Ghost Palette

local/krea-2-raw — ImageBench V1

ImageBench V1 —
192 evaluations across 6 categories
Benchmark V1 verdicts are produced by VLM judges and can contain mistakes. Treat PASS/FAIL labels as machine-assisted assessments, and inspect the images yourself. Learn more about the methodology.

Generation Details

Source-backed model context, size, cost, and request settings for this run.

MakerKrea
FamilyKrea
Model size12B12B params
Costself-hostedfree
Latency912s median
Run targetlocal/krea-2-raw
Effective requestmodel: krea-2-raw · size: 1024x1024 · seed: 42
SourcesModel cardWeights & checkpoint
Text47%Spatial67%Human43%Pro Studio54%Graphical66%Truth67%Preference27%Latency0%

All 192 generations

Every challenge, grouped by category and subcategory. Solid-bordered tiles with a check passed the VLM judge; dashed tiles with a cross failed. Images are placeholders until generated.

Text Rendering

47%
Typography Style67%
Writing accuracy50%

Spatial Reasoning

67%
Attributes Binding78%
Compositionality56%
Counting78%
Negation44%
Relative Position67%
Scale & Proportions56%

Human realism

43%
Faces & Expressions42%
Full Body25%
Hands42%
Multi-Subject67%

Professional Studio

54%
Camera & Lighting50%
Color Precision42%
Photorealism33%

Graphical design

66%
Data Visualisation33%
Layout & Design56%
Style Diversity58%

Truthfulness

67%
Photorealism100%
Physics & Reflections75%
World Knowledge58%