[PR #11468] feat(ml): add more search models #13249

Closed
opened 2026-02-05 15:11:32 +03:00 by OVERLORD · 0 comments
Owner

Original Pull Request: https://github.com/immich-app/immich/pull/11468

State: closed
Merged: Yes


Description

This PR adds a number of OpenCLIP search models that give very competitive results, especially considering their efficiency.

Also updates the code for exporting models to ONNX, mainly making the search models static and adding a line to ensure none of the ops are in training mode.

Monolingual:
ViT-B-16-SigLIP-256__webli
ViT-B-16-SigLIP-384__webli
ViT-B-16-SigLIP-512__webli
ViT-B-16-SigLIP__webli
ViT-L-16-SigLIP-256__webli
ViT-L-16-SigLIP-384__webli
ViT-SO400M-14-SigLIP-384__webli

This is how they stack up (models in red are new):

clip_quality_vs_efficiency

Source for recall@k scores, considering only image retrieval.

Multilingual:
ViT-B-16-SigLIP-i18n-256__webli
nllb-clip-base-siglip__mrl
nllb-clip-large-siglip__mrl

This is how they stack up (models in red are new):

multilingual_clip_quality_vs_efficiency

multilingual_clip_quality_by_language

Source for recall@k scores, considering only image retrieval.

**Original Pull Request:** https://github.com/immich-app/immich/pull/11468 **State:** closed **Merged:** Yes --- ## Description This PR adds a number of OpenCLIP search models that give very competitive results, especially considering their efficiency. Also updates the code for exporting models to ONNX, mainly making the search models static and adding a line to ensure none of the ops are in training mode. Monolingual: ViT-B-16-SigLIP-256__webli ViT-B-16-SigLIP-384__webli ViT-B-16-SigLIP-512__webli ViT-B-16-SigLIP__webli ViT-L-16-SigLIP-256__webli ViT-L-16-SigLIP-384__webli ViT-SO400M-14-SigLIP-384__webli This is how they stack up (models in red are new): ![clip_quality_vs_efficiency](https://github.com/user-attachments/assets/8eef18dc-36f8-4353-a9d8-f147f9f063cd) [Source](https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_retrieval_results.csv) for recall@k scores, considering only image retrieval. Multilingual: ViT-B-16-SigLIP-i18n-256__webli nllb-clip-base-siglip__mrl nllb-clip-large-siglip__mrl This is how they stack up (models in red are new): ![multilingual_clip_quality_vs_efficiency](https://github.com/user-attachments/assets/fd70e159-2a08-4330-9d00-6992a9f259a0) ![multilingual_clip_quality_by_language](https://github.com/user-attachments/assets/6e3080b3-32d9-4990-9e2b-9a03f974eb3e) [Source](https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_multilingual_retrieval_results.csv) for recall@k scores, considering only image retrieval.
OVERLORD added the pull-request label 2026-02-05 15:11:32 +03:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: immich-app/immich#13249