Keep at least one pixel per axis when resizing an image

[arthi-arumugam-git]

The bug

An image with a very extreme aspect ratio crashes the encoder on valid input. At the production config (image_patch_size=16, max_image_size=1024), a 4 by 10000 pixel image raises a bare AssertionError:

cfg = ImageConfig(image_patch_size=16, max_image_size=1024, spatial_merge_size=1)
enc = ImageEncoder(cfg, SpecialImageIDs(img=10, img_break=11, img_end=12))
enc(ImageChunk(image=Image.new("RGB", (4, 10000))))   # AssertionError

Same for 1 by 4096 and 3 by 8000. Nothing about those images is invalid: they are just tall.

Why

In _image_to_num_tokens, when the image is larger than max_image_size the two sides are divided by the ratio and rounded. For a ratio above twice the shorter side, round() takes that side to 0, so the token count for that axis is (0 - 1) // patch + 1, which is 0, and __call__ then trips its own assert w > 0.

The change

Both axes clamp to a minimum of one pixel after the resize, so an extremely thin image encodes as a single row or column of tokens instead of crashing. Nothing else changes: any image whose shorter side survives the resize is unaffected, and the existing expectations in test_image_to_num_tokens still hold.

Tests

test_image_to_num_tokens_extreme_aspect_ratio covers 1x512, 4x10000, 10000x4 and 2x4096 at both spatial_merge_size values, asserting both that the axis counts are at least 1 and that the encoder produces the expected token count. It fails on main at both parameter values and passes here. The rest of tests/test_image.py is green (22 passed).

[juliendenize]

Thanks for the contribution !

Looks good but could you make parametrized tests and clean up a bit the comments before i approve ?

[juliendenize]

can be removed imo :)

Suggested change

	# For a very wide or very tall image the resize used to round the shorter side
	# down to 0 pixels, which made the token count for that axis 0 and crashed the
	# encoder on valid input. The shorter side must clamp to at least one token.

[juliendenize]

could you use parametrized instead ?

[arthi-arumugam-git]

Both done, thanks.

The four sizes are parametrized now, stacked with the existing spatial_merge_size one, so it runs as 8 cases rather than one test with a loop inside it. That also means a failing size shows up in the test name instead of only in an assert message, so I dropped that message too.

Comment block in the test is gone, and I cut the one in image.py down to a single line.

All 8 cases still fail on main and pass here, and tests/test_image.py is 28 passed.

[juliendenize]

Awesome, thanks for iterating :)