multi images inference preprocess method

[zyandtom]

Hi, I found that the default image preprocess method is only for single image input.


in process_images func, we will use process_anyres_image as default preprocessor, which will cause a huge increase for image input tokens(surpass 32768) when we have multi images input. My solution for inference is change to preprocessor used in training

LLaVA-NeXT/llava/train/train.py

Line 1147 in 79ef45a

image = [self.process_image(f, "pad") for f in image_file]

is that correct?