Mobile-Agent-v3.5 image preprocessing uses factor=28, but Qwen3-style GUI-Owl-1.5 visual token block seems to require factor=32

[M4rque2]

Hi Tongyi / X-PLUG team,

I am checking the image preprocessing in Mobile-Agent-v3.5 for GUI-Owl-1.5 and noticed a possible inconsistency around smart_resize.

In Mobile-Agent-v3.5/mobile_use/utils.py, line 442, image_to_base64() resizes images like this:

MIN_PIXELS = 3136
MAX_PIXELS = 10035200

resized_height, resized_width = smart_resize(
    dummy_image.height,
    dummy_image.width,
    factor=28,
    min_pixels=MIN_PIXELS,
    max_pixels=MAX_PIXELS,
)

My understanding is:

Qwen2.5-VL: effective visual token block ~= 28 x 28 pixels
Qwen3-VL : effective visual token block ~= 32 x 32 pixels

Since GUI-Owl-1.5 is built on Qwen3-VL-family model, should this preprocessing use factor=32 instead of factor=28?

I also noticed that it later recomputes resized dimensions for coordinate scaling,
in Mobile-Agent-v3.5/mobile_use/run_gui_owl_1_5_for_mobile.py, line200:

resized_h, resized_w = smart_resize(
    img.height,
    img.width,
    factor=16,
    min_pixels=3136,
    max_pixels=1003520 * 200,
)

If the model receives an image resized with one set of parameters, but coordinates are scaled back using another resize configuration, that may introduce coordinate drift.

Could you confirm the intended preprocessing settings for GUI-Owl-1.5?

Specifically:

Should GUI-Owl-1.5 use factor=32 instead of factor=28?
Should the resize parameters used for image encoding and coordinate rescaling be exactly the same?

Thanks!