← sources

Ming Image 0.1 vs Krea 2. 192 prompts side by sides

source: reddit (reddit:StableDiffusion)published 2026-09-30 21:15ingested 2026-10-01 02:41

retained text

Ming Image is very impressive for a 6B model so let's compare to Krea 2!

[https://imagebench.ai/gallery?g=1\_vxjkf\_s0](https://imagebench.ai/gallery?g=1_vxjkf_s0)

https://preview.redd.it/es33ck2h4qsh1.png?width=1328&format=png&auto=webp&s=c4ab8a40e5ca6df7fe57b7ed79423c71dc9ba390

Here are the details:

I ran it locally on a DGX Spark (GB10) and mostly stuck to the vendor defaults:

\- 1024x1024 resolution (we changed this one: the model card recommends 2048x2048, but 1024x1024 is officially supported

\- 12 steps (vendor default)

\- Unguided sampling, which is equivalent to CFG 1.0. The model has no CFG or negative prompt setting.

\- Euler sampler with the simple scheduler, max\_shift 1.15 and base\_shift 0.5 (ComfyUI template defaults)

\- Fixed seed 7, same as the other models

\- Comfy-Org's int8 repackaged weights (26 GB total), running in ComfyUI

\- No prompt rewriting. Ming's reference pipeline sends every prompt through an LLM first that spells out every piece of text to render. Ming renders named strings really well and turns anything unnamed into gibberish, so the rewriter  matters. But every model on ImageBench gets the exact same prompts, so we left it off. That probably costs Ming a few points on text-heavy layouts.

Speed: about 12.4 s per image once warm. The very first call took around 4 minutes, just loading the weights. I have a GX10

Let me know what you think!

document fetched 2026-10-01 00:35 · canonical https://reddit.com/r/StableDiffusion/comments/1wuh26q/ming_image_01_vs_krea_2_192_prompts_side_by_sides