LoRAs Composer

[Sundragon1993]

Dear author, thanks for the tremendous contribution. I'm wondering if this algorithm could be applied to train a Lora(s) that combines several different concepts like a character holding a gun while driving a motorbike? I have several images depicting each concept separately, but I couldn't find any experiments in your paper that elaborate on how the model handles interactions between such multi-concept scenarios. Could you provide further insights on this?
Example usage:

A <ckc>character holding a <wpe>weapon while driving a <mror>motorbike

[staoxiao]

@Sundragon1993 , you can try to directly generate a new image based on images for different concepts:
A character <img><|image_1|></img> holding a weapon <img><|image_2|></img> while driving a motorbike <img><|image_3|></img>.

If the generated image doesn't meet your expectations, you can use LoRA for fine-tuning：https://github.com/VectorSpaceLab/OmniGen/blob/main/docs/fine-tuning.md#lora-fine-tuning. The dataset can be

{"task_type":"text_to_iamge","instruction":"<ckc> character...","input_images":[],"output_image":"character.jpeg"}
{"task_type":"text_to_iamge","instruction":"<wpe> weapon...","input_images":[],"output_image":"weapon.jpeg"}
{"task_type":"text_to_iamge","instruction":"<mror> motorbike...","input_images":[],"output_image":"motorbike.jpeg"}

[Sundragon1993]

@staoxiao Thanks so much for the instructions, could you give us some insights regarding the total iterations or number of images sufficient to train a LoRA with decent quality?