The multimodal warehouse · Images
Find the image you need by describing it, or by showing one like it.
Search by text, by a reference image, or by a face or a logo inside the picture. Every image is stored as embeddings, faces and labels in the same warehouse as your video and documents, so a product photo can find the video scene and the PDF page it appears in.
From $25/mo. Index once. Query every way.
| embedding | What the picture shows, searchable by text or image | whole image |
| face | One face, with a quality score | box 412, 88 |
| label | “outerwear › trench coat” from your taxonomy | whole image |
| metadata | sku, campaign and photographer, as you supplied them | fields |
One warehouse
Your images sit beside every other file you own.
Each file is broken into rows that point back to the file they came from. Video scenes, images, audio segments and document blocks live in one namespace, so a question can start in one kind of file and finish in another.
Joins your images make
Find every video moment that shows this product photo.
product photosvideo scenes
Runs on the Feature search stageWhich uploads carry one of our registered logos?
uploaded imagesyour logo set
Runs on the Cross compare stageAdd price and stock from the catalog database to every match.
product photosyour SQL table
Runs on the SQL lookup stageWhat you can ask
Ask in plain words, or with a file.
› a beige trench coat on a city street
Images that match the description, ranked
› this photo
Near duplicates and derivative versions
› this face
Every image the person appears in
› outdoor shots from the spring campaign
Semantic matches filtered by your own fields
Start in one click
Point it at your images
An empty namespace
Connect the bucket your images already live in and pick the extractors. Studio indexes it and your first query runs against your own files.
Create a namespace →In code
One call to search what was indexed.
The extractors that turn images into rows:
- Image →SigLIP embeddings for images, pages and frames
- Face identity →Detect and match faces
- Multimodal →One embedding space shared with video and text
from mixpeek import Mixpeek
client = Mixpeek(api_key="YOUR_API_KEY")
# Search by image URL
results = client.retrievers.execute(
"img-search-retriever",
inputs={"query_image_url": "https://example.com/query-image.jpg"},
filters={
"AND": [
{"key": "category", "value": "product", "operator": "eq"}
]
},
limit=20
)
for result in results["documents"]:
print(f"Score: {result.score}, ID: {result.document_id}")Questions
What people ask about images search
What is an image search API?
An image search API allows developers to build applications that search, compare, and retrieve images programmatically. Instead of relying on filenames or manual tags, it uses computer vision and vector embeddings to understand visual content, enabling searches by image similarity, natural language descriptions, or a combination of both.
How does visual similarity search work?
Visual similarity search works by converting images into high-dimensional vector embeddings using deep learning models like CLIP or SigLIP. These embeddings capture semantic features such as objects, colors, textures, and composition. When you query with an image, the system finds vectors closest to it in the embedding space using approximate nearest neighbor search, returning the most visually similar results.
What is reverse image search?
Reverse image search takes an image as input and finds matching or similar images in a database. Unlike text-based search, it analyzes the visual content directly to find exact duplicates, near-duplicates (cropped, resized, or filtered versions), and semantically similar images. Common use cases include source attribution, duplicate detection, and copyright enforcement.
What image formats does Mixpeek support?
Mixpeek supports all major image formats including JPEG, PNG, WebP, TIFF, BMP, and GIF (first frame). Images are automatically preprocessed and normalized during ingestion, so you do not need to convert formats before uploading. The API also handles varying resolutions and aspect ratios without manual resizing.
Can I use custom embedding models for image search?
Yes. Mixpeek supports bringing your own models via the plugin system. You can deploy custom PyTorch or ONNX models alongside builtin extractors like CLIP, SigLIP, and ResNet. Custom models run on the same GPU infrastructure and integrate directly into the feature extraction pipeline with no additional setup.
How does image search scale to millions of images?
Mixpeek uses distributed processing on Ray GPU clusters for feature extraction and Mixpeek Vector Store for vector indexing, which provides sub-millisecond approximate nearest neighbor search even at billions of vectors. Ingestion pipelines automatically scale horizontally, and storage tiering moves cold data to S3 while keeping hot vectors in memory for fast retrieval.
What is the difference between image search and image recognition?
Image recognition identifies and classifies objects, scenes, or attributes within a single image (e.g., 'this image contains a cat'). Image search uses those recognized features to find similar or matching images across a corpus. Mixpeek combines both: recognition happens during feature extraction, and the resulting embeddings power search and retrieval.
Can I combine image search with text and metadata filters?
Yes. Mixpeek retrievers support multi-stage pipelines that combine vector similarity search with metadata filtering. You can search by image similarity and then filter results by date, category, tags, or any custom metadata field. You can also combine image and text queries in a single search for more precise results.
Put your images in the warehouse.
Index them once, then ask them anything, alone or joined to the rest of your files.
Other file types:VideoAudioDocumentsHow the warehouse works