ceruleandeep / ComfyUI-LLaVA-Captioner

A ComfyUI extension for chatting with your images with LLaVA. Runs locally, no external services, no filter.
GNU General Public License v3.0
98 stars 11 forks source link

ComfyUI LLaVA Captioner

A ComfyUI extension for chatting with your images. Runs on your own system, no external services used, no filter.

Uses the LLaVA multimodal LLM so you can give instructions or ask questions in natural language. It's maybe as smart as GPT3.5, and it can see.

Try asking for:

llava_captioner

NSFWness (FAQ #1 apparently)

The model is quite capable of analysing NSFW images and returning NSFW replies.

It is unlikely to return an NSFW response to a SFW image, in my experience. It seems like this is because (1) the model's output is strongly conditioned on the contents of the image so it's hard to activate concepts that aren't pictured and (2) the LLM has had a hefty dose of safety-training.

This is probably for the best in general. But you will not have much success asking NSFW questions about SFW images.

Installation

  1. git clone https://github.com/ceruleandeep/ComfyUI-LLaVA-Captioner into your custom_nodes folder
    • e.g. custom_nodes\ComfyUI-LLaVA-Captioner
  2. Open a console/Command Prompt/Terminal etc
  3. Change to the custom_nodes/ComfyUI-LLaVA-Captioner folder you just created
    • e.g. cd C:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-LLaVA-Captioner or wherever you have it installed
  4. Run python install.py
  5. Download models from 🤗 into models\llama:

Usage

Add the node via image -> LlavaCaptioner

Supports tagging and outputting multiple batched inputs.

Requirements

This is easy to install but getting it to use the GPU can be a saga.

GPU inference time is 4 secs per image on a RTX 4090 with 4GB of VRAM to spare, and 8 secs per image on a Macbook Pro M1. CPU inference time is 25 secs per image. If your inference times are closer to 25 than to 5, you're probably doing CPU inference.

Unfortunately the multimodal models in the Llama family need about a 4x larger context size than the text-only ones, so the llama.cpp promise of doing fast LLM inference on their CPUs hasn't quite arrived yet. If you have a GPU, put it to work.

See also