ComfyUI Qwen-Image + DiffSynth (Enhanced README)

This README consolidates and updates all features in this project, including the latest DiffSynth integration, Distill behavior, VRAM strategies, and Blockwise ControlNet Inpaint support.

Highlights (What’s New)

• Inpaint ControlNet: Full support for DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Inpaint
• Smarter sampler rules
  • Distill models: never auto-increase steps due to enhanced_quality; cap steps at ≤15 only if user sets >15
  • Lightning LoRA: respect user steps (e.g., 4/8), enhanced_quality only boosts CFG a bit
• VRAM policy refined for Pipeline Loader
  • Without ControlNet: keep Text Encoder (TE) and VAE resident on GPU; only offload transformer (DiT)
  • With ControlNet: manage TE/VAE as before (offload pipe includes transformer+TE+VAE) to avoid 99% VRAM stalls
• Example workflow added: 09_qwen_controlnet_inpaint.json

Installation

1. Clone into ComfyUI/custom_nodes
   • git clone https://github.com/xuchenxu168/ComfyUI_Qwen-Image.git
2. Install dependencies
   • pip install -r requirements.txt
3. Install DiffSynth-Studio (required for DiffSynth nodes)
   • git clone https://github.com/modelscope/DiffSynth-Studio.git
   • cd DiffSynth-Studio && pip install -e .
4. (Optional) MMGP offload (if available on your platform). If not present, the plugin falls back to built‑in VRAM management.
5. Restart ComfyUI

Notes

• First run will auto-download missing model weights via ModelScope/model ids used by DiffSynth.
• A CUDA-capable GPU is recommended; bf16 is the default precision.

Nodes (DiffSynth section)

• QwenImageDiffSynthLoRALoader: Load one LoRA
• QwenImageDiffSynthLoRAMulti: Merge two LoRA inputs
• QwenImageDiffSynthControlNetLoader: Load Blockwise ControlNet
  • Types: canny, depth, pose, normal, seg, inpaint
  • You can choose a local .safetensors or just pick a type to use the official repo id
• QwenImageDiffSynthPipelineLoader: Main pipeline with VRAM optimization and LoRA/ControlNet integration
  • base_model: auto | Qwen-Image | Qwen-Image-EliGen | Qwen-Image-Distill-Full
  • torch_dtype: bfloat16 | float16 | float32 (bf16 recommended)
  • offload_to_cpu: true/false
  • vram_optimization: No_Optimization | HighRAM_HighVRAM | HighRAM_LowVRAM | LowRAM_HighVRAM | LowRAM_LowVRAM | VerylowRAM_LowVRAM
• QwenDiffSynthSetControlArgs: Create control_args from reference image (+ optional inpaint_mask)
• QwenImageDiffSynthControlNetInput: One-stop preprocessing for control images (canny/depth/pose/normal/seg)
• QwenImageDiffSynthAdvancedSampler: Advanced sampler with EliGen + Blockwise ControlNet support
• QwenImageDiffSynthMemoryManager / pass-through helpers

ControlNet Inpaint Usage

There are two recommended ways:

A) Minimal wiring (recommended)

• Use QwenDiffSynthSetControlArgs
  • ref_image: the reference image to inpaint around
  • inpaint_mask: white=areas to modify, black=areas to preserve
• In QwenImageDiffSynthControlNetLoader, set controlnet_type to inpaint
  • If you don’t provide a local file, the loader will add the official model id DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Inpaint
• Connect control_args to Advanced Sampler

B) Custom preprocessing path

• Build your own mask image and pass it through QwenDiffSynthSetControlArgs, then proceed as above.

Advanced Sampler will construct ControlNetInput(image=..., inpaint_mask=...) and handle size/mode normalization for the mask.

Sampler Behavior (Quality/Speed)

• Distill models
  • cfg_scale fixed ~1.0 recommended internally (handled automatically)
  • Steps: respect user setting; capped at 15 if a higher value is set; enhanced_quality does NOT force higher steps
• Lightning LoRA (detected by filename containing "lightning")
  • Respect user steps (e.g., 4/8). enhanced_quality only slightly increases CFG
• Original models
  • enhanced_quality may raise steps to a minimum of 25 and boost CFG; fast reduces both moderately

VRAM Optimization Policy

• Without ControlNet
  • TE + VAE stay on GPU; only transformer is offloaded/profiled
  • MMGP budgets around transformer=90%, workingVRAM≈70%
• With ControlNet
  • TE + VAE managed as original (included in offload/profile set when offload_to_cpu true)
  • MMGP budgets default back to transformer/text_encoder/vae and a more conservative workingVRAM≈25%
• If MMGP isn’t available, we automatically fall back to DiffSynth’s built‑in enable_vram_management()

Example Workflows

• example_workflows/09_qwen_controlnet_inpaint.json — Inpaint ControlNet end‑to‑end
• example_workflows/qwen_diffsynth_controlnet_workflow.json — Multi‑ControlNet and Distill pipeline example
• example_workflows/03_qwen_controlnet_canny.json — Canny control
• example_workflows/04_qwen_controlnet_depth.json — Depth control
• example_workflows/05_qwen_eligen_entities.json — EliGen entity prompts + masks

Tips

• If VRAM hits 99% and slows down while using ControlNet, keep offload_to_cpu=true and use the provided profiles; avoid pinning TE/VAE to GPU when ControlNet is active
• Inpaint masks: use binary masks where possible; white=edit area, black=keep
• For higher speed on Distill, consider 8–12 steps; for original models, enhanced_quality at ≥25 steps for best quality

Troubleshooting

• “Pipeline has no Blockwise ControlNet loaded; ignoring control_args”
  • Ensure you selected a ControlNet model or set controlnet_type in the loader
• Mask alignment issues
  • The sampler will auto-resize the mask to the control image size; provide reasonably aligned inputs for best results
• Extremely low VRAM usage without ControlNet
  • This is expected if offloading is enabled; raise working size (resolution) or steps slightly, or disable offloading if you have headroom

License

Apache 2.0. See LICENSE.

Acknowledgements

• DiffSynth-Studio by ModelScope
• Community repos and prior workflows that inspired this integration

ComfyUI Qwen-Image Plugin v2.1

🎨 A comprehensive ComfyUI plugin for Qwen-Image model integration using ComfyUI's standard separated model loading architecture. Features exceptional Chinese text rendering, advanced image generation capabilities, and the new Advanced Diffusion Loader with comprehensive optimization options.

🌟 Features

⚡ New in v2.1: Advanced Features

• 🚀 All-in-One Loading: Integrated UNet, CLIP, and VAE loading in a single node
• 🎯 Performance Optimization: Advanced weight and compute data type selection
• 🧠 SageAttention Support: Memory-efficient attention mechanisms
• ⚙️ cuBLAS Integration: Hardware-accelerated linear layer computations
• 💾 Memory Management: FP16 accumulation and auto-detection features
• 🔧 Advanced Configuration: Extra state dictionary support for power users
• 🎭 Dual CLIP Support: Load and use two CLIP models simultaneously for enhanced text understanding
• 🔀 Flexible Blending: Multiple modes to combine dual CLIP outputs (average, concat, etc.)

New Architecture (v2.0+)

• Separated Model Loading: Uses ComfyUI's standard UNet/CLIP/VAE loading system
• Standard Workflow Integration: Fully compatible with ComfyUI's native sampling and conditioning
• Optimized Performance: Better memory management and faster loading times
• Model Flexibility: Support for different precision formats (bf16, fp8, fp8_e4m3fn, fp8_e5m2)

Core Capabilities

• Advanced Image Generation: High-quality text-to-image generation with Qwen-Image model
• Exceptional Chinese Text Rendering: Industry-leading Chinese character and text rendering in images
• 🎯 DiffSynth Integration: ControlNet support, LoRA integration, and advanced memory management
• ControlNet Support: Canny, Depth, Pose, Normal, and Segmentation control
• LoRA Integration: Load and combine multiple LoRA models with weight control
• Lightning LoRA: Fast inference optimization for improved performance
• Intelligent Image Editing: Style transfer, object manipulation, and detail enhancement (legacy)
• Multi-modal Understanding: Image analysis, object detection, and content comprehension (legacy)
• Professional Text Rendering: High-quality text overlay with customizable fonts and styles (legacy)

Key Advantages

• Chinese Language Optimization: Specifically optimized for Chinese text processing and rendering
• Multiple Aspect Ratios: Support for various aspect ratios optimized for different use cases
• Flexible Configuration: Extensive customization options for generation parameters
• ComfyUI Integration: Seamless integration with ComfyUI's node-based workflow system
• High Performance: Optimized for both quality and speed

📦 Installation

Prerequisites

• ComfyUI installed and running
• Python 3.8 or higher
• CUDA-compatible GPU (recommended)
• At least 8GB VRAM for optimal performance

Required Models

Place these models in the corresponding ComfyUI directories:

Diffusion Models (ComfyUI/models/diffusion_models/):

• qwen_image_bf16.safetensors
• qwen_image_fp8_e4m3fn.safetensors

Text Encoders (ComfyUI/models/text_encoders/):

• qwen_2.5_vl_7b.safetensors
• qwen_2.5_vl_7b_fp8_scaled.safetensors

VAE (ComfyUI/models/vae/):

• qwen_image_vae.safetensors

ControlNet Models (ComfyUI/models/controlnet/) - For DiffSynth features:

• qwen_image_blockwise_controlnet_canny.safetensors
• qwen_image_blockwise_controlnet_depth.safetensors

LoRA Models (ComfyUI/models/loras/) - For DiffSynth features:

• qwen_image_distill.safetensors
• qwen_image_lightning.safetensors

Method 1: ComfyUI Manager (Recommended)

1. Open ComfyUI Manager
2. Search for "Qwen-Image"
3. Click Install
4. Restart ComfyUI

Method 2: Manual Installation

1. Navigate to your ComfyUI custom_nodes directory:

   cd ComfyUI/custom_nodes

2. Clone this repository:

   git clone https://github.com/your-repo/ComfyUI_Qwen-Image.git

3. Install dependencies:

   cd ComfyUI_Qwen-Image
   pip install -r requirements.txt

4. Restart ComfyUI

🚀 Quick Start

Advanced Workflow (v2.1 - Recommended)

1. Add 🎨 Qwen-Image Advanced Diffusion Loader node
2. Configure optimization settings (weight dtype, SageAttention, etc.)
3. Add 🎨 Qwen-Image Text Encode nodes for positive and negative prompts
4. Add 🎨 Qwen-Image Empty Latent node
5. Add 🎨 Qwen-Image Sampler node
6. Add VAE Decode and Save Image nodes
7. Connect the workflow and execute

Standard Workflow (Individual Loaders)

1. Add 🎨 Qwen-Image UNet Loader node
2. Add 🎨 Qwen-Image CLIP Loader node
3. Add 🎨 Qwen-Image VAE Loader node
4. Add 🎨 Qwen-Image Text Encode nodes for positive and negative prompts
5. Add 🎨 Qwen-Image Empty Latent node
6. Add 🎨 Qwen-Image Sampler node
7. Add VAE Decode and Save Image nodes
8. Connect the workflow and execute

Example Workflow Files

• qwen_diffsynth_controlnet_workflow.json - 使用 DiffSynth 管线，已支持 base_model（auto/Qwen-Image/Qwen-Image-EliGen/Qwen-Image-Distill-Full）
• 05_qwen_eligen_entities.json - EliGen 实体控制示例（将 base_model 切到 Qwen-Image-EliGen）
• dual_clip_workflow.json - Dual CLIP text encoding with enhanced understanding
• advanced_diffusion_loader_workflow.json - Advanced loader with optimizations
• qwen_image_standard_workflow.json - Basic text-to-image generation
• chinese_text_rendering_workflow.json - Optimized for Chinese calligraphy

Example Prompt (Chinese)

一只可爱的小猫咪坐在樱花树下，春天的阳光洒在它的毛发上，背景是传统的中式庭院

Example Prompt (English)

A beautiful landscape with Chinese text '你好世界' written in elegant calligraphy

📚 Node Reference

Model Loaders

🎨 Qwen-Image Advanced Diffusion Loader ⭐ NEW

All-in-one model loader with advanced optimization options.

• Inputs:
  • Model name, weight dtype, compute dtype
  • SageAttention mode, cuBLAS modifications
  • Auto-detection for CLIP and VAE
  • Extra state dictionary for advanced configs
• Outputs: MODEL, CLIP, VAE
• Features:
  • Memory optimization for various GPU tiers
  • Performance tuning options
  • Auto-detection of compatible models
• See: ADVANCED_DIFFUSION_LOADER_GUIDE.md for detailed usage

🎨 Qwen-Image UNet Loader

Loads the Qwen-Image diffusion model.

• Input: UNet model file (qwen_image_bf16.safetensors, etc.)
• Output: MODEL
• Weight Types: default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2

🎨 Qwen-Image CLIP Loader ⭐ Enhanced

Loads Qwen text encoder model(s) with dual CLIP support.

• Inputs:
  • Primary CLIP model file
  • load_dual_clip: Enable dual CLIP loading
  • Secondary CLIP model file (optional)
  • Device selection
• Outputs: Primary CLIP, Secondary CLIP
• Features: Single or dual CLIP loading based on user choice

🎨 Qwen-Image VAE Loader

Loads the Qwen VAE model.

• Input: VAE model file (qwen_image_vae.safetensors)
• Output: VAE

Text Processing

🎨 Qwen-Image Text Encode

Encodes text prompts with Chinese optimization.

• Inputs: CLIP, text
• Output: CONDITIONING
• Features: Magic prompt enhancement, language detection

🎨 Qwen-Image Dual Text Encode ⭐ NEW

Advanced text encoder supporting dual CLIP models.

• Inputs:
  • Primary CLIP, text prompt
  • Secondary CLIP (optional)
  • Blend mode selection
  • Magic prompt and language options
• Output: CONDITIONING
• Features:
  • Dual CLIP text encoding
  • Multiple blend modes (average, concat, primary_only, secondary_only)
  • Enhanced text understanding
• See: DUAL_CLIP_GUIDE.md for detailed usage

Latent Operations

🎨 Qwen-Image Empty Latent

Creates empty latent images with optimized dimensions.

• Inputs: width, height, aspect_ratio, batch_size
• Output: LATENT
• Aspect Ratios: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3

🎨 Qwen-Image Sampler

Samples images using ComfyUI's standard sampling system.

• Inputs: MODEL, positive, negative, latent_image
• Output: LATENT
• Parameters: seed, steps, cfg, sampler_name, scheduler, denoise

🎯 Optimized Aspect Ratios

Ratio	Dimensions	Use Case
1:1	1328×1328	Square images, social media
16:9	1664×928	Landscape, wallpapers
9:16	928×1664	Portrait, mobile screens
4:3	1472×1104	Traditional photos
3:4	1104×1472	Portrait photos
3:2	1584×1056	Photography standard
2:3	1056×1584	Book covers, posters

🔧 Advanced Configuration

Magic Prompts

The plugin automatically enhances prompts with quality improvements:

• Chinese: "超清，4K，电影级构图"
• English: "Ultra HD, 4K, cinematic composition."

Language Detection

Automatic language detection based on Unicode character ranges:

• Chinese characters (U+4E00-U+9FFF) trigger Chinese optimizations
• Other characters use English optimizations

🆕 Migration from v1.0

Breaking Changes

• Old QwenImageModelLoader and QwenImageGenerate nodes are deprecated
• New separated loading architecture required
• Workflow files need to be updated

Legacy Support

Legacy nodes are still available for backward compatibility:

• QwenImageEdit (Legacy)
• QwenImageTextRender (Legacy)
• QwenImageUnderstand (Legacy)

🐛 Troubleshooting

Common Issues

1. Model not found: Ensure models are in correct ComfyUI directories
2. CUDA out of memory: Use fp8 models or reduce batch size
3. Text encoding errors: Check CLIP model is loaded correctly

Performance Tips

• Use fp8 models for lower VRAM usage
• Enable magic prompts for better quality
• Use appropriate aspect ratios for your use case
• For DiffSynth features: Choose appropriate VRAM optimization strategy
• Use Lightning LoRA for faster inference when quality is acceptable

🎯 Advanced Features

DiffSynth Integration

For advanced ControlNet and LoRA features, see the DiffSynth Guide.

Key DiffSynth Features:

• ControlNet Support: Structure control with Canny, Depth, Pose, Normal, Segmentation
• LoRA Integration: Load and combine multiple LoRA models
• Memory Management: Advanced VRAM optimization strategies
• Lightning LoRA: Fast inference optimization

DiffSynth Nodes:

• QwenImageDiffSynthLoRALoader: Load LoRA models
• QwenImageDiffSynthControlNetLoader: Load ControlNet models
• QwenImageDiffSynthPipelineLoader: Main pipeline with memory management（新增 base_model：auto / Qwen-Image / Qwen-Image-EliGen / Qwen-Image-Distill-Full）
• QwenImageDiffSynthDistillPipelineLoader: 专用 Distill-Full 管线加载器（保持不变）
• QwenImageDiffSynthSampler: Generate images with ControlNet/LoRA
• QwenImageDiffSynthMemoryManager: Advanced memory optimization

📄 License

Apache 2.0 License - see LICENSE file for details.

🤝 Contributing

Contributions welcome! Please read our contributing guidelines and submit pull requests.

📞 Support

• GitHub Issues: Report bugs and feature requests
• Documentation: Check example workflows
• Community: Join ComfyUI Discord for discussions