WildID › Models

The models behind WildID: BioCLIP 2 and BioCLIP 2.5

WildID contains no recognition model of its own. It runs BioCLIP, an open vision model for biology, exactly as its authors published it. This page explains what the model does, how its two versions differ and which one to choose.

What BioCLIP is

BioCLIP 2 is a vision model for photographs of living things. It was published by researchers at the Imageomics Institute and trained on TreeOfLife-200M, a very large collection of images labelled with their position in the tree of life. It is released under the MIT licence, and the WildID setup downloads it unchanged from its page on Hugging Face.

How it names a photo

The model turns your photo into a list of numbers. It holds a matching list of numbers for every name in its catalogue, 867,455 taxa in the case of BioCLIP 2. The names whose numbers lie closest to those of the photo come out on top.

The score next to each name compares the candidates with each other. It is not the probability that a name is right, and it depends on the candidate list: a score of 0.9 is easier to reach among 400 regional species than among all 867,455 names. Scores from runs with different place lists cannot be compared.

The two versions side by side

BioCLIP 2BioCLIP 2.5
Role in WildIDDefaultOptional, chosen per run
Image encoderViT-L/14ViT-H/14, larger
Names it can give867,455 taxa794,878 taxa
ReptilesYesNone at all
Lemur photos named correctly in our test77.1 %80.8 %
Graphics memory while working in Lightroomabout 7 GBabout 10 GB
Memory while working, on the tested Macabout 12 GBabout 13 GB
Speed on a folder of JPEGs, RTX 4090144 photos/s98 photos/s
Speed on a folder of JPEGs, Mac with an M5 Pro33 photos/s17 photos/s
Download4.5 GB7.3 GB
The lemur figure is from 401 RAW photos of four lemur groups, with the crop on and the Madagascar place list. All test results

BioCLIP 2.5 does not know reptiles

The name catalogue published for BioCLIP 2.5 contains no reptiles: no lizards, no snakes, no turtles and no crocodiles. It cannot answer “chameleon”, so it answers with something else and gives that answer a confident score. On 63 of our Madagascar reptile photos BioCLIP 2 named all 63 correctly with the Madagascar list, and BioCLIP 2.5 named none.

This is why BioCLIP 2 is the default. The Lightroom dialog warns you when you pick BioCLIP 2.5 while reptiles could be among the candidates.

Which one to choose

  • BioCLIP 2 for anything that might include a reptile, and as the first pass over a new import.
  • BioCLIP 2.5 for shoots of mammals or birds only, if your graphics card or your Mac has the memory. On our lemur photos it was ahead by 2 to 4 percentage points with a place list and by 6 to 10 without one.

BioCLIP 2.5 was not ahead everywhere. On 22 photos of Kenyan mammals BioCLIP 2 named 21 correctly and BioCLIP 2.5 named 19. Its authors report better results than BioCLIP 2 on their own benchmarks, and they note that it needs more computing resources.

A routine that uses both: identify a new import with BioCLIP 2, then select the Mammalia collection set and run Identify again with BioCLIP 2.5. Photos whose species changes move to the new collection.

Switching models in Lightroom

The Identify dialog offers both models for every run. The first use of a model in a session loads it, which takes a few seconds. With “Keep both models loaded” ticked, switching costs no waiting time. That option needs about 14 GB of graphics memory and is offered when the card reports 20 GB or more. On a Mac it needs about 20 GB of memory and is offered when the Mac reports 30 GB or more as usable.

BioCLIP 2.5 must have been downloaded. If you chose only BioCLIP 2 during the setup, run WildID Setup again and answer all.

The two helper models

MegaDetector is an open animal detector from Microsoft’s AI for Good Lab. It finds the animal in the frame so that BioCLIP looks at the animal and not at the landscape around it. WildID uses the most confident detection, with a little margin around it. If nothing is detected, the whole frame is used.

SpeciesNet is Google’s classifier for camera-trap images. It is strongest on mammals and weak on birds and reptiles. When you switch on the second opinion, WildID asks it about the same crop and records whether it agrees with BioCLIP, lands in the same genus, family, order or class, or disagrees.

WildID does not train, tune or alter any of these models.

Back to How it worksNext: GPU, VRAM and speed