[hmellor]

What does this PR do?

Note

Stacked on #50671. Only the last commit belongs to this PR; I'll rebase once #50671 merges.

Adds BaseImageProcessor._merge_and_standardize_kwargs, which fills in the kwargs that weren't passed from the instance attributes and then standardizes them (as suggested in #50671 (comment)). Video processors inherit it.

It's used wherever passed kwargs override the instance defaults:

• preprocess: replaces the hand-written defaulting loop and _standardize_kwargs call in the image and video base classes, and in the three video processors that copied them (Ernie 4.5 VL MoE, MiniCPM-V 4.6, MiniCPM-OCR).
• Image/video processor counting methods: 82 methods (mostly get_number_of_image_patches / get_num_of_video_patches) now look up images_kwargs["k"] instead of images_kwargs.get("k", self.k) / .get("k", None) or self.k.
• Processors: ~40 _get_num_multimodal_tokens methods (and LFM2-VL's _get_image_num_tokens) call self.image_processor._merge_and_standardize_kwargs(...) (or the video processor's) instead of reading defaults off it.

The method name combines the existing ProcessorMixin._merge_kwargs (defaults overridden by passed kwargs) and _standardize_kwargs.

The conversion was done with an AST script that only rewrites a lookup when the key is one of the class's valid_kwargs and the default is the attribute of the same name. Modular files were updated and the generated files regenerated.

Behaviour changes

• Passed kwargs in the counting methods are now standardized like in preprocess, e.g. a user-passed size becomes a SizeDict (Llama4's manual conversion is removed).
• An explicit None/0 passed to a counting method is no longer replaced by the default (_merge_kwargs never produces None values, so this only affects manual calls).
• Video preprocess standardizes kwargs before frame sampling rather than after.
• 13 get_num_of_video_patches methods and MuseGlimmerImageProcessor.get_number_of_image_patches no longer crash when called without videos_kwargs/images_kwargs.

Not converted: lookups whose key differs from the attribute (e.g. get("patch_size", self.size) in GOT-OCR2 and Cohere2-Vision) and unrelated kwargs.get(..., self.x) outside counting methods.

Testing

• pytest on every test_processing_*, test_image_processing_* and test_video_processing_* file of the 56 touched models (112 files): 3611 passed, 0 failed.
• All test_image_processing_* and test_video_processing_* files: 3967 passed, 1 failed (pp_doclayout_v3 test_post_process, which also fails without this PR and isn't touched by it).
• utils/check_modular_conversion.py and utils/check_copies.py report nothing for the touched files.

This PR was written with AI assistance (Claude Code).

🤖 Generated with Claude Code

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: aria, aya_vision, blip_2, chameleon, cohere2_vision, cohere_compass, colqwen2, cosmos3_edge, deepseek_ocr2, deepseek_vl, deepseek_vl_hybrid, dots3_note, embedding_gemma2, ernie4_5_vl_moe, exaone4_5, gemma4