Intro

CLIP (Contrastive Language-Image Pretraining), Predict the most relevant text snippet given an image

it is ML (machine learning) model used in Immich to do text search on media (photo and video) files

Immich support a lot of model, than can be found at https://huggingface.co/immich-app and than configured in settings

Different models a takes different time to index media. Some model are Multilanguage, some are not.

default CLIP model is ViT-B-32__openai (not multilanguage)

Time of media storage processing

Time to process

first test stats on library with 320 000 photos and 17 000 videos with size of 5.4 TB

latest test process library with 331000 photos 22800 videos with size of 6 TB (same library with some new photos added)

ViT-B-16-SigLIP-512__webli

DevOps > Immich CLIP model test > 2024-08-22_205621.png

nllb-clip-large-siglip__mrl takes even more time than ViT-B-16-SigLIP-512__webli too process media

nllb-clip-large-siglip__v1

takes about 20 days

DevOps > Immich CLIP model test > nllb-clip-large-siglip__v1_time_to_index.png

XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k

on i7-1355U

DevOps > Immich CLIP model test > XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k.png

ViT-SO400M-16-SigLIP2-256__webli

Executed on 2 devices with same library. Using OpenVINO (this time). Results the same for English text. Some difference in Ukrainian.

on Intel Core i7 1355U

DevOps > Immich CLIP model test > ViT-SO400M-16-SigLIP2-256__webli.png

on Intel N150

DevOps > Immich CLIP model test > ViT-SO400M-16-SigLIP2-256__webli_Intel-N150.png

ViT-SO400M-14-SigLIP2-378__webli

Core i7-1355 + openVINO

DevOps > Immich CLIP model test > ViT-SO400M-14-SigLIP2-378__webli-openvino-i7-1355U.png

ViT-B-16-SigLIP-i18n-256__webli

Intel N150 + OpenVINO

DevOps > Immich CLIP model test > 2025-05-05_231850.jpg

ViT-H-14-378-quickgelu__dfn5b

Core i7-1355 + openVINO

DevOps > Immich CLIP model test > image-2025-5-24_20-15-37.png

Some official data

to be found at https://immich.app/docs/features/searching/#clip-models

Quality of search by some random queries

Since those models are tested mostly on predefined sets, I tested the quality of the search on my personal media folder and some semi-random queries in both English and Ukrainian.

That is some present of "quality" marks deviation based on "feel" results is good, not just formal.
The resulting search of "apple" like the Apple Computers logo is formally 100% correct, but at least 50% of apple fruit feels like more correct output

Summary

Word/Phrase	XLM-Roberta- Base-ViT-B-32__ laion5b_s13b_b90k	ViT-SO400M-16- SigLIP2-256 __webli	ViT-SO400M-16- SigLIP2-384 __webli	ViT-SO400M-14- SigLIP2-378 __webli	ViT-H-14-378- quickgelu __dfn5b	ViT-B-16- SigLIP-i18n-256 __webli	nllb-clip- large-siglip __mrl	nllb-clip- large-siglip __v1
Total score	65%	72%	74%	73%	56%	59%	56%	49%
English Score (/18)	85%	94%	98%	91%	95%	71%	71%	63%
anxiety	0	100	100	95	100	100	0	0
apple	98	95	95	95	95	20	80	75
apple tree	90	95	95	95	95	80	80	80
arguing people	95	95	95	75	100	95	90	90
car	95	80	95	75	75	60	80	80
car on a parking lot	95	90	95	85	100	70	100	100
city fog	95	95	100	100	100	75	100	80
exchange rate	93	95	95	90	80	70	80	70
feel sad	95	95	90	80	95	5	0	4
fog	95	70	100	70	95	70	70	70
green fence	100	100	100	100	98	100	50	50
honda civic	100	100	100	100	100	100	100	85
peach	65	100	100	100	95	95	55	50
person feel anxiety	98	97	98	96	98	95	95	65
Range Rover	35	95	100	90	90	35	50	60
river fog	100	100	100	100	100	100	85	90
serenity	85	100	100	100	95	65	75	0
TV	100	100	100	100	100	100	95	85
Ukrainian score (/31)	45%	51%	51%	56%	18%	48%	40%	35%
відчувати радість	50	0	0	0	0	0	25	28
відчувати сум	0	0	0	0	0	0	0	0
заправка wog	0	100	95	100	0	95	0	0
заправка кло	50	65	95	90	0	50	0	0
зелений забор	85	85	95	90	5	85	95	80
курс обміну валют	95	65	85	65	0	65	65	65
люди в лісі	100	100	100	100	85	100	5	5
люди які сперечаються	0	65	5	85	0	0	90	90
людина що відчувають розчарування	0	0	0	0	0	30	80	5
мати сумніви	0	0	0	3	0	0	0	0
машина на стоянці	100	100	95	100	60	45	3	3
машина	100	65	95	100	20	100	3	3
персик	95	100	5	95	70	0	55	30
почуття суму	0	0	0	0	0	0	0	0
радість	5	0	0	0	0	70	0	0
Рейндж Ровер	5	90	70	90	0	10	50	80
Розчарування	0	0	0	0	0	0	0	0
спокій	0	0	0	0	50	3	0	0
сум	0	0	0	0	0	0	0	0
сумніватись	0	0	0	0	0	0	0	0
сумувати	0	0	0	0	0	0	0	0
тв	90	80	97	95	80	90	95	80
туман на річці	95	95	95	95	0	95	95	95
туман у місті	100	95	95	95	0	100	50	40
туман	0	95	95	95	0	95	40	40
Хонда Сівік	100	90	90	100	0	100	100	90
Хонда Цивік	100	97	97	100	0	100	100	90
Хонда Цівік	100	97	97	100	0	100	100	90
ціна на бензин	80	40	92	50	80	65	100	90
яблуко	0	0	8	5	60	0	80	85
яблуня	70	70	70	80	85	90	20	0

Image found Examples

Car (English)

Model	Search Result
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k
XLM-Roberta-Large-Vit-B-16Plus
ViT-B-16-SigLIP-512__webli
nllb-clip-large-siglip__mrl
nllb-clip-large-siglip__v1 (half year later)
ViT-SO400M-16-SigLIP2-256__webli
ViT-H-14-378-quickgelu__dfn5b
ViT-SO400M-14-SigLIP2-378__webli
ViT-SO400M-16-SigLIP2-384__webli
ViT-B-16-SigLIP-i18n-256__webli

Green Fence (English)

Model	Search Result
XLM-Roberta-Large-Vit-B-16Plus
ViT-B-16-SigLIP-512__webli
nllb-clip-large-siglip__mrl
nllb-clip-large-siglip__v1 (half year later)
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k
ViT-H-14-378-quickgelu__dfn5b
ViT-SO400M-16-SigLIP2-256__webli
ViT-SO400M-14-SigLIP2-378__webli
ViT-SO400M-16-SigLIP2-384__webli
ViT-B-16-SigLIP-i18n-256__webli

туман на річці ("River fog" Ukrainian)

Model	Search Result
XLM-Roberta-Base-ViT-B-32__laion5b_s13b_b90k
XLM-Roberta-Large-Vit-B-16Plus
ViT-B-16-SigLIP-512__webli
nllb-clip-large-siglip__mrl
nllb-clip-large-siglip__v1 (half year later)
ViT-H-14-378-quickgelu__dfn5b	random people photos. ( 75% people and river)
ViT-SO400M-16-SigLIP2-256__webli	first output screen is random screenshots (need couple of page down to find the rivers) - considering that as a fail
ViT-SO400M-14-SigLIP2-378__webli
ViT-SO400M-16-SigLIP2-384__webli
ViT-B-16-SigLIP-i18n-256__webli