[ngxson]

Overview

• remove all non-pillow resize variants because HF impl always uses pillow
• track down all models's config.json and correct them one by one, see a complete table in comment

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: most code written + verified by AI. this is a mechanical task

[ngxson]

model (projector)	need correct?	correct algo	link to config
LLaVA 1.5 (MLP)	yes (was BILINEAR)	BICUBIC	preprocessor_config.json
LLaVA 1.6 (MLP, uhd)	yes (ov was BILINEAR)	BICUBIC (main/rf/ov)	preprocessor_config.json
MobileVLM (LDP/LDPV2)	yes (was BILINEAR)	BICUBIC	image_processing_clip.py (CLIP default)
CogVLM2 (COGVLM)	yes (was BILINEAR)	BICUBIC	modeling_cogvlm.py
GLM-Edge (GLM_EDGE)	yes (was BILINEAR)	BICUBIC	preprocessor_config.json (resample=3)
MiniCPM-V 2.x/4.x (MINICPMV)	yes (ov was BILINEAR)	BICUBIC (all resizes)	image_processing_minicpmv.py
MiniCPM-V 4.6 (MINICPMV4_6)	yes (ov was BILINEAR)	BICUBIC (all resizes)	same processor family as MiniCPM-V 2.6 above
InternVL (INTERNVL)	yes (ov was BILINEAR)	BICUBIC	preprocessor_config.json (resample=3)
Nemotron Nano V2 VL	no	BICUBIC	image_processing.py (torchvision BICUBIC)
Idefics3 / SmolVLM (IDEFICS3)	yes (was default BICUBIC)	LANCZOS	preprocessor_config.json (resample=1)
LFM2-VL (LFM2)	yes (was BILINEAR x3)	BICUBIC (main/rf/ov)	preprocessor_config.json (resample=3)
Phi-4-multimodal (PHI4)	no	BILINEAR	processing_phi4mm.py (torchvision resize, default interpolation)
Pixtral (PIXTRAL)	yes (was BILINEAR)	BICUBIC	preprocessor_config.json (resample=3)
LightOnOCR (LIGHTONOCR)	no	BICUBIC	preprocessor_config.json (resample=3)
dots.ocr (DOTS_OCR)	no	BICUBIC	preprocessor_config.json (Qwen2VL processor, default BICUBIC)
dots3-note (DOTS3NOTE_V)	no	BICUBIC	preprocessor_config.json (Qwen2VL processor, default BICUBIC)
Kimi-VL (KIMIVL)	yes (was BILINEAR)	BICUBIC	image_processing_kimi_vl.py
Kimi-K2.5 (KIMIK25)	no	BICUBIC	media_utils.py
Gemma 3 (GEMMA3)	no	BILINEAR	preprocessor_config.json (resample=2)
Gemma 4 (GEMMA4V/GEMMA4UV)	yes (was BILINEAR)	BICUBIC	processor_config.json (resample=3)
Gemma 3n (GEMMA3NV)	yes (was default BICUBIC)	BILINEAR	preprocessor_config.json (resample=2)
Qwen2-VL / Qwen2.5-VL / Qwen3-VL	yes (was BILINEAR)	BICUBIC	image_processing_qwen2_vl.py (default BICUBIC)
MiniMax-M3 (MINIMAX_M3)	no	BICUBIC	image_processor.py
Muse Glimmer (MUSE_GLIMMER)	no	LANCZOS	processor_config.json (resample=1)
MiMo-VL (MIMOVL)	no	BICUBIC	preprocessor_config.json (Qwen2VL processor, default BICUBIC)
Step3-VL (STEP3VL)	no (algo hardcoded in preprocessor, hparams unused)	BILINEAR (PIL stage)	processing_step3.py
Youtu-VL (YOUTUVL)	no	BILINEAR	preprocessor_config.json (resample=2)
Reka Edge (YASA2)	no	BICUBIC	preprocessor_config.json (resample=3)
GLM-4.xV (GLM4V)	yes (was BILINEAR)	BICUBIC	image_processing_glm4v.py (default BICUBIC)
Llama 4 (LLAMA4)	yes (was default BICUBIC)	BILINEAR	preprocessor_config.json (resample=2)
PaddleOCR-VL (PADDLEOCR)	yes (was BILINEAR)	BICUBIC	preprocessor_config.json (resample=3)
DeepSeek-OCR / OCR-2	no	BICUBIC (PIL resize default)	modeling_deepseekocr.py
HunyuanOCR (HUNYUANVL)	yes (was BICUBIC)	LANCZOS	preprocessor_config.json (resample=1)
EXAONE 4.5 (EXAONE4_5)	no	BICUBIC	preprocessor_config.json (resample=3)
Janus Pro (JANUS_PRO)	yes (was BILINEAR)	BICUBIC	preprocessor_config.json (resample=3)
Granite 4 vision (GRANITE4_VISION)	yes (ov was BILINEAR)	BICUBIC (main/rf/ov)	preprocessor_config.json (resample=3)

[ngxson]

Speed optimization (last commit)

Bicubic

case	master legacy bicubic	master pillow bicubic	this PR	vs legacy	vs master pillow
4032x3024 -> 1024x1024	60.6	18.4	12.6	4.8x	1.5x
1920x1080 -> 784x784	33.9	5.4	3.2	10x	1.7x
800x600 -> 336x336	6.6	1.16	0.75	8.8x	1.5x
512x512 -> 1024x1024 (up)	62.0	5.0	2.2	28x	2.3x
336x336 -> 3072x3072 (up)	536	29.4	10.0	53x	2.9x

Lanczos

case	master	this PR	speedup
4032x3024 -> 1024x1024	26.3	19.1	1.4x
1920x1080 -> 784x784	7.6	5.3	1.4x
800x600 -> 336x336	1.81	1.26	1.4x
512x512 -> 1024x1024 (up)	6.0	3.0	2.0x
336x336 -> 3072x3072 (up)	38.6	13.0	3.0x

Bilinear

Master's legacy bilinear is not a fair baseline: it had no antialiasing and used align-corners mapping,
so it did far less work and did not match Pillow. Numbers for reference:

case	master legacy bilinear (not PIL-correct)	this PR (PIL-exact)
4032x3024 -> 1024x1024	2.6	8.7
1920x1080 -> 784x784	1.5	3.0
800x600 -> 336x336	0.28	0.69
512x512 -> 1024x1024 (up)	2.6	1.6
336x336 -> 3072x3072 (up)	22.8	7.5