The Art of Prompt Engineering: A Deep Dive into the CLIP Interrogator

The CLIP Interrogator AI model is a powerful prompt engineering tool.

The Art of Prompt Engineering: A Deep Dive into the CLIP Interrogator
The CLIP Interrogator—a powerful prompt engineering tool that blends art and AI to create masterpieces. Discover how it works, and how you can use it to unleash your creativity.

There's something remarkable about the intersection of art and technology. When the two worlds collide, the boundaries of what's possible expand, and new forms of expression emerge. In recent times, AI models have emerged as powerful tools for artists, researchers, and enthusiasts alike, allowing for the creation of stunning and unique visual experiences.

One such tool that has captured my imagination is the CLIP Interrogator—an innovative prompt engineering tool that combines OpenAI's CLIP and Salesforce's BLIP to optimize text prompts that match a given image. The results are astounding, and the possibilities are limitless.

Before we begin our journey, I want to take a moment to acknowledge the incredible work done by the team at Replicate and the Replicate Codex community. Replicate provides a platform that makes running machine learning models in the cloud as easy as writing code, with no servers to set up. The Replicate Codex, built in collaboration with Replicate, is the most comprehensive resource for exploring and discovering AI models available on Replicate. You can find models like Stable Diffusion, Text-to-Pokemon, GFPGAN, and Codeformer on the platform. And as someone who created Replicate Codex, I couldn't be prouder of what we've built together.

So, without further ado, let's dive in and explore the magic of the CLIP Interrogator.

The Essence of the CLIP Interrogator

The CLIP Interrogator, created by pharmapsychotic, is a marvelous tool for artists and prompt engineering enthusiasts. It leverages the power of OpenAI's CLIP models to test a given image against a variety of artists, mediums, and styles. The goal is to study how different AI models perceive the content of the image, and the results are nothing short of fascinating.

But that's not all. The CLIP Interrogator goes a step further, combining the results with BLIP captioning to suggest a text prompt that can be used to create more images similar to the input. The implications of this are profound, as it opens the door to artistic exploration and creativity.

For those interested in using text-to-image models like Stable Diffusion to create cool art, the CLIP Interrogator is a game-changer. It's not just about generating images—it's about understanding the nuances of AI perception and harnessing that knowledge to create something beautiful.

How to Run the Model: A Step-by-Step Guide

To help you get started with the CLIP Interrogator, I've prepared a detailed, step-by-step guide on how to run the model. Whether you're a beginner or an intermediate user, I'll walk you through the process in a simple and conversational tone.

Step 1: Authenticate with Your API Token

To begin, you'll need to authenticate with Replicate using your API token. Copy your API token and set it as an environment variable using the following command:

export REPLICATE_API_TOKEN=[token]

Step 2: Call the HTTP API with cURL

With authentication in place, you're ready to call the HTTP API directly using cURL. Provide the input image and the desired parameters in the data payload. You can pipe the output into a command-line tool like jq to pretty-print it. Here's the command you'll use:

curl -s -X POST \
  -d '{"version": "a4a8bafd6089e1716b06057c42b19378250d008b80fe87caa5cd36d40c1eda90", "input": {"image": "YOUR_IMAGE_HERE", "clip_model_name": "ViT-L-14/openai", "mode": "best"}}' \
  -H "Authorization: Token $REPLICATE_API_TOKEN" \
  -H 'Content-Type: application/json' \
  "" | jq

Replace "YOUR_IMAGE_HERE" with the base64 encoded string of the image you want to test. Choose the clip_model_name based on whether you're using Stable Diffusion 1 (ViT-L-14/openai) or Stable Diffusion 2 (ViT-H-14/laion2b_s32b_b79k). The mode can be set to either "best" (takes 10-20 seconds) or "fast" (takes 1-2 seconds).

Step 3: Review the API Response

The API response is your new prediction as a JSON object. Initially, the status is "starting," and there's no output yet. Here's an example of what the response looks like:

  "completed_at": null,
  "created_at": "2023-03-08T17:54:26.385912Z",
  "error": null,
  "id": "j6t4en2gxjbnvnmxim7ylcyihu",
  "input": {"image": "..."},
  "logs": null,
  "metrics": {},
  "output": null,
  "started_at": null,
  "status": "starting",
  "version": "a4a8bafd6089e1716b06057c42b19378250d008b80fe87caa5cd36d40c1eda90"

Step 4: Refetch the Prediction

Since the prediction may take some time to complete, you'll need to refetch it from the API using the prediction ID from the previous response. Use the following command:

curl -s -H "Authorization: Token $REPLICATE_API_TOKEN" \
  -H 'Content-Type: application/json' \
  "" | jq "{id, input, output, status}"

If the prediction has completed, you'll see a response like this:

  "id": "j6t4en2gxjbnvnmxim7ylcyihu",
  "input": {"image": "..."},
  "output": "...",
  "status": "succeeded"

The output contains the suggested text prompt based on the input image.

For models that take longer to return a response, you may need to poll the API periodically for an update. Alternatively, you can specify a webhook URL to be called when the prediction is complete. Check out the webhook docs for details on setting that up.

Understanding the Inputs and Outputs

The CLIP Interrogator requires specific inputs, and it produces an output in a specific format. Let's take a closer look at these aspects.


  1. image (image file): The input image that you want to test. This image will be analyzed by the CLIP Interrogator to generate an optimized text prompt. The image should be provided as a base64 encoded string.
  2. clip_model_name (string): The CLIP model you want to use for the analysis. You can choose between two options: ViT-L-14/openai for Stable Diffusion 1, and ViT-H-14/laion2b_s32b_b79k for Stable Diffusion 2.
  3. mode (string): The prompt generation mode. You can choose between "best" and "fast." The "best" mode provides higher quality results but takes 10-20 seconds to complete, while the "fast" mode is quicker, taking only 1-2 seconds.

Output Schema

The output of the CLIP Interrogator is provided in the following JSON schema:

  "type": "string",
  "title": "Output"

The output is a string representing the suggested text prompt based on the input image. This prompt can be used with text-to-image models like Stable Diffusion to create more images similar to the input image.

Exploring the Possibilities with CLIP Interrogator

The true beauty of the CLIP Interrogator lies in the possibilities it opens up for artists, researchers, and enthusiasts. By understanding how AI models perceive images and generating optimized text prompts, you can create stunning visual art that pushes the boundaries of creativity.

Imagine transforming an image of a serene landscape into a masterpiece inspired by Van Gogh's style, or using a photograph of a city skyline to generate futuristic concept art. The CLIP Interrogator empowers you to explore different styles, mediums, and interpretations, all with a few simple commands.

What's more, the CLIP Interrogator is just one of many AI models available on the Replicate Codex platform. Whether you're interested in restoring old photos with GFPGAN, creating your own Pokémon with Text-to-Pokemon, or exploring code transformations with Codeformer, Replicate Codex offers a wealth of models to explore and experiment with.


As we reach the end of this comprehensive guide, I hope you've gained a deeper understanding of the CLIP Interrogator and the potential it holds for artistic expression. The journey doesn't end here—there's much more to explore, learn, and create.

The CLIP Interrogator is a powerful tool that bridges the gap between art and AI, allowing us to generate text prompts that match a given image and use them to create beautiful, unique art. And as part of the Replicate Codex community, you have access to a diverse array of AI models that can inspire and fuel your creativity.

In the ever-evolving world of AI and art, the only limit is your imagination. So go forth, experiment, and unleash your creative spirit. The canvas awaits.

Subscribe or follow me on Twitter for more content like this!