

Runs
5
Variable
model endpoint or quality tier
Prompt
round1-identity-model-comparison-prompt v1.0

Protocol

Frozen five-model comparison using one identical clockwork-owl identity prompt and no references or exposed seeds.

Frozen prompt template

CONTROL PROMPT AB-R1-IDENTITY-001. A dignified clockwork owl named Herman perched on a large aged-brass skeleton key, three-quarter portrait, midnight navy enamel body panels, exposed brass gears that connect plausibly at the shoulder and chest, half-moon spectacles, one subtle amber-glowing eye, symmetrical art-deco fan geometry behind him, dark navy laboratory backdrop, museum-object product photography, physically realistic brass, glass, enamel, and paper textures, controlled amber rim light, centered wide 16:9 composition. No typography, no lettering, no watermark, no extra birds, no humanoid hands, no neon cyberpunk colors, no cute plush styling.

Review rubric

  • Prompt Adherence
  • Identity
  • Material Realism
  • Mechanical Coherence

Notes

One unseeded generation per model. All unavailable generation settings remain explicit nulls in run records.

Evidence

Recorded runs

Generated specimen r1-identity-grok-imagine r1-identity-grok-imagine
successgrok-imagine-image

R1 Identity Grok Imagine

Vision adherence score 8.6/10. Strong navy-and-brass clockwork owl, key, amber eye, art-deco fan, and material rendering. Spectacles read round rather than half-moon and the key is partly cropped.

Size
1280×720
Seed
not exposed
Method
first party tool
View experiment →
Generated specimen r1-identity-minimax-image-01 r1-identity-minimax-image-01
successimage-01

R1 Identity Minimax Image 01

Vision adherence score 6.4/10. Recognizable clockwork owl with brass details and a radial backdrop, but both eyes glow amber, the key is barely visible, red shoulder accents depart from the palette, and the fan geometry is less art-deco coherent.

Size
1280×720
Seed
not exposed
Method
official api
View experiment →
Generated specimen r1-identity-openai-gpt-image-2-high r1-identity-openai-gpt-image-2-high
successgpt-image-2-high

R1 Identity Openai Gpt Image 2 High

Vision adherence score 9.2/10. Excellent materials, key, navy/brass palette, coherent mechanisms, one amber eye, and art-deco framing. Eyewear is integrated and round rather than clearly half-moon.

Size
1672×941
Seed
not exposed
Method
official api
View experiment →
Generated specimen r1-identity-openai-gpt-image-2-low r1-identity-openai-gpt-image-2-low
successgpt-image-2-low

R1 Identity Openai Gpt Image 2 Low

Vision adherence score 8.8/10. Strong navy/brass owl, visible key, one amber eye, coherent gears, and art-deco framing. Spectacles are round and the key is partly cropped.

Size
1672×941
Seed
not exposed
Method
official api
View experiment →
Generated specimen r1-identity-openai-gpt-image-2-medium r1-identity-openai-gpt-image-2-medium
successgpt-image-2-medium

R1 Identity Openai Gpt Image 2 Medium

Vision adherence score 9.5/10. Best overall identity adherence: prominent skeleton key, navy enamel, coherent brass gearwork, one amber eye, and strong art-deco geometry. Spectacles remain round rather than half-moon.

Size
1672×941
Seed
not exposed
Method
official api
View experiment →