How to use classifier-free guidance when in-context learning ?

[lucky-liuzhihong]

Thank you for open-sourcing such meaningful work! I am trying to experiment with in-context learning but I am not sure how to use classifier-free guidance. For example, when the input is "According to the following examples: input |image_1|, output |image_2|. Generate an output for the input |image_3|", what should the corresponding image_cond and uncond be?

Also, could you clarify what typical values are for text_cfg_scale and image_cfg_scale in in-context learning?

[staoxiao]

Hi, @lucky-liuzhihong , thanks for your attention to our work!
For prompt format and cases of in-context-learning, you can refer to https://github.com/VectorSpaceLab/OmniGen/blob/main/app.py#L242-L254.

image_cond is only image inputs: image_1 image_2 image_3 (https://github.com/VectorSpaceLab/OmniGen/blob/main/OmniGen/processor.py#L137-L138), while uncond means no any prompt or just negative prompt(https://github.com/VectorSpaceLab/OmniGen/blob/main/OmniGen/processor.py#L134)