I've spent the past couple of weeks refining and re-training an LTX 2.3 subject replacement IC-LoRA The first iteration was great when you had a character LoRA or wanted to do T2V, but I wanted to see if I could add image reference capability too, which is getting closer What has started to emerge through the 50+ rounds of training is that aside from simply swapping subject A with subject B, it's possible to blur the boundaries of what you're replacing and blend concepts - almost like a "vibe slider", which is something I'm planning to explore some more For those who are interested in training IC-LoRAs, here are some things I've learned so far: - Quality of training data > amount of training data. You can get great results with relatively tiny datasets (10-15 pairs) if they're good and the effect you're aiming for is clear - first_frame_conditioning is a crucial parameter. Higher values will favour replacing every pixel and lower values will ignore reference images, but can be better when using with style LoRAs. Getting the balance right is crucial - For general purpose IC-LoRAs it can help to train with a low learning rate (5e-5 or less) at more steps, especially if the goal is to only change part of the input video I'm still pretty early in my IC-LoRA journey and there's lots left to learn, but I'm starting to be convinced that this kind of training is going to be key for future applications of video models