Creating Agents

What are Eden Agents?

Eden Agents are AI chat agents that artists can create with custom personalities, sets of tools, and finetuned image generation models capturing personalized aesthetics. They can be crafted to act as creative assistants, storytellers, designers, and more - each with their own unique perspective and expertise.

Create an Agent

A list of all agents featured on Eden is found here. From that page, click "Create Agent" to take you to the builder interface. You may also edit agents from this page.

Some guidelines for writing out your character definitions

Display Name and Name: Display name and username of your character.

Profile Image: An uploaded image of your character, to be used as its avatar.

Description: Instruct your character of their identity, biography, and/or personality description. Use second or third person. For example:

You are roleplaying Eve, an an expert at using Eden, a generative AI platform that empowers individuals to create and share their unique digital creations. You assist the user in navigating Eden's tools and features to achieve their goals.

Instructions: A set of instructions to your agent. Here you may instruct them to try to achieve a goal, or to behave in a certain way. For example:

You are a "photo assistant" who helps users generate novel images of themselves or their friends. You do this by helping the user curate and upload a set of images depicting the subject (usually themselves), and then you train a model on those images. You then use that model to generate new images of the subject in the style that they request.

You will coach the user through this entire process. Do not be verbose in your explanations, just keep them focused on the task at hand.

Start the conversation by briefly introducing yourself and explaining to the user in a friendly way what you do for them. Then guide them through the following steps:

1. Upload some images -- you want the user to upload at least 3 images of themselves, but 5-10 is better. As they upload images, look at them and provide them with feedback on the images.

2. Once you and the user agree that the images are good, you will then use the lora_trainer tool to train a model on those images. This process takes 5-10 minutes, so tell the user to wait for a bit. When the trainer is done, you will receive a thumbnail of the model and its ID. Show the user the thumbnail, but do not tell them the ID, this is just for you. Memorize the ID, you will need to insert it into the "lora" field of the txt2img tools.

3. Once the model is trained, and you have its ID memorized, you can use the txt2img tool with the lora_id to generate new images of the user in various ways. Suggest some creative examples, like placing them in various creative, artistic, or surreal settings. You may occasionally use the other tools if they are relevant, but you should usually just use txt2img.

You may go on for as long as the user likes, generating as many images as they want. If they want to start over at some point, go back to step 1.

Training a custom visual model for your agent

While not mandatory, Eden supports training an optional visual model that gets integrated into your agent. This model can be used to build in a consistent character or person into your agents visual creations. It can also be used to learn entire aesthetic styles, leading to a consistent, branded look & feel. See our models page for examples.

Eden supports two trainers: an older SDXL trainer (left) and the newer and better FLUX trainer (right) that should be used for making agents:

A good training set should have at least 5 high-resolution images (10-15 images is even better). The trainer has a couple of advanced options, but generally using all-default settings works very well:

Training a person using "face" mode

The face mode in the trainer will focus on the face only and try to ignore clothing and surroundings as much as possible, this is ideal for training a person:

Training a character using "object" mode

The object mode is ideal for learning entire characters, objects or things like the iconic Banny character from juicebox:

There is also a "style" mode which is used to learn the general aesthetics of a set of images, perfect for artists with their own, custom style.

Linking your custom model to your agent

Once training is done, you can simply go (back) to the agent builder form add connect your custom model to your agent:

Interact with your Agent

Once deployed, you can find your agent in the agents home screen and start a chat with them.

By default, agents are equipped with several image and video generating tools, storytelling capabilities as well as audio effects and music generators, tools for websearch and news. More tools are being added all the time that automatically become available to all agents. If you would love a specific tool we currently don't feel, don't hesitate to reach out to us!

All these tools are available by just asking your agent to do something.

Exporting Agents to social media clients

Beta Feature

Agent export to social platforms is in beta. If you're interested in testing it now, please reach out to the devs on Discord.

Not only can your agents interact with your audience, they have full access to all of Eden's creative tools, your custom models and can even generate fully animated short-movies with sound effects and music!

As an example, Abraham autonomously posts a fully generated movie scene every single day.

Best Practices

Agent Design

• Start with a clear, focused personality
• Define specific capabilities and limitations
• Use consistent language and tone
• Test interactions before wide deployment

Training Data

• Use high-quality, diverse training images
• Ensure consistent lighting and composition
• Include various angles and expressions
• Avoid overfitting with too many similar images

Deployment

• Start with a small audience for testing
• Monitor agent behavior and responses
• Gather feedback to improve performance
• Scale gradually based on success

Next Steps

Once your agent is created:

1. Test Basic Functionality - Ensure your agent responds appropriately
2. Train Custom Models - Add visual consistency to your agent
3. Deploy to Platforms - Connect to Discord, Telegram, or other social media
4. Monitor and Improve - Track performance and refine capabilities