EvViz2 - Embedding Vector Visualizer 2

What is this?

This is an extension for stable-diffusion-webui which visualizes the embedding vector generated by CLIP.

Example

Usage

1. Input prompt as you like.
2. Click Run button.
3. Wait a second.

Detail

Embedding vectors (upper scatter)

This figure shows the embedding vector for each token. Each vector has 768 (for SDv1) or 1024 (for SDv2) dimensions.

Correlations of each token (lower heatmaps)

This figure shows correlations between each token. Calculation is carried out as follows:

1. Compute an embedding vector v from the given prompt. v is typically has dimension (77, 768). For xattn, v is converted by to_k linear layer.
2. For each token t, create a new prompt with the t replaced by padding token. Then compute its embedding vector v_{t}.
3. Let d_{t} = v - v_{t}.
4. Let d_{t,n} is nth row vector of d_{t}. d_{t,n} is a 768(or 1024)-dimensional vector representing t's effect on nth token. Then compute |d_{t,n}| where |x| is norm of a vector x.
5. Repeat procedure 2..3 for all t in the given prompt.

By default, padding token is _</w> (ID=318).