hyper3labs/hyper3-clip-v0.5

Hyper3-CLIP v0.5

Hyper3-CLIP v0.5 is an open-weight hyperbolic vision-language checkpoint from hyper³labs. It places image and text representations in a Lorentz space and was trained with compositional entailment constraints for hierarchy-sensitive image-text retrieval.

This v0.5 release is intended as an open baseline and research artifact.

Model

• Architecture: ViT-B scale vision-language model
• Vision backbone: vit_base_patch16_224
• Text backbone: openai/clip-vit-base-patch32
• Embedding dimension: 512
• Training steps: 500,000
• Global batch size: 768
• Weights artifact: model.safetensors

The original full training checkpoint included optimizer, scheduler, AMP scaler, RNG state, config, and step metadata. This repository publishes the weights-only model.safetensors artifact for inference and downstream research.

Quick Start: Sentence Transformers

The default way to use this checkpoint is through Sentence Transformers. The adapter in this repository returns 512-dimensional L2-normalized tangent-space embeddings for standard cosine/dot-product vector stores.

Install the runtime dependencies:

pip install "sentence-transformers>=5.5.1" timm safetensors pyyaml Pillow

If you are using the gated Hugging Face repository from a fresh machine, accept access on the model page and set HF_TOKEN.

from PIL import Image
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("hyper3labs/hyper3-clip-v0.5", trust_remote_code=True)

image_embedding = model.encode([Image.open("/path/to/image.jpg")], normalize_embeddings=True)
text_embedding = model.encode(["machined metal part"], normalize_embeddings=True)

Transformers

from PIL import Image
import torch
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("hyper3labs/hyper3-clip-v0.5", trust_remote_code=True).eval()
tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32")

image = model.preprocess_image(Image.open("/path/to/image.jpg")).unsqueeze(0)
text = tokenizer(
    ["machined metal part"],
    padding=True,
    truncation=True,
    max_length=model.config.max_text_length,
    return_tensors="pt",
)

with torch.no_grad():
    outputs = model(
        pixel_values=image,
        input_ids=text["input_ids"],
        attention_mask=text["attention_mask"],
    )

image_embedding = outputs.image_embeds
text_embedding = outputs.text_embeds

Haystack image retrieval pipeline

For indexing images in a Haystack retrieval pipeline, use SentenceTransformersDocumentImageEmbedder with image paths in Document.meta["file_path"], paired with SentenceTransformersTextEmbedder for text queries.

pip install "haystack-ai>=2.30.1" "sentence-transformers>=5.5.1" timm safetensors pyyaml Pillow

from haystack import Document
from haystack.components.embedders import SentenceTransformersTextEmbedder
from haystack.components.embedders.image import SentenceTransformersDocumentImageEmbedder

model_id = "hyper3labs/hyper3-clip-v0.5"

documents = [
    Document(
        content="front view of a machined metal part",
        meta={"file_path": "/path/to/image.jpg"},
    )
]

image_embedder = SentenceTransformersDocumentImageEmbedder(
    model=model_id,
    trust_remote_code=True,
    batch_size=8,
    normalize_embeddings=True,
)
documents = image_embedder.run(documents=documents)["documents"]

text_embedder = SentenceTransformersTextEmbedder(
    model=model_id,
    trust_remote_code=True,
    normalize_embeddings=True,
)
query_embedding = text_embedder.run("machined metal part")["embedding"]

Evaluation

The numbers below use the official evaluator convention for R@10. Higher is better except for TIE and LCA.

Model	Comparable setting	ImageNet top-1	COCO text R@10	COCO image R@10	Flickr text R@10	Flickr image R@10	TIE	LCA	Jaccard	H-Prec	H-Rec
MERU-B/16	same-family baseline	40.1	82.0	68.6	96.2	90.0	3.630	2.220	0.780	0.850	0.850
HyCoCLIP-B/16	official checkpoint	45.8	82.0	69.3	95.4	90.3	3.172	2.047	0.814	0.874	0.874
UNCHA-B/16	official checkpoint	48.8	82.6	71.0	95.9	91.2	2.945	1.961	0.828	0.883	0.884
PHyCLIP-B/16	related reported result	44.4	80.4	68.7	95.6	89.9	3.285	2.088	0.807	0.868	0.868
Hyper3-CLIP v0.5	this release	48.5	84.0	72.8	97.5	92.4	2.972	1.986	0.828	0.882	0.883

Raw evaluation files are included:

• eval_coco_karpathy_final.json
• eval_flickr30k_final.json
• eval_imagenet_final.json
• eval_hycoclip_uncha_intersection_final.json

License And Attribution

The model materials in this repository are released under OpenMDW-1.0. See LICENSE.

Redistributions should preserve NOTICE, LICENSE, and the original model card when practical. Modified or derived checkpoints should use a distinct name and must not imply endorsement by hyper³labs.

Please cite and link to the original hyper³labs model repository when publishing benchmarks, papers, derivative checkpoints, or public demos based on this model.

Intended Use

This release is intended for:

• hierarchy-sensitive image-text retrieval research
• zero-shot and retrieval evaluation
• multimodal embedding baselines
• downstream experiments with hyperbolic representation learning

This model has not been validated for safety-critical use.

Citation

If you use Hyper3-CLIP v0.5, cite the original model repository and hyper³labs.