CLIP ViT-B/32 model
Embeddings are compatible with CLIP text embeddings and with Image Embedding (CLIP) vectors for cross-modal search.
CatalogLatency Synchronous; depends on frame count and video length Async false Rate Limit Per API key Max Input No hard limit, but very large frame counts slow the response; use max_frames to cap
cn8/studio-embed-video
A single synchronous endpoint powered by CLIP ViT-B/32 (via sentence-transformers).
Learn more about Video Embedding (CLIP) VideosyncInstant previewvideoanalysisembedding
Try it
Video to embed
Sample rate (frames / second)
Max frames (optional)
Result
The result shows up here
Upload your video on the left and hit Run.
Overview
A single synchronous endpoint powered by CLIP ViT-B/32 (via sentence-transformers). studio-embed-video decodes a video, samples frames at a configurable rate (sample_fps), encodes each sampled frame with CLIP, and returns a list of 512-dimensional unit-normalised vectors — one per frame. Use the vectors for semantic video-scene retrieval, nearest-neighbour search, and multimodal RAG. Embeddings share a semantic space with CLIP text embeddings and with the Image Embedding (studio-embed-image) endpoint for cross-modal search.
Key capabilities
Video frame embedding
The video is decoded; frames are sampled at sample_fps (default 1 FPS). Each sampled frame is encoded in a single batch call. Returns one embedding per frame.
Configurable sampling
Control sampling density with sample_fps (e.g. 0.5 = one frame every 2 seconds). Use max_frames to cap the total number of frames for very long videos.
Synchronous response
Embeddings are returned directly in the JSON body — no polling.
When to use it
Video scene retrieval
Sample frames from a video, embed them, and search for specific scenes or objects across a video library.
Multimodal RAG
Store frame embeddings in a vector database; retrieve relevant video segments for RAG pipelines.
Video similarity
Compute cosine similarity between frame embeddings to measure visual similarity across videos.
Input & output
input
Video URL with optional sample_fps / max_frames
JSON body
output
List of 512-dimensional float vectors (one per sampled frame)
JSON
Guides & tips
CLIP embedding basics
- CLIP (ViT-B/32) produces 512-dimensional float vectors normalised to unit length.
- Vectors from the same model are directly comparable with cosine similarity.
- Text CLIP embeddings (from an external CLIP text encoder) are compatible: use them to query a frame vector index.
Choosing sample_fps for video
- 1.0 FPS (default): good for typical 1–10 minute educational or presentation videos.
- 0.5 FPS: coarser; useful for long videos (>30 min) to keep frame count manageable.
- 2.0–4.0 FPS: finer; useful for action video where scenes change quickly.
- Use max_frames as a safety cap to avoid unexpectedly large responses.
Storing and querying embeddings
- Store embeddings in a vector database (e.g. Qdrant, Pinecone, Weaviate, pgvector).
- Store each frame embedding with its frame index / timestamp as metadata.
- At query time, encode the query image or text with CLIP, then run nearest-neighbour search.
Specs
Schema
Request body
videoUrlrequiredstringPublic URL of the source video (mp4, mov, mkv, webm, avi)
sample_fpsnumber · default: 1Frame sampling rate in frames-per-second (e.g. 1.0 = 1 frame/sec, 0.5 = 1 frame/2 sec). Lower values reduce cost and response size.
max_framesintegerOptional cap on total frames to encode. Useful for very long videos. null = no limit.
Pricing
Billed per video processed, regardless of frame count.
| Service | Unit | Price |
|---|---|---|
| Video Embedding (CLIP) | video | 2 credits/video |
Price is per video regardless of frame count or sample_fps. Use max_frames to control response size and latency.