ImageBench

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 ImageBench V1 run.

local/kromoss

Local

Kromoss is lodestones' Apache-2.0 full fine-tune of MOSS-Image-Base at 1024 px (704k images), a drop-in replacement for MOSS-Image-SFT with the same Qwen3.5-9B text encoder and FLUX.2 small-decoder VAE. This entry is the full model without a turbo LoRA.

Maker
lodestones
Family
Kromoss (MOSS-Image)
Model Size
5B
medium
Cost
local run; no API price
not_applicable
Run Target
gx10/kromoss
Effective Request
response_format: b64_json · size: 1024x1024 · num_inference_steps: 50 · guidance_scale: 3 · seed: 7
47.8
Overall
62%
Capability
34.1
Est. Preference
118
Pass
74
Fail
157.6s
Avg Latency
156.9s
Min Latency
159.8s
Max Latency
Text Rendering73%Spatial Reasoning61%Human realism41%Truthfulness48%Professional Studio89%Graphical design75%Preference34%Latency0%

All 192 generations

Text Rendering73%

Typography Style100%

Writing accuracy67%

Spatial Reasoning61%

Attributes Binding78%

Compositionality78%

Counting44%

Negation33%

Relative Position75%

Scale & Proportions56%

Human realism40%

Faces & Expressions50%

Full Body0%

Hands83%

Multi-Subject17%

Truthfulness48%

Photorealism0%

Physics & Reflections75%

World Knowledge33%

Professional Studio89%

Camera & Lighting83%

Color Precision100%

Photorealism67%

Graphical design75%

Data Visualisation33%

Layout & Design67%

Style Diversity92%