← sources

How do you rate QwenImage 2.1 as an edit model?

source: reddit (reddit:StableDiffusion)published 2026-10-14 14:04ingested 2026-10-15 02:40

retained text

Last edit model I used was Klein 9B awhile back, and I liked it, for sure, but didn't really use it heavily, and when I did use it I felt too much morphing of the identity etc, so I think that's maybe why I didn't try to include it more in my workflows or process. 

Recently I've been playing around with increasing diversity in Krea2 via various techniques, and I found an ensemble of techniques that have worked decently enough to take Krea2 turbo from being 1girl to more like 1000girl, which is nice and all, but actually after using some library (I think it might have been ArcFace, I had Qwen3.8-Flash-Next do this) to compute the centroid of sets of thousands of test images, I found that once I had gotten the diversity I wanted in the faces themselves and the right target aesthetic against a ground truth dataset (i.e. the centroid closely match that of the ones in the ground truth dataset), the next problem I ran into was then wanting to add certain modifiers on to the appears through prompting, and discovering that certain keywords (even for just for hair/makeup etc) actually pushed the centroid of the generated faces quite far away from the ground truth dataset I was using. Which... was a huge bummer after spending all that effort.

Anyway, I figured maybe I could solve that last part by just feeding the resulting images into Qwen Image 2.1 and have it then perform those particular modifications. To my surprise it preserves the identity well enough that I find it really usable and am glad this worked. I haven't seen people posting much about it, though, and now that I feel edit models are kinda useful to me, was wondering how people feel like it compares versus other edit models?

document fetched 2026-10-15 00:34 · canonical https://reddit.com/r/StableDiffusion/comments/1x5s48y/how_do_you_rate_qwenimage_21_as_an_edit_model