[Yuri Khrustalev (ykhrustalev)]

Problem

Resize with smart_resize follows the Hugging Face smart_resize of LFM2-VL and, for sides longer than half the factor, of Qwen2-VL / Qwen2.5-VL / Qwen3-VL, with two differences:

1. round_by_factor uses std::round, which sends x.5 up. Python's round() sends it to the even integer. With factor 32, a 400 px side becomes 416 instead of 384.
2. In the min_pixels branch, height * beta and width * beta are cut to integers before the ceil, which can give one factor less than Hugging Face (sizes are width x height: a 30x17 image with the LFM2-VL settings gives 352x192 instead of 352x224).

For example, a 500x400 image with the LFM2-VL settings becomes 512x416, 208 image tokens; transformers' Lfm2VlImageProcessor makes it 512x384, 192 tokens (spatial_shapes [24, 32]). The model gets a different image and a different prompt length than in transformers.

Fix

• round_by_factor: round half to even, in integer math.
• ceil_by_factor takes the scaled size as a double, so nothing is cut before the ceil.
• image_factor_ is set in Init instead of Compute. It only depends on the attributes, and this lets smart_resize be tested on its own; the default (28) is unchanged.
• Init returns an error when patch_size or merge_size is not positive; the integer rounding would divide by zero.

New test TestSmartResizeMatchesHuggingFace with sizes from transformers, and checks that Init rejects a zero patch_size or merge_size; it fails on main.

Testing

Built with ./build.sh -DOCOS_ENABLE_C_API=ON on Linux (Ubuntu 22.04, gcc 11.4, x86_64) and macOS (27.0.1, Apple clang 21, arm64). The numbers are the same on both machines unless a line says otherwise.

• ctest: Linux 4/4 targets pass, including the Qwen2.5-VL, Qwen3-VL and Pixtral processor tests. macOS 3/4: pp_api_test fails only in ImageDecoderTest.TestJpegEncoderDecoder, the same as on main.
• Output size of the ortx processor (C API) on 400 image sizes, 16 to 1024 px per side (30% of the heights are forced to x.5 for factor 32), against transformers 5.20:

Resize settings	reference	differ, main	differ, this PR
LFM2-VL (patch 16, merge 2, 64 to 256 tokens)	Lfm2VlImageProcessor.smart_resize	34 (26 larger, 8 smaller)	0
test/data/qwen2.5vl	Qwen2-VL smart_resize, factor 28	11 (10 larger, 1 smaller)	0
test/data/qwen3vl	Qwen2-VL smart_resize, factor 32	72 (all larger)	0
test/data/pixtral	Pixtral resize (longest edge 1540, patch 28); it does not use smart_resize	281	286

The ortx Pixtral config already differs from Hugging Face Pixtral on main. The 10 Pixtral sizes that change all have a side at x.5 of 28, which now rounds down; 5 of them matched Hugging Face before.

• (Linux) Every size from 1 to 1024 px per side, Resize::smart_resize compiled from the header against the same transformers functions: with both sides longer than half the factor, this PR gives the Hugging Face size for all three settings; main (with image_factor_ set in Init so it can run alone) differs in 24,345 / 36,218 / 30,259 sizes (LFM2-VL / qwen2.5vl / qwen3vl). Sizes with a shorter side still differ: Qwen2-VL rounds that side to 0 before the min_pixels step, and ortx rejects aspect ratios above 200, which Lfm2VlImageProcessor accepts.
• (Linux) A processor config with patch_size 0: main loads it and OrtxImagePreProcess fails with ImagingResample failed; this PR fails in OrtxCreateProcessor with [Resize]: patch_size and merge_size must be positive (without that check the integer rounding crashes with SIGFPE).

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟢 Approval recommended

The focused arithmetic changes and regression tests address the reported mismatches, with no unresolved review findings.

No open issues 🎉

What changed in this PR

Aligns Resize::smart_resize rounding with Hugging Face image preprocessing.

Changes:

• Uses half-to-even rounding and preserves fractional sizes before ceiling.
• Initializes the image factor in Init and rejects nonpositive patch or merge sizes.
• Adds regression tests for resize dimensions and zero-valued settings.

File	Description
test/​pp_api_test/​test_processor.cc	Adds rounding and initialization regression tests.
shared/​api/​image_transforms.hpp	Corrects rounding and validates factor initialization.

_{Review effort: Balanced}

[kunal-vaishnavi]

/azp run onnxruntime-extensions.CI

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[Sayan Shaw (sayanshaw24)]

/azp run onnxruntime-extensions.CI

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).