ComfyUI Qwen-Image DFloat11 Nodes

ComfyUI custom nodes for the DFloat11 compressed Qwen-Image model. Generate high-quality images with 32% less memory through lossless DFloat11 compression — bit-identical outputs to the original model.

Nodes

Node	Description
DFloat11 Qwen-Image Loader	Load the compressed model with optional CPU offloading
Qwen-Image Text Encode	Process text prompts with optional quality-boosting magic prompt
Qwen-Image Sampler	Generate images with full parameter control (steps, CFG, size, batch)
Qwen-Image Aspect Ratio	Quick preset aspect ratios (1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9)
Qwen-Image Preset (All-in-One)	Single-node generation with quality presets (draft/fast/balanced/quality/max)

Installation

1. Install Custom Nodes

Via ComfyUI Manager (recommended): Search for "Qwen-Image-DF11" in ComfyUI Manager and install.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/verIdyia/ComfyUI-Qwen-Image-DF11.git
cd ComfyUI-Qwen-Image-DF11
pip install -r requirements.txt

2. Download Models

Set up the model directory structure in ComfyUI/models/diffusion_models/:

Qwen-Image/
├── model_index.json              # From Qwen/Qwen-Image
├── transformer/
│   ├── config.json               # From Qwen/Qwen-Image
│   └── diffusion_pytorch_model.safetensors  # From DFloat11/Qwen-Image-DF11
├── text_encoder/                 # From Qwen/Qwen-Image
├── vae/                          # From Qwen/Qwen-Image
├── tokenizer/                    # From Qwen/Qwen-Image
└── scheduler/                    # From Qwen/Qwen-Image

Step 1 — Download the base model:

cd ComfyUI/models/diffusion_models
git lfs clone https://huggingface.co/Qwen/Qwen-Image

Step 2 — Replace the transformer with the DFloat11 version:

cd Qwen-Image/transformer
mv diffusion_pytorch_model.safetensors diffusion_pytorch_model.safetensors.backup
wget https://huggingface.co/DFloat11/Qwen-Image-DF11/resolve/main/diffusion_pytorch_model.safetensors

Step 3 — Merge config files (required for DFloat11 loading):

import json
from huggingface_hub import hf_hub_download

# Download DFloat11 config
df11_config_path = hf_hub_download("DFloat11/Qwen-Image-DF11", "config.json")
with open(df11_config_path) as f:
    df11_config = json.load(f)

# Merge with existing Qwen transformer config
config_path = "ComfyUI/models/diffusion_models/Qwen-Image/transformer/config.json"
with open(config_path) as f:
    qwen_config = json.load(f)

merged = {**qwen_config, **df11_config}
with open(config_path, "w") as f:
    json.dump(merged, f, indent=2)

3. Restart ComfyUI

Memory Requirements

VRAM (GPU)

Mode	VRAM Required	Peak VRAM
CPU Offload ON	16 GB+	~16.7 GB
CPU Offload OFF	32 GB+	~29.7 GB

System RAM

	Minimum	Recommended
RAM	96 GB	128 GB

Important: During model loading, the system needs ~68 GB of RAM simultaneously (41 GB model shell + 27 GB compressed weights). 64 GB is insufficient and will crash on Windows. 96 GB works with headroom; 128 GB is ideal. Linux with 64 GB may work due to mmap overcommit, but is not guaranteed.

• Qwen-Image: 20.4B parameter model (bf16 = 41 GB)
• DFloat11 compressed transformer: 28.42 GB (32% smaller than original)
• Text encoder (Qwen2.5-VL-7B): ~16 GB additional
• Lossless compression — produces bit-identical outputs

Preset Quality Modes

Preset	Steps	CFG	Use Case
draft	12	3.0	Quick previews
fast	20	3.5	Rapid iteration
balanced	50	4.0	General use
quality	80	4.5	High quality
max_quality	100	5.0	Best possible output

Requirements

• System RAM: 64 GB minimum, 128 GB recommended
• GPU: NVIDIA with 16+ GB VRAM (32+ GB without CPU offload)
• CUDA: 12.1+
• Python: 3.10+
• Disk: ~43 GB for model files
• diffusers >= 0.35.0, transformers >= 4.51.3, dfloat11[cuda12]

License

Apache License 2.0 — see LICENSE.

• Qwen-Image — Apache 2.0
• DFloat11/Qwen-Image-DF11 — See model license