AI image generator in LoveCore AI - Guide
Published on 2024-08-15 by Audrey Wójcik
LoveCore AI features an AI image generator with minimal restrictions. Users can request images and selfies from their AI companions, including NSFW content.
We support hyper realistic images as well as anime.
Currently, only female characters are supported, but we are working on male characters, non-binary and other humanoid support.
How it works
In LoveCore AI, you can chat with your AI girlfriend, boyfriend, or friend, and ask them to send you a picture during your conversation. The AI generates an image based on your request and sends it to you. The image can depict the AI character in various poses, locations, outfits, and expressions.
Any image you've received is also viewable in the character gallery, not just in your chat history. You can mark images as favorites.
During character creation, in addition to the usual options like race, hair color, eye color, body type, you can customize the image prompt for all images, by alterign the image prompt field to create unique characters.
You can also define the face of your character by uploading an reference image. This only works for realistic characters at the moment.
How image requests are detected
Image requests can be detected in both user messages (typically) and AI character messages. Trigger phrases like "send a picture" or "can I see" in user messages, and "send a picture" in character messages, will initiate an image request.
- Pose is determined based on words in the message that triggered the request.
- Location is the current setting of the chat. You can infer the current location from the chat background or by asking the character. You can also ask the character to change locations during the chat.
- Clothes are based on what the character is wearing or on words in the prompt. Future features may allow users to buy and change the character's clothes using hearts. We may also add accessories in the future.
- Expression is based on the character's mood.
Tech details
LoveCore AI's image generator uses Stable Diffusion as its underlying engine. Stable Diffusion is open-source, has no output restrictions, and has a large community of users training models and LoRAs for it.
A drawback of Stable Diffusion and most AI image generators is that they don't understand the laws of reality and human anatomy, so the generated images can be unusual and sometimes unsettling. A notable example is its difficulty in generating accurate hands and feet, often creating too many or too few fingers and toes. You might also encounter multiple sets of hands, legs, or other strange anomalies.
Typically, you need to generate many images and select the best ones. While this is acceptable for an offline process when creating images as an artist, it's not ideal for a chat application. Therefore, we need to provide tools and restrictions to increase the rate of good image generation. Anecdotally, depending on the prompt, checkpoint, and LoRAs used, you might get 1 good image out of 3 with standard Stable Diffusion, while our system aims to improve that to 9 out of 10.
Our secret sauce
In LoveCore AI, we have an intermediary server that translates image generation requests into prompts that are then fed to Stable Diffusion.
- pose: the pose to use. This generally involves using appropriate Controlnet images and/or LoRAs. If not provided, the server selects an image from the pose list based on weights - some poses are more common than others. Currently, there are about 400 poses for female characters.
- location: the location to use. If not provided, a random location is selected based on weights.
- clothes: what the character is wearing. If not provided, it can be selected from a list.
- expression: the character's expression. If not provided, it can be selected from a list.
- prompt_extra: this is added to the prompt as text. Currently, this parameter is not populated from chat messages because we cannot guarantee a good experience with current technology.
To optimize resources and response time, the server first searches the database for images matching the criteria that haven't been sent to the user before. If no image is found, the server generates a new image and sends it to the user. There are daily limits on the number of images that can be generated by a single user, and the server will refuse to generate new images if the limit is reached, but can respond with a pre-generated image using random criteria.
We are always working on improving our system and your experience and these details are based on the current implementation as of writing of this article. More improvements will be coming to the future. You can reach out with your feedback or suggestions at [email protected].