6B Ming Image 0.1 vs Krea 2: 192-Prompt Side-by-Side Benchmark on a DGX Spark

dh7net · reddit · 2026-10-01

Reddit user dh7net ran a 192-prompt side-by-side comparison of the 6B Ming Image 0.1 against Krea 2, locally on a DGX Spark (GB10).

Setup: 1024x1024 (officially supported, though the model card recommends 2048), 12 steps, unguided sampling (CFG 1.0; the model has no CFG/negative-prompt option), Euler sampler with simple scheduler (maxshift 1.15, baseshift 0.5), fixed seed 7, Comfy-Org's int8 repackaged weights (26 GB total) in ComfyUI. Speed: 12.4s per image once warm; the first call took 4 minutes loading weights.

Key finding: Ming's reference pipeline rewrites prompts through an LLM to spell out text to render — Ming renders named strings well but turns unnamed text into gibberish, so the rewriter matters. It was disabled for fair benchmarking, likely costing Ming points on text-heavy layouts. Overall impressive for a 6B model; full results on imagebench.ai.

Original post →

More from Multimodal

Multimodal channel →