ComfyUI_NAIDGenerator

A ComfyUI extension for generating image via NovelAI API.

Installation

• git clone https://github.com/bedovyy/ComfyUI_NAIDGenerator into the custom_nodes directory.
• or 'Install via Git URL' from Comfyui Manager

Setting up NAI account

Before using the nodes, you should set NAI_ACCESS_TOKEN on ComfyUI/.env file.

NAI_ACCESS_TOKEN=<ACCESS_TOKEN>

You can get persistent API token by User Settings > Account > Get Persistent API Token on NovelAI webpage.

Otherwise, you can get access token which is valid for 30 days using novelai-api.

Usage

The nodes are located at NovelAI category.

Txt2img

Simply connect GenerateNAID node and SaveImage node.

Note that all generated images via GeneratedNAID node are saved as output/NAI_autosave_12345_.png for keeping original metadata.

Img2img

Connect Img2ImgOptionNAID node to GenerateNAID node and put original image.

Note that width and height of the source image will be resized to generation size.

Inpainting

Connect InpaintingOptionNAID node to GenerateNAID node and put original image and mask image.

Note that both source image and mask will be resized fit to generation size.

(You don't need MaskImageToNAID node to convert mask image to NAID mask image.)

Vibe Transfer

Connect VibeTransferOptionNAID node to GenerateNAID node and put reference image.

You can also relay Img2ImgOption on it.

Note that width and height of the source images will be resized to generation size. This will change aspect ratio of source images.

Multiple Vibe Transfer

Just connect multiple VibeTransferOptionNAID nodes to GenerateNAID node.

ModelOption

The default model of GenerateNAID node is nai-diffusion-3(NAI Diffusion Anime V3).

If you want to change model, put ModelOptionNAID node to GenerateNAID node.

NetworkOption

You can set timeout or retry option from NetworkOption node. Moreover, you can ignore error by ignore_errors. In that case, the result will be 1x1 size grayscale image. Without this node, the request never retry and wait response forever, and stop the queue when error occurs

Note that if you set timeout too short, you may not get image but spend Anlas.

PromptToNAID

ComfyUI use () or (word:weight) for emphasis, but NovelAI use {} and []. This node convert ComfyUI's prompt to NovelAI's.

Optionally, you can choose weight per brace. If you set weight_per_brace to 0.10, (word:1.1) will convert to {word} instead of {{word}}.

Director Tools

You can find director tools like LineArtNAID or EmotionNAID on NovelAI > director_tools.

V4 Support (Preview)

The node now supports NAI's V4 architecture through the nai-diffusion-4-curated-preview model. This is a preview release of V4 with some limitations:

• Important Notes:
  • This is a preview version of V4 and some features are limited
  • Inpainting will automatically use V3 model (but works with V4-generated images)
  • Vibe transfer is not yet supported with V4 preview (will be available with full V4 release)
  • Full V4 feature support will come with the official V4 release

New Model Option

NAI Diffusion V4 Curated Preview is now available in the ModelOptionNAID node:

model = "nai-diffusion-4-curated-preview"

V4 Prompt Handling

Two new nodes have been added for V4 prompt handling:

V4BasePrompt

A node for handling V4 positive prompts:

V4BasePrompt -----> positive
               GenerateNAID

V4NegativePrompt

A node for handling V4 negative prompts:

V4NegativePrompt -> negative
                 GenerateNAID

Example V4 Workflow

Here's a basic V4 setup:

V4BasePrompt -----> positive
V4NegativePrompt -> negative  GenerateNAID
ModelOption ------> option

Work In Progress Features

The following V4 features are currently in development:

"""
- V4PromptConfig: Advanced prompt configuration
  - Coordinate-based prompting
  - Order-based prompting
- V4CharacterCaption: Character-specific prompting with positioning
"""

Note: Basic img2img functionality works with V4 preview. For inpainting, the node will automatically use V3 model but can still work on V4-generated images. Vibe transfer will be supported once V4 fully releases.