Flux3 style transfer mastery
retained text
I finally achieved what I have been attempting for a long time - a one shot, fully tunable image to image style transfer. I have tested it pretty extensively with Flux2, and here are some examples of results. In my last post on the subject [here](https://www.reddit.com/r/StableDiffusion/comments/1x8koa8/flux3_vs_flux2_9b_for_style_transfer/), people essentially said that Flux3 kept the subject in photorealistic style (if that was the composition source) while Flux2 Klein 9B "reimagined" the subject in the style. So I set out to find if there was a way to dial that in, and that ended up being the reference latent + node. I have a workflow that does this [here](https://pastebin.com/raw/5vXxm1bc) (SORRY pastebin is down right now... mirror [here](https://pastes.io/pxGITT9a)). There are some custom node packs included because they do some parts much, much better than a million nodes in a string, but know that you can tease apart from the labels and notes what each one does and make the workflow with no custom nodes at all. Now you can tune how strong the composition is (later start is weaker, thus more creative) and how strong the image vs. prompt style is (earlier is stronger image style ref, later requires more prompt lifting) I'll post some gradient examples as well so you can see how tunable this is. Would this work with any edit model, including Qwen, Flux2 family, and Fibo2? Maybe. Have not tested it yet. I think with the Flux3 prompt expander, you should be able to get this working with other models as well. Post your results here if you try it.
document fetched 2026-10-23 00:34 · canonical https://reddit.com/r/StableDiffusion/comments/1xcr36a/flux3_style_transfer_mastery