OrchidCLIP

orchid-clip-v8 is a CLIP model fine-tuned from BioCLIP 2 for fine-grained orchid identification. Orchidaceae is one of the largest plant families and one of the most heavily long-tailed domains in biological vision — a handful of cultivated genera dominate every public image source while thousands of tropical-epiphyte species have fewer than 30 labeled images on the entire internet. v8 lifts top-1 accuracy from 0.873 (BioCLIP 2 baseline) to 0.911 on a stratified 4,000-image holdout, with the gains concentrated exactly where they should be: long-tail Pleurothallidinae genera gain +14 to +28 pp.

Headline

The +3.8 pp top-1 lift comes at no cost in genus-top-1, confirming that the gains are real species discrimination within genera, not a coarsening of the decision boundary.

The long tail is the point

The biggest lifts come on the smallest, longest-tailed genera — Stelis (~1,300 species worldwide, n=25 in our holdout) gains +24 pp; Lepanthes (~1,200 species, n=40) gains +27.5 pp. Head genera like Ophrys (n=2,754) gain modestly but never regress.

What lifted the long tail (and what didn’t)

The training pool is 1.14M images covering 5,124 species after WCVP synonym dedup, a previous-generation cosine quality filter, and a min_species ≥ 3 threshold. The class-frequency distribution is heavily skewed:

The recipe that worked:

1. Inverse-square-root long-tail sampler — sample each (binomial, image) pair with weight ∝ 1/√n_rows_in_class, with a per-species cap of 2,000. Less aggressive than uniform-by-class (which over-corrects and hurts head accuracy), much more tail-friendly than uniform-by-row.
2. WCVP 2026 synonym dedup — of 26,928 unique binomials in the raw label space, 4,504 binomials covering 69,467 image rows resolved to a different accepted name. The largest single confusion in our previous-generation v7 model was Ophrys fuciflora → Ophrys holosericea (29% of all v7 errors), which collapsed entirely under WCVP because they are the same accepted species.
3. Previous-generation cosine filter — drop the bottom percentile of training rows by previous-generation image↔binomial cosine. Rows that score poorly against their claimed binomial under a previous orchid-specific model are most likely label errors or off-target images.

Three substantial ablations against this recipe each underperformed:

• v9 — backbone swap to BioCLIP 2.5-H ViT-H/14 — regressed −2.5 pp top-1.
• v10 — hierarchical genus-species sampler — regressed on macro-genus (cardinality-blind across genera).
• v11 — auxiliary genus classification head — lifted genus-top-1 by +0.6 pp but regressed top-1 by −0.7 pp.

The lesson: at this scale, in this domain, the dominant variable is the label distribution. Architectural and auxiliary-objective changes that would help on a balanced dataset can actively hurt when the underlying label space is heavily skewed and noisy.

Errors are taxonomy-shaped

When v8 is wrong, it’s wrong in a structured way. Bucketing errors by WCVP rank distance between the true and predicted class:

Errors at d=1 are 58× more common than chance; cross-subfamily mistakes are essentially absent. The right framing for downstream consumers of v8’s top-1 prediction is “this genus, probably this species” rather than as a hard species label. The model’s effective competence is at the genus level, with a residual species-level disambiguation problem in cryptic-species complexes.

Status

The frozen image encoder is being prepared for public release as a foundation embedding for downstream orchid tasks. A draft paper and the full v9–v11 negative-result ablations are in musharna/orchid-sdxl. The cover figure on this page is a UMAP projection of v8 species centroids colored by WCVP subfamily — Cypripedioideae and Vanilloideae form clean islands while the two megadiverse subfamilies (Epidendroideae and Orchidoideae) partially overlap. That overlap is the natural target for the v12 hierarchical-distance-weighted contrastive loss currently in training.