A set of custom nodes for ComfyUI designed to natively initialize and run the historic original Latent Diffusion Model (LDM) F8 Large architecture from CompVis.

Unlike newer models (like Stable Diffusion 1.5), this specific pipeline relies on a heavy BERT-based text encoder (dim: 1280, vocab_size: 30522) and an aggressive f8 latent autoencoder. This repository enables standalone support for this unique model without relying on external .yaml configuration blueprints or old, deprecated codebases.

🚀 Key Features

Native LDM F8 Large Support: Full execution pipeline for text-to-image (txt2img) generation using the original architecture weights.
Lightweight Dependencies: No bloatware or heavy framework wrappers. Built on top of core packages (torch, safetensors, transformers) already shipped with standard ComfyUI setup.
Zero-Config Loading: The architecture is handled directly within the custom nodes backend, preventing configuration errors common with standard checkers.

📦 Installation

Support install in ComfyUI Manager or manual install for Windows and Linux

Navigate to your ComfyUI custom nodes directory inside your terminal or PowerShell.

Clone this repository:

git clone https://github.com/Winlensky/comfyui_compvis_ldm.git

Install requirements (Most likely not needed):

cd comfyui_compvis_ldm
pip install -r requirements.txt

Restart ComfyUI. The nodes will be registered automatically using your embedded Python environment.

🗂️ Model Downloads

Download Checkpoints: Latent Diffusion F8 Large on Hugging Face
Available Flavors:
- latent-diffusion-f8-large-fp32.safetensors (Original precision, ~6GB)
- latent-diffusion-f8-large-bf16.safetensors (Optimized precision, ~3GB)
- latent-diffusion-f8-large-fp16.safetensors (Optimized precision, ~3GB)
- latent-diffusion-f8-large-jack000-finetuned-fp16.safetensors (Finetuned version by jack000 for better structural composition)

📂 Where to place the files:

Place the downloaded .safetensors files strictly into your ComfyUI models directory under the dedicated LDM folder: ComfyUI\models\ldm

🛠️ Usage & Node Workflows

Resolution Notice: The LDM f8-large model was explicitly trained on downscaled, tight matrix resolutions. Ensure your target generation canvas is constrained (e.g., 256x256).

📜 Credits & License

Core architecture by CompVis (LMU Munich).
Finetuned variations by jack000.
This project is distributed under the MIT License.