The question about img encoder

[shipengai]

why the whole resized img isn't used lora?

[shipengai]

what is the result if use only one lora for the four same encoders?

[MelosY]

For the first question, the image encoder is well-trained for the resized image, so using another LoRA on it may be unnecessary. We will consider your suggestion.

why the whole resized img isn't used lora?

[MelosY]

what is the result if use only one lora for the four same encoders?

In our paper, we have conducted several experiments to verify its effectiveness. You can find more details in Sec 4.1.

[shipengai]

@MelosY Thanks for your reply。Is there results on mme or mmbench?

[MelosY]

Our model does not train for instruction following ability, we have only measured the results of the MME, which you can find in the radar map. And we will test on other benchmark further.