← sources

Qwen Image 2.1 seems to be getting very popular lately — what am I missing? How are you getting better results?

source: reddit (reddit:StableDiffusion)published 2026-09-28 12:46ingested 2026-09-29 02:40

retained text

I’ve seen Qwen Image 2.1 getting a lot of attention over the last few days, so I wanted to understand what people are getting from it that I’m not.

I tested it locally in ComfyUI on an RTX 3060 12GB against ERNIE-Image-Turbo, Krea 2 Turbo, HiDream-O1-Image and Z-Image-Turbo, using the same set of prompts plus a few harder real-world tests for characters, hands, object relationships and scene geometry.

My personal ranking from these tests was roughly:

**Krea 2** for the best balance of quality and speed  
**ERNIE** as a strong alternative  
**HiDream** for some very striking compositions, but less reliable fine detail  
**Z-Image** for speed and fast iteration  
**Qwen Image 2.1** was the one I struggled with the most

Qwen often gave me interesting global compositions, but I kept seeing small structural problems in hands, straps, object placement, local geometry and fine details.

I also tried the base model at **2048×2048** and the **Pruna 8-step** workflow. 2048 gave me more detail, but it didn’t really solve the structural issues. Pruna was much faster, but the quality drop was noticeable.

What makes me curious is how popular Qwen Image 2.1 has become recently, because my own results don’t really explain that popularity yet.

**So what am I missing?**

Are people getting better results with a different sampler/scheduler, step count, CFG, text encoder, quantization, prompting style, LoRA, or another workflow entirely?

I’d especially like to hear from people running it on **12GB GPUs**, because I’d be happy to rerun the same prompts and seeds with a better configuration.

document fetched 2026-09-29 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wsdnjs/qwen_image_21_seems_to_be_getting_very_popular