CLIP Encoding - ManimGL Explainer

ManimGL animation source code for explaining how CLIP maps images and text into a shared embedding space.

This repo is a visual explainer for the core CLIP idea: image patches and text tokens are encoded separately, projected into the same vector space, normalized, compared with cosine similarity, and then used for retrieval or classification-like decisions.

What This Teaches

• How image and text encoders produce vectors.
• Why a shared embedding space makes image-text matching possible.
• How L2 normalization and cosine similarity shape CLIP comparisons.
• How similarity matrices, UMAP-style projections, and decision boundaries can make embedding behavior easier to see.

Repo Map

• clip_encoding.py - main CLIP encoding overview scene.
• clip_encoding_scene_1.py - shared embedding space.
• clip_encoding_scene_2.py - CLIP similarity matrix.
• clip_encoding_scene_3.py - L2 normalization and cosine similarity.
• clip_encoding_scene_4.py - UMAP-style embedding visualization.
• clip_encoding_scene_5.py - MLP decision boundary scene.
• clip_encoding_scene_6.py - bad sinks and cluster behavior.
• example_photos/ - small reference images used by the scenes.
• requirements.txt - pinned ManimGL-related dependencies.

Setup

Create and activate a virtual environment:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

If the PyPI manimgl package has import issues on your Python version, install ManimGL from source instead:

pip install /path/to/manimgl

Running A Scene

Interactive mode with live preview:

manimgl clip_encoding.py CLIPEncoding -i --autoreload

Render the main scene to video:

manimgl clip_encoding.py CLIPEncoding -w --hd

Render a specific scene:

manimgl clip_encoding_scene_3.py L2NormalizationCosineSimilarity -w --hd

Rendered output is written under videos/.

Why This Exists

CLIP is often introduced with a single diagram, but the intuition lives in the geometry: normalization, similarity, clusters, and boundaries. This repo keeps the animation source public so the explanation is inspectable and reusable.

Related Explainers

• LLM Arithmetic
• Hybrid Search + BM25
• Neural Superposition