WingCLIP-0.3

A 38.7M-parameter bird image encoder for on-device identification, the third step in a distillation chain that starts at BioCLIP-2. It reaches 86.90 top-1 on NABirds while being small enough to run in a browser tab.

It is the model behind WingDex.

Lineage

Read this before comparing numbers: the model has a teacher and a grand-teacher, and they are not the same thing.

	model	visual params	NABirds top-1
grand-teacher	BioCLIP-2, ViT-L/14	304.0M	86.41
teacher	WingCLIP-0.1, ViT-B-16	86.6M	89.93
this model	WingCLIP-0.3, TinyCLIP-39M	38.7M	86.90

WingCLIP-0.1 was distilled from BioCLIP-2 and fine-tuned past it. WingCLIP-0.3 was then distilled from WingCLIP-0.1, not from BioCLIP-2.

So: this model does not beat its teacher. It retains 96.6% of WingCLIP-0.1's NABirds accuracy at 45% of the parameters. It does edge out the grand-teacher BioCLIP-2, at 7.9x fewer visual parameters, which is the more useful headline but a comparison two steps removed from what it actually learned from.

Why distil from the student rather than the original teacher

Both were tried, on a 401-species NABirds pilot:

teacher	val_cos	NABirds top-1
BioCLIP-2	0.9616	83.44
WingCLIP-0.1	0.9612	89.09

BioCLIP-2 wins on embedding-copy fidelity and loses the task by 5.65 points. A teacher already specialised to birds transfers better than a larger general biology model, and val_cos will not tell you that.

What it is

A visual tower whose output is projected into the 768-d BioCLIP-2 embedding space and L2-normalized. forward() is the whole exportable graph: no text encoder runs at inference time.

Classification is a cosine similarity against a frozen 11,167 x 768 matrix of BioCLIP-2 text embeddings, shipped here as text_classifier_fp32.npy. So the model covers 11,167 bird species even though only 7,555 had enough photos to distil on. A species needs a name to be predictable, not training images.

Files

file	what
wingclip-0.3.pt	the original training checkpoint, with args and the WiSE-FT metadata
wingclip-0.3.safetensors	the same weights, without the pickle
wingclip-0.3-alpha.pt	after distillation, before fine-tuning. val cosine 0.9436
wingclip-0.3-beta.pt	after fine-tuning, before the WiSE-FT merge
text_classifier_fp32.npy	11,167 x 768 frozen BioCLIP-2 text embeddings
labels.json	11,167 rows of [common name, scientific name, eBird code], in classifier row order
onnx/wingclip_visual_fp32.onnx	fp32 export, parity-checked against PyTorch
onnx/wingclip_visual_int8.onnx + .data	int8, 39 MB across two files, what WingDex ships to the web

Reconstructing the alpha sweep

The release is an exact linear interpolation of the two stages:

released = (1 - a) * alpha + a * beta        # a = 0.60, verified to 0.0e+00

So wingclip-0.3-alpha.pt and wingclip-0.3-beta.pt regenerate any point on the sweep. Measured NABirds top-1, for reference:

a	0.25	0.40	0.50	0.60	0.75	0.90
top-1	86.27	86.64	86.82	86.90	86.90	86.56

0.60 and 0.75 tie; 0.60 ships. The optimum sits below WingCLIP-0.1's 0.90 because this fine-tune moves more of a smaller model.

Usage

import json
import numpy as np
import timm
import torch
import torch.nn.functional as F
from huggingface_hub import hf_hub_download
from PIL import Image
from timm.data import create_transform, resolve_data_config

REPO = "johnlian/WingCLIP-0.3"


class WingCLIP(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.visual = timm.create_model(
            "vit_medium_patch16_clip_224.tinyclip_yfcc15m",
            pretrained=False, num_classes=0)
        self.proj = torch.nn.Linear(512, 768)

    def forward(self, x):
        return F.normalize(self.proj(self.visual(x)), dim=-1)


model = WingCLIP().eval()
ckpt = torch.load(hf_hub_download(REPO, "wingclip-0.3.pt"), map_location="cpu")
model.load_state_dict(ckpt["model"])

cfg = resolve_data_config({}, model=model.visual)
preprocess = create_transform(**cfg, is_training=False)

classifier = np.load(hf_hub_download(REPO, "text_classifier_fp32.npy"))
labels = json.load(open(hf_hub_download(REPO, "labels.json")))

img = preprocess(Image.open("bird.jpg").convert("RGB")).unsqueeze(0)
with torch.no_grad():
    emb = model(img).numpy()

sims = (emb @ classifier.T)[0]
for i in sims.argsort()[-5:][::-1]:
    print(f"{sims[i]:.4f}  {labels[i][0]}  ({labels[i][1]})")

Preprocessing is standard CLIP: 224x224, bicubic, crop_pct 0.9, mean (0.48145466, 0.4578275, 0.40821073), std (0.26862954, 0.26130258, 0.27577711).

The text embeddings were built with the prompt "a photo of {common name}, {scientific name}, a species of bird.", so keep that template if you rebuild the classifier.

How it was trained

1. Distil. A TinyCLIP-39M visual tower regresses WingCLIP-0.1 embeddings over 2,503,107 iNaturalist photos across 7,555 species, 25 epochs at lr 8.1e-5. Reaches val cosine 0.9436.
2. Fine-tune. 12 epochs of supervised fine-tuning on held-out ground-truth photos, lr 1e-5, weight decay 0.1, label smoothing 0.1, light augmentation.
3. Merge. WiSE-FT interpolation between the distilled and fine-tuned weights at alpha = 0.60, chosen by a sweep on NABirds.

Step 3 matters. The fine-tune alone trades away out-of-distribution accuracy; the merge buys it back and then some.

Quantization

Measured on all 24,633 NABirds images. agree is top-1 agreement with fp32.

precision	top-1	tower MB	cos(fp32)	agree
fp32	86.91	155	1.000000	100.00%
int8	86.82	38.9	0.999923	99.27%
int4 block 32	84.61	21.9	0.988648	91.15%
int4 block 64	84.06	20.7	0.982274	88.99%
int4 block 128	81.50	20.1	0.974508	85.97%

int8 is effectively free. int4 at block 128 is a different model rather than a noisier one: 0.974 cosine and 86% agreement is a real behaviour change.

Reranking with an occurrence prior

Raw zero-shot argmax is not the end of the story. WingDex reranks the top 25 candidates against an iNaturalist occurrence prior that is conditioned on both the grid cell and the month:

score(species) = sim / T + beta * log P(species | cell, month)

On a 3,322-photo validation split, against a 97.14 recall ceiling:

stage	top-1
raw argmax, vision only	81.10
+ occurrence prior, pooled over months	93.80
+ month-aware prior (what ships)	95.09

Month is worth a further +1.0 to +1.2 points, with a paired-bootstrap 95% confidence interval of [+0.78, +1.60] over 2,000 resamples, so the interval excludes zero. It matters because a species can be common in a cell in July and absent in January, and a prior with no time dimension scores both the same.

The prior blob and the fitted T, beta and k live in the WingDex repo, not here.

Limitations

• Birds only. It has no notion of "not a bird". Feed it a dog and it returns a bird. Abstention has to be handled by the caller.
• Low confidence means species ambiguity, not bad framing. Top-1 confidence against relative bird area is Pearson 0.051. Cropping and retrying will not reliably rescue a low-confidence prediction.
• North-American evaluation. NABirds is the deciding benchmark, so accuracy outside North America is less well characterised.
• Long tail. Of 11,167 predictable species, 7,555 were distilled and 3,850 were fine-tuned. The rest ride entirely on the text embedding of their name.
• Two distillation steps from the original. Errors in WingCLIP-0.1 are inherited, and there is no path back to BioCLIP-2's behaviour through this model.

Licence and attribution

Weights are CC BY-NC 4.0. The WingDex source code is MIT, but the weights are trained on iNaturalist photos of which 1,923,704 are CC-BY-NC, so non-commercial propagates to the weights. WingDex is a strictly non-commercial project and this model inherits that.

Training data: iNaturalist Open Data. 2,503,107 images, 7,555 species, 62,423 credited observers. ShareAlike-licensed photos were excluded. Per-photo attribution is recorded in attributions.csv in the WingDex repo.

licence	photos
CC-BY-NC	1,923,704
CC-BY	389,106
CC-BY-NC-ND	126,460
CC0	57,960
CC-BY-ND	5,877

Upstream model licences: TinyCLIP-39M via timm (MIT) supplied the initial weights; BioCLIP-2 (MIT) supplied the embedding space.

Citation

WingCLIP has no paper of its own; cite the repository and the work below.

TinyCLIP, the backbone architecture and initial weights:

@inproceedings{wu2023tinyclip,
  title     = {{TinyCLIP}: CLIP Distillation via Affinity Mimicking and Weight Inheritance},
  author    = {Wu, Kan and Peng, Houwen and Zhou, Zhenghong and Xiao, Bin and Liu, Mengchen and Yuan, Lu and Xuan, Hong and Valenzuela, Michael and Chen, Xi and Wang, Xinggang and Chao, Hongyang and Hu, Han},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2023},
  eprint    = {2309.12314},
  archivePrefix = {arXiv}
}

BioCLIP 2, the grand-teacher and the source of the embedding space:

@inproceedings{gu2025bioclip2,
  title     = {{BioCLIP} 2: Emergent Properties from Scaling Hierarchical Contrastive Learning},
  author    = {Gu, Jianyang and Stevens, Sam and Campolongo, Elizabeth and Thompson, Matthew and Zhang, Net and Wu, Jiaman and Kopanev, Andrei and Mai, Zheda and White, Alexander and Balhoff, James and Dahdul, Wasila and Rubenstein, Daniel and Lapp, Hilmar and Berger-Wolf, Tanya and Chao, Wei-Lun and Su, Yu},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {38},
  pages     = {102778--102811},
  year      = {2025},
  eprint    = {2505.23883},
  archivePrefix = {arXiv}
}

BioCLIP, which BioCLIP 2 builds on:

@inproceedings{stevens2024bioclip,
  title     = {{BioCLIP}: A Vision Foundation Model for the Tree of Life},
  author    = {Stevens, Samuel and Wu, Jiaman and Thompson, Matthew J and Campolongo, Elizabeth G and Song, Chan Hee and Carlyn, David Edward and Dong, Li and Dahdul, Wasila M and Stewart, Charles and Berger-Wolf, Tanya and Chao, Wei-Lun and Su, Yu},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {19412--19424},
  year      = {2024}
}

WiSE-FT, the alpha-0.60 weight interpolation in step 3:

@inproceedings{wortsman2022robust,
  title     = {Robust fine-tuning of zero-shot models},
  author    = {Wortsman, Mitchell and Ilharco, Gabriel and Kim, Jong Wook and Li, Mike and Kornblith, Simon and Roelofs, Rebecca and Gontijo-Lopes, Raphael and Hajishirzi, Hannaneh and Farhadi, Ali and Namkoong, Hongseok and Schmidt, Ludwig},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2022},
  eprint    = {2109.01903},
  archivePrefix = {arXiv}
}

MobileCLIP2, which supplied the distillation recipe bundle:

@article{faghri2025mobileclip2,
  title   = {{MobileCLIP2}: Improving Multi-Modal Reinforced Training},
  author  = {Faghri, Fartash and Vasu, Pavan Kumar Anasosalu and Koc, Cem and Shankar, Vaishaal and Toshev, Alexander and Tuzel, Oncel and Pouransari, Hadi},
  journal = {Transactions on Machine Learning Research},
  year    = {2025},
  eprint  = {2508.20691},
  archivePrefix = {arXiv}
}

NABirds, the evaluation dataset:

@inproceedings{vanhorn2015nabirds,
  title     = {Building a Bird Recognition App and Large Scale Dataset With Citizen Scientists: The Fine Print in Fine-Grained Dataset Collection},
  author    = {Van Horn, Grant and Branson, Steve and Farrell, Ryan and Haber, Scott and Barry, Jessie and Ipeirotis, Panos and Perona, Pietro and Belongie, Serge},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {595--604},
  year      = {2015}
}

CLIP and OpenCLIP, the underlying method and the training library:

@inproceedings{radford2021clip,
  title     = {Learning Transferable Visual Models From Natural Language Supervision},
  author    = {Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya},
  booktitle = {Proceedings of the 38th International Conference on Machine Learning (ICML)},
  year      = {2021},
  eprint    = {2103.00020},
  archivePrefix = {arXiv}
}

@software{ilharco2021openclip,
  title  = {OpenCLIP},
  author = {Ilharco, Gabriel and Wortsman, Mitchell and Wightman, Ross and Gordon, Cade and Carlini, Nicholas and Taori, Rohan and Dave, Achal and Shankar, Vaishaal and Namkoong, Hongseok and Miller, John and Hajishirzi, Hannaneh and Farhadi, Ali and Schmidt, Ludwig},
  year   = {2021},
  doi    = {10.5281/zenodo.5143773}
}