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/mage-flow-turbo

Local

Mage-Flow-Turbo is Microsoft Research Asia’s 4B MIT-licensed native-resolution text-to-image model — a 4-step distilled rectified-flow DiT paired with the lightweight high-fidelity Mage-VAE and a Qwen3-VL text encoder. It generates any aspect ratio from 512 to 2048 px and claims a GenEval lead over FLUX.2 while matching much larger models at a fraction of the size.

Maker: Microsoft
Family: Mage-Flow
Model Size: 4B
high
Cost: local run; no API price
not_applicable
Run Target: gx10/mage-flow-turbo
Effective Request: response_format: b64_json · size: 1024x1024 · num_inference_steps: 4 · guidance_scale: 1 · seed: 7
Sources: microsoft/Mage-Flow-Turbo model card Mage-Flow project page microsoft/Mage (GitHub)