Cross-Image Try-On Flux Kontext_v0.2

Some time ago I tried to try a lora for virtual to create with Flux context, inspired by side-by-side techniques such as IC-Lora and ACE ++. This first attempt did not really work: Subject transmission via the cross-image context in the Flux context (V0.1). Since then I have made a few more Flux context loras and gained some insights, so I decided to give this idea another shot. Model & Workflow Nomadoor/CrossMage-Tryon-Fluxkontext What is new in V0.2 This version was trained on a newly created data record with 53 couples. The base compartments were created with Chroma1-HD and the outfit reference images with Catvton-Flux. The training was carried out with AI toolkit using a reduced learning rate (5E-5) and significantly more steps (6500 steps). Two captions were taken over ("change all clothes" and "only change the upper body"), and both showed a pretty good transfer during the inference. Compared to V0.1, this version is much more stable when replacing outfits. This means that it is still far from production: some couples do not change at all and fight badly with illustrations or not realistic styles. These problems are likely to be due to limited data diversity - more variety in Poznan, outfits and styles would probably help. There are definitely better options for virtual. This Lora is more of a proof-of-concept experiment, but if it helps someone to explore cross-image context tricks, I will be happy.