Extended ComfyUI Plugin for DFloat11

Update: ComfyUI's recent dynamic VRAM feature causes problems with DFloat11 loading, so you will need to use the --disable-dynamic-vram flag when launching ComfyUI, in order to use DFloat11 models. I will investigate this issue, and hopefully I will be able to come up with a fix soon.

Update: Flux.2-klein is now supported, but there is currently an unexplainable bug causing outputs to be slightly different from BF16 output. I am currently looking into this issue, and in the meantime I have added a new decompressor node that fully decompresses the model weights, so it acts like a regular Load Diffusion Model node.

Currently supported model architectures:

• Flux.1-dev and schnell
• Flux.2-klein (currently buggy, will look into this)
• Chroma
• Chroma-Radiance
• SDXL (UNET component only, requires the --bf16-unet command-line flag)
• Cosmos Predict 2 Text2Image
• Lumina Image 2.0
• Z-Image
• Anima
• ACE-Step v1.5

Currently supported features:

• LoRA support for Chroma, Flux and Z Image
• Compression node for compressing BF16 models into DFloat11 format (for all supported model architectures)
• Cpu offloading options (available under the "DFloat11 Model Loader (Advanced)" node)
• Compatibility node for loading diffusers-native Flux DF11 models (using the "DFloat11 diffusers-native Model Loader" node)

Thanks to @tonyzhang617 for implementing the base DF11 compression and inference code. Unfortunately, it seems that the original developer is rather sporadic in his efforts to maintain the codebase and add features, so I decided to fork the repo and attempt to support it myself.

My fork adds support for more model types (compared to just Flux-dev based models in the original repo), and I plan to support more base architectures in future. Also, I have added a node ("DFloat11 diffusers-native Model Loader") for loading existing Flux DF11 models that are native to the diffusers library by converting them on-the-fly, which reduces the need to provide DF11 compression for both ComfyUI and diffusers separately with the same model. Sadly, it appears that my current runtime conversion code causes ComfyUI to use up slighly more VRAM as compared to using a ComfyUI-native DF11 model. In any case, VRAM consumption still stays below 24 GB (technically even the full Flux BF16 weights can also run with 24GB of VRAM, just that using literally 100% of the available VRAM affects system responsiveness). Another drawback of the on-the-fly conversion process is that there is a noticeable speed penalty (~1.30 it/s vs ~1.55 it/s).

Finally, the "DFloat11 Model Compressor" node allows users to generate their own ComfyUI-native DF11 compressions for Flux and Chroma models. The compression process uses only the CPU for compression, while the GPU is only used for verification, so less than 4GB of VRAM is used in the process. This means almost anyone should be able to create their own compressions, provided they have sufficient system RAM. Compressing a 12B model (i.e. Flux-based models) takes up almost 48 GB of RAM. Currently, the compression is single-threaded, but I guess waiting an hour or two for the compression process should be fine.

Check out my HuggingFace profile here: https://huggingface.co/mingyi456. I have uploaded a few DF11 models that are natively compatible with the original DF11 custom node, while the rest are diffusers-native and require my own added node to work with ComfyUI. Feel free to create an issue to request other models for compression as well (either for diffusers or ComfyUI), although models that use architectures I am unfamiliar with might be more difficult.

DFloat11 reduces model size by more than 30% while producing bit-for-bit identical outputs to the original. Unlike quantization techniques which trade quality for size, DFloat11 is a lossless compression method, preserving model output quality fully while supporting efficient inference.

---

Features

• 🚀 Fully Lossless – 100% identical outputs to the original model
• 📦 >30% smaller model size – lower VRAM requirements than the original model
• ⚡ Compatible with ComfyUI – drop-in support with custom nodes
• 🔧 GPU-accelerated inference – optimized for CUDA 12.1+

---

Installation

Requirements

• ComfyUI installed
• NVIDIA GPU with CUDA 12.1+

Setup

1. Install dependencies:

   pip install -r requirements.txt

2. Install the DFloat11 custom nodes in ComfyUI:

   cd <ComfyUI_installation_path>/custom_nodes
   git clone https://github.com/mingyi456/ComfyUI-DFloat11-Extended

---

Usage (Inference)

1. Once installed, the DFloat11 nodes show up under the DFloat11 folder in the Node Library.
2. Download a DFloat11 model for ComfyUI from Hugging Face and place it under <ComfyUI_installation_path>/models/diffusion_models.
3. Select a template DFloat11 workflow in the templates menu, under the ComfyUI-DFloat11-Extended section.
4. (Optional) Use the DFloat11 Model Loader node to load the model in *.safetensors format, which acts as a drop-in replacement for the Load Diffusion Model node.

Usage (Compression)

1. Ensure that your source model is in BFloat16 precision, and place it in the <ComfyUI_installation_path>/models/diffusion_models directory.
2. Create a "DFloat11 Model Compressor" node, and select the source model
3. Connect the node to a "Preview as Text" node.
4. Run the workflow, and wait for the process to complete.
5. In the <ComfyUI_installation_path>/models/diffusion_models directory, check that there is a new folder named after the source model, with a "DF11" suffix.
6. Move the model.safetensors file out of the folder, and rename it to any name you prefer.
7. Delete the folder, as well as the config.json file inside it. The config file is not required for the purposes of ComfyUI.

Note: If your source model is from produced from a "ModelSave" or "Save Checkpoint" node in ComfyUI, it might not be in BF16 precision. If so, restart ComfyUI with the --bf16-unet command line flag, and save the model again.

---

Resources

• 📖 DFloat11 Paper (arXiv)
• 🤗 DFloat11 ComfyUI Models on Hugging Face

---

Contributing

Contributions are welcome!

• Open an issue to request new model support
• Submit pull requests for bug fixes or improvements