Orchid Vision
A computer-vision program built on orchids - generation, recognition, and guided hybridization - because the long tail is where fine-grained models actually break.
Orchid Vision is the umbrella for the orchid-focused computer-vision work on this site.
Orchids are an unusually hard test for fine-grained botanical vision. WCVP accepts over 30,000 species, more than any other plant family, the distribution is heavily long-tailed, and a large share of online image labels are wrong in subtle, load-bearing ways: right genus with the wrong species, obsolete synonyms treated as distinct taxa, cultivar names standing in for species. Most of the work in this program has gone into curation rather than architecture.
Three pieces, built in this order
OrchidGAN — generative
StyleGAN2-ADA fine-tuned on a curated set of Cattleya. Established that orchid floral morphology is learnable and produced a navigable latent "morphospace" supporting seed sampling, interpolation and style mixing. This is where the program started.
orchid-clip-v8 — recognition
A long-tail-aware CLIP for fine-grained orchid ID: +7.6 pp top-1 averaged per genus over the BioCLIP 2 baseline (0.844 vs 0.768), and up to +28 pp on the rarest. Across six independent attempts, genus structure transfers while within-genus species identity stays locked.
Cattleya Hybrid Visualizer — guided hybridization
SDXL plus an ancestry LoRA, prompted from a botanical phenotype engine that blends parent traits as independent pigment channels with dominance rules and generation-dependent recessive thresholds. Predicts what a cross would plausibly look like, four to seven years before it flowers.
Where it goes next: closing the loop
The three pieces are not yet a system. The next step is to let recognition gate generation: score each generated bloom against a per-species prototype and reject the ones that miss, so generation becomes a measurable loop instead of something a human eyeballs. That loop is designed but not built.
The through-line
Each piece exists because the previous one exposed a limit.
OrchidGAN showed the morphology was learnable but had no notion of which orchid it had drawn. That demanded recognition, so orchid-clip-v8 was trained, and it found a ceiling: genus transfers, species does not. The live demo serves a calibrated genus and names a species only when the top-1/top-2 margin earns it, so it abstains instead of committing to a confidently wrong binomial.
The visualizer inherits both results. It can generate, but the recognition model cannot fully referee the output yet, so the phenotype engine carries the botanical constraints instead of trusting the diffusion model to infer them.
Data
The corpus behind orchid-clip-v8 was assembled from iNaturalist research-grade observations, GBIF, the Smithsonian NMNH, Wikimedia Commons, CC-licensed Flickr and OrchidRoots, then put through a quality-filter pipeline: CLIP-cosine outlier rejection, perceptual-hash deduplication, and a GBIF lineage filter to catch synonym collapse before it reached training.