musharna/orchid-species-embeddings

Orchid species embedding gallery

18,858 orchid species as 768-d unit-normalized embeddings, one mean vector per binomial, aggregated from 101,713 image embeddings by musharna/orchid-clip-v8 — a BioCLIP-2 ViT-L/14 fine-tuned for orchid identification.

Orchidaceae is one of the largest plant families and one of the most heavily long-tailed domains in biological vision. This is the retrieval gallery from that work, released so the species-level vectors are usable without re-running the model — or obtaining the images, which cannot be redistributed.

Schema

column	type	description
binomial	string	species name, e.g. Cattleya trianae. Unique — no duplicates
genus	string	349 genera. 100 % coverage
tribe	string	nullable — 271 rows unknown
subfamily	string	nullable — 256 rows unknown
n_images	int32	image embeddings averaged into this vector
embedding	list[float32, 768]	L2-normalized mean embedding

Every vector has ‖v‖ = 1.0000, so cosine similarity is a plain dot product. No NaNs or infinities.

⚠️ n_images is the most important column

The distribution is severely long-tailed, and this is the point rather than a flaw:

n_images	species	share
exactly 1	7,562	40.1 %
exactly 2	3,212	17.0 %
exactly 3	1,095	5.8 %
≥ 30	87	0.5 %

Median is 2; maximum is 61.

For 40 % of species, "centroid" is a misnomer — the vector is a single image's embedding wearing a species label, carrying that photograph's pose, lighting, background and any misidentification. Treat n_images == 1 rows as weak evidence. Filter or weight by n_images for any quantitative claim; an unweighted mean over this gallery is dominated by species that were only ever photographed once.

Genus coverage is likewise uneven: Bulbophyllum (1,922), Dendrobium (1,243), Stelis (1,200), Lepanthes (1,110) and Epidendrum (835) account for a large share of all rows.

Usage

from datasets import load_dataset
import numpy as np

ds = load_dataset("musharna/orchid-species-embeddings", split="train")
E  = np.array(ds["embedding"], dtype="float32")     # (18858, 768), unit-norm
names = ds["binomial"]

# nearest species to a query embedding (already L2-normalized)
sims = E @ query_vec                                 # cosine, since ‖v‖=1
top  = np.argsort(-sims)[:10]
print([(names[i], float(sims[i])) for i in top])

Restricting to better-supported species:

ds_solid = ds.filter(lambda r: r["n_images"] >= 5)   # 18858 -> far fewer, but each is a real mean

Query vectors must come from the same encoder — musharna/orchid-clip-v8. Embeddings from stock BioCLIP-2 or OpenCLIP live in a different space and will return nonsense.

What retrieval quality to expect

Measured on this gallery, 200 randomly drawn species queried against all 18,858:

	real species queries	random unit vectors
mean top-1 cosine	0.914	0.082 (max 0.223)
mean share of top-5 sharing the query's genus	0.655	—

So the space carries real structure — an 11× separation from noise — but top-5 congener agreement is ~66 %, not ~100 %. Roughly a third of near neighbours cross genus boundaries. Expect that: some orchid genera are genuinely hard to separate visually, and the 40 % of rows built from a single image are noisy by construction. Use this for candidate generation and similarity search, not as a taxonomic oracle.

Licence and provenance

The embeddings are released under MIT, matching the model that produced them.

The source photographs are not included and are not redistributable. They were assembled from public listings and orchid-registry pages whose images remain under their respective owners' terms. What is published here is a derived feature: a lossy 768-d projection, not a reproduction of any photograph, and for most rows a mean over several of them. Species-level averaging puts these a further step from any individual image.

Stated precisely rather than absolutely: nothing in this repository renders or reconstructs a source photograph, and no pixels are redistributed. The claim is that these are derived features, not that embedding inversion is impossible in general — note that the 7,562 rows with n_images == 1 are a single image's embedding rather than an average, so they sit closest to their source.

If you need images, collect them yourself; this gallery is deliberately the part that can be shared.

Related

• Model: musharna/orchid-clip-v8 (MIT)
• Live demo: musharna/orchid-genus-id
• Code and the honest write-up of where fine-grained transfer hits a wall: https://github.com/musharna/orchid-clip
• Base model: imageomics/bioclip-2

Citation

@misc{arnold_orchid_species_embeddings,
  author = {Arnold, Jaret},
  title  = {Orchid species embedding gallery (orchid-clip-v8)},
  year   = {2026},
  url    = {https://huggingface.co/datasets/musharna/orchid-species-embeddings}
}