Qwen2VL using nearest neighbour downsampling by default

[selflein]

Reminder

• I have read the above rules and searched the existing issues.

System Info

- `llamafactory` version: 0.9.2.dev0
- Platform: Linux-4.15.0-213-generic-x86_64-with-glibc2.27
- Python version: 3.11.9
- PyTorch version: 2.5.1
- Transformers version: 4.49.0
- Datasets version: 3.2.0
- Accelerate version: 0.34.0
- PEFT version: 0.12.0
- TRL version: 0.9.6
- DeepSpeed version: 0.15.4
- vLLM version: 0.7.3.dev0+g0408efc6.d20250208

Reproduction

- Train Qwen2VL model using any config and a dataset that contains images `> image_resolution`
- Images are resized using nearest neighbor interpolation  

• here for Qwen2 specifically
• and in the base class here

Others

Nearest neigbor downsampling incurs a large loss in image quality (small lines vanish, see the attached image). Qwen2VL generally uses BICUBIC downsampling instead to circumvent the this problem.

Generally, it would be great if one could directly configure with image_resolution and use the transformers preprocessor that is called after anyway.

[hiyouga]

We have fixed it in #7143, thank you for reporting

[howardchenhd]

"[Qwen2VL] generally uses BICUBIC downsampling instead to circumvent the this problem." Does that mean Qwen2VL should use BICUBIC? but now it seems Qwen2vl still uses the nearest neighbor in Qwen2vlPlugin._preprocess_image.

[hiyouga]

@howardchenhd my bad, I'll fix it soon