Custom nodes for ComfyUI.
The workflow above is embedded in the image. You can load it in ComfyUI.
Clone the repository to custom_nodes in your ComfyUI directory and install the requirements:
git clone https://github.com/Zuellni/ComfyUI-Custom-Nodes custom_nodes\Zuellni
pip install -U -r custom_nodes\Zuellni\requirements.txt
To update execute the following command in the same directory:
git -C custom_nodes\Zuellni pull
All required models are downloaded to to the .cache directory automatically.
| Name | Description |
|---|---|
| Aesthetic Loader | Loads models for use with Aesthetic Filter. |
| Aesthetic Filter | Returns x best images and a list of their indexes based on cafe_aesthetic/cafe_waifu scoring. If no models are loaded then acts like LatentFromBatch and returns 1 image with 1-based index. |
| Aesthetic Select | Takes latents and a list of indexes from Aesthetic Filter and returns only the selected latents. |
A poor man's implementation of DeepFloyd IF. Text encoder requires more than 8GB of VRAM. To download the models you will have to agree to the terms of use on the site, create an access token and log in with it.
| Name | Description |
|---|---|
| IF Loader | Loads models for use with other IF nodes. Device can be used to move the models to specific devices, eg cuda:0, cuda:1. Leaving it empty enables cpu offloading. |
| IF Encoder | Encodes prompts for use with IF Stage I and IF Stage II. Setting unload to True removes the model from memory. Prompts can be reused without having to reload the encoder. |
| IF Stage I | Takes the prompt embeds from IF Encoder and returns images which can be used with IF Stage II or other nodes. |
| IF Stage II | As above, but also takes Stage I or other images and upscales them x4. |
| IF Stage III | Upscales Stage II or other images using Stable Diffusion x4 upscaler. Doesn't work with IF Encoder embeds, has its own encoder accepting string prompts instead. Setting tile to True allows for upscaling larger images than normally possible. |
| Name | Description |
|---|---|
| Latent Decode | Combines VAEDecode and VAEDecodeTiled. Probably not necessary since VAEDecodeTiled is now used on error, but just here for the sake of completeness. |
| Latent Encode | As above, but adds batch_size. Allows loading 1 image and denoising it x times without having to create multiple sampler nodes. |
| Multi Crop | Center crops images/latents to specified dimensions. |
| Multi Noise | Adds random noise to images/latents. |
| Multi Repeat | Allows for repeating images/latents x times, similar to Latent Encode. |
| Multi Resize | Similar to LatentUpscale but uses scale instead of width/height. Works with both images and latents. |
| Load Folder | Loads all images in a specified directory. The images will be cropped/resized if their dimensions aren't equal. |
| Share Image | Saves images without metadata in a specified directory. Counter resets on restart. Useful for sharing images without having to remove prompts manually. |