NovelAI-API

A lightweight asynchronous Python wrapper for NovelAI image generation API.

Features

• Lightweight - Focuses on image generation only, providing a simple and easy-to-use interface.
• Concurrent - Supports both API and web backend, allowing to run two generating tasks simultaneously.
• Parameterized - Provides a Metadata class to easily set up generation parameters with type validation.
• Asynchronous - Utilizes asyncio to run generating tasks and return outputs efficiently.

Important

Unfortunately, NovelAI has announced that image generation function of the API endpoint will be depreciated starting from Mar 21, 2024. As a result, concurrent generation feature will no longer be available.

Source

Installation

Install with pip:

pip install novelai

Note that this package requires Python 3.12 or higher. For Python 3.7-3.11, install the legacy version instead:

pip install novelai-legacy

Legacy branch has the same features as master branch on user side, the only difference is code compatibility.

Usage

Initialization

Import required packages and initialize a client with your NovelAI account credentials.

import asyncio
from novelai import NAIClient

# Replace argument values with your actual account credentials
username = "Your NovelAI username"
password = "Your NovelAI password"

async def main():
    client = NAIClient(username, password, proxy=None)
    await client.init(timeout=30)

asyncio.run(main())

Image Generation

After initializing successfully, you can generate images with the generate_image method. The method takes a Metadata object as the first argument, and an optional host argument to specify the backend to use.

By passing verbose=True, the method will print the estimated Anlas cost each time a generating request is going to be made.

The full parameter list of Metadata can be found in the class definition.

from novelai import Metadata, Host, Resolution

async def main():
    metadata = Metadata(
        prompt="1girl",
        negative_prompt="bad anatomy",
        res_preset=Resolution.NORMAL_PORTRAIT,
        n_samples=1,
    )

    print(f"Estimated Anlas cost: {metadata.calculate_cost(is_opus=False)}")

    # Choose host between "Host.API" and "Host.WEB"
    # Both of two hosts work the same for all actions mentioned below
    output = await client.generate_image(
        metadata, host=Host.WEB, verbose=False, is_opus=False
    )

    for image in output:
        image.save(path="output images", verbose=True)

asyncio.run(main())

Image to Image

To perform img2img action, set action parameter in Metadata to Action.IMG2IMG, and image parameter to your base image. The base image needs to be converted into Base64-encoded format. This can be achieved using base64 module.

import base64
from novelai import Metadata, Action

async def main():
    with open("tests/images/portrait.jpg", "rb") as f:
        base_image = base64.b64encode(f.read()).decode("utf-8")

    metadata = Metadata(
        prompt="1girl",
        negative_prompt="bad anatomy",
        action=Action.IMG2IMG,
        width=832,
        height=1216,
        n_samples=1,
        image=base_image,
        strength=0.5,
        noise=0.1,
    )

    output = await client.generate_image(metadata, verbose=True)

    for image in output:
        image.save(path="output images", verbose=True)

asyncio.run(main())

Inpainting

To perform inpaint action, set action parameter in Metadata to Action.INPAINTING, and image parameter to your base image, and mask parameter to the black and white mask image, where white is the area to inpaint and black to keep as is. Both base image and mask need to be converted into Base64-encoded format. This can be achieved using base64 module.

import base64
from novelai import Metadata, Model, Action, Resolution

async def main():
    with open("tests/images/portrait.jpg", "rb") as f:
        base_image = base64.b64encode(f.read()).decode("utf-8")

    with open("tests/images/inpaint_left.jpg", "rb") as f:
        mask = base64.b64encode(f.read()).decode("utf-8")

    metadata = Metadata(
        prompt="1girl",
        negative_prompt="bad anatomy",
        model=Model.V3INP,
        action=Action.INPAINT,
        res_preset=Resolution.NORMAL_PORTRAIT,
        image=base_image,
        mask=mask,
    )

    output = await client.generate_image(metadata, verbose=True)

    for image in output:
        image.save(path="output images", verbose=True)

asyncio.run(main())

Vibe Transfer

Vibe transfer doesn't have its own action type. Instead, it is achieved by adding a reference_image parameter in Metadata. The reference image needs to be converted into Base64-encoded format. This can be achieved using base64 module.

import base64
from novelai import Metadata, Resolution

async def main():
    with open("tests/images/portrait.jpg", "rb") as f:
        base_image = base64.b64encode(f.read()).decode("utf-8")

    metadata = Metadata(
        prompt="1girl",
        negative_prompt="bad anatomy",
        res_preset=Resolution.NORMAL_PORTRAIT,
        reference_image=base_image,
        reference_infomation_extracted=1,
        reference_strength=0.6,
    )

    output = await client.generate_image(metadata, verbose=True)

    for image in output:
        image.save(path="output images", verbose=True)

asyncio.run(main())

Concurrent Generation

By default, NovelAI only allows one concurrent generating task at a time. However, this wrapper provides the ability to simultaneously run two concurrent generating tasks by sending requests to API and web backend respectively.

Note that API and web backend both have limit on concurrent generation. Therefore, running more than two concurrent tasks will result in a 429 Too Many Requests error.

Full usage example is provided under /docs.

async def task_api():
    await client.generate_image(metadata, host=Host.API)
    print("API task completed")

async def task_web():
    await client.generate_image(metadata, host=Host.WEB)
    print("Web task completed")

async def main():
    tasks = [
        asyncio.create_task(task_api()),
        asyncio.create_task(task_web()),
    ]
    await asyncio.wait(tasks)

asyncio.run(main())

Use in CLI

Optionally, a module function is also provided to directly generate access token in CLI.

Once a access token is generated, it will be valid for 30 days. Token can be used as the authentication header to make requests to NovelAI.

# Replace argument values with your actual account credentials
python3 -m novelai login <username> <password>

References

NovelAI Backend

Aedial/novelai-api

NovelAI Unofficial Knowledgebase