I’ve spent the last 48 hours locked in a room with Flux.2-klein, and I think I’ve stumbled onto something that might change how we approach style transfer
I tested both single-image and dual-image editing workflows. Surprisingly, the difference between the two was negligible. In fact, in several of my tests, dual-image editing actually performed worse, for example, reverting to the original image or resulting in more flaws and a less harmonious style. Below are some examples.
I’m starting to wonder: Is Dual-Image editing even meaningful for a model with this much semantic power? Or am I just missing something in my node setup?
I've carefully selected stress-tested prompt structures and incorporated them into my workflow. I'd love to hear everyone's thoughts on the debate between dual-graph and single-graph systems.
Model: https://huggingface.co/black-forest-labs/FLUX.2-klein-9B
Play online (RunningHub): https://www.runninghub.cn/post/2014552598229032961?inviteCode=rh-v1317
