ComfyUI-IPAdapter-WAN 2.0

This extension adapts the InstantX IP-Adapter for SD3.5-Large to work with Wan 2.1 and other UNet-based video/image models in ComfyUI.

Unlike the original SD3 version (which depends on joint_blocks from MMDiT), this version performs sampling-time identity conditioning by dynamically injecting into attention layers — making it compatible with models like Wan 2.1, AnimateDiff, and other non-SD3 pipelines.

---

🚀 Features

Core Functionality

• 🔁 Injects identity embeddings during sampling via attention block patching
• 🧠 Works with Wan 2.1 and other UNet-style models (no SD3/MMDiT required)
• 🛠️ Built on top of ComfyUI's IPAdapter framework
• 🎨 Enables consistent face/identity across frames in video workflows

Version 2.0 Upgrades

• ⏱️ Timestep-based Scheduling: Fine-grained control over when IP conditioning is active during denoising (start/end percentage)
• 🎯 Type-Safe Architecture: Extensive type hints, dataclasses, and runtime validation for better reliability
• 💾 Enhanced Memory Management: FP16 by default, explicit device placement, optimized batched processing
• 🔧 Structured Configuration: Centralized config system with IPAdapterConfig and VisionEncoderConfig
• 📊 Time-Conditioned Resampler: Advanced TimeResampler with Adaptive Layer Normalization (adaLN) for timestep-aware conditioning
• 🛡️ Fail-Fast Validation: Input validation at construction time to catch errors early
• 📚 Comprehensive Documentation: Detailed docstrings with computational complexity analysis
• 🎨 True Dynamic Resolution Support: Perceiver-style resampler architecture enables native support for variable-resolution inputs from SigLIP2 so400m naflex - no fixed sequence length assumptions

---

📦 Installation

1. Clone the repo into your ComfyUI custom nodes directory:

git clone https://github.com/your-username/ComfyUI-IPAdapter-WAN.git

2. Download the required model weights:

• Download the IP-Adapter weights:

  • ip-adapter.bin

  • Place it in: ComfyUI/models/ipadapter/

• Download the CLIP Vision model:

  • SigLIP2 so400m patch16 naflex (recommended): google/siglip2-so400m-patch16-naflex

  • Place the model file in: ComfyUI/models/clip_vision/

  • This model supports dynamic resolution with native flexible (naflex) attention, providing better performance for variable aspect ratios.

(Note: This node only supports SigLIP2 so400m patch16 naflex. The vision encoder type is automatically set to SigLIP2 - no configuration needed.)

Dynamic Resolution: The IPAdapter architecture fully supports variable-resolution inputs. The Perceiver-style resampler compresses any sequence length from SigLIP2 to a fixed number of queries, enabling true dynamic resolution support without hardcoded size assumptions.

---

🧠 How It Works

Wan models use a UNet structure instead of the DiT transformer blocks used in SD3. To make IPAdapter work with Wan:

• The extension scans all attention blocks (modules with .to_q and .to_k) dynamically.

• It injects IPAdapter's attention processors (IPAttnProcessor) directly into those blocks.

• Version 2.0 Enhancement: Identity embeddings are updated based on the current sampling timestep using a time-conditioned resampler (TimeResampler) with:

  • Adaptive Layer Normalization (adaLN) for timestep-aware feature modulation
  • Perceiver-style cross-attention for efficient compression of vision embeddings - this enables true dynamic resolution support by compressing variable-length sequences to fixed queries
  • Timestep embedding integration for denoising-aware conditioning

• Dynamic Resolution Support: The resampler architecture is specifically designed to handle variable-length vision encoder outputs (from SigLIP2 so400m naflex). It compresses any sequence length to a fixed number of queries (64), making it resolution-agnostic. This means:

  • ✅ Works with any input resolution/aspect ratio from SigLIP2
  • ✅ No hardcoded sequence length assumptions
  • ✅ Efficient O(n) scaling with variable token counts

• Timestep Scheduling: The extension respects start_percent and end_percent parameters, only applying IP conditioning during the specified denoising range.

This means it works without requiring joint_blocks or specific architectural assumptions — making it plug-and-play for many custom models.

---

🛠 Usage

1. In ComfyUI, use the following nodes:

   • Load IPAdapter WAN Model - Loads the IP-Adapter checkpoint with SigLIP2 so400m configuration

   • Apply IPAdapter WAN Model - Applies IP conditioning to your diffusion model

2. Connect the CLIP Vision embedding (from a face image) and your diffusion model to the adapter.

   • Note: This node automatically uses SigLIP2 so400m configuration - no vision encoder selection needed.

3. New in 2.0: Configure timestep scheduling:

   • start_percent (default: 0.0): When to start applying IP conditioning (0.0 = beginning of denoising)
   • end_percent (default: 1.0): When to stop applying IP conditioning (1.0 = end of denoising)
   • This allows fine-grained control over when identity conditioning is active during the denoising process

4. Use a weight of ~0.5 as a good starting point (adjustable from -1.0 to 5.0).

5. You can apply this in video workflows to maintain consistent identity across frames.

---

📁 Example Workflows

Example .json workflows will be available soon in the workflows/ folder.

---

✅ Compatibility

| Model | Status | | ---------------- | ------------------- | | Wan 2.1 | ✅ Works | | AnimateDiff | ✅ Works | | SD3 / SDXL | ❌ Use original repo | | Any UNet variant | ✅ Likely to work |

---

✨ Version 2.0 Technical Improvements

Architecture Enhancements

• Type-Safe Design: Protocol-based interfaces, dataclasses for configuration, comprehensive type hints
• Memory Optimization: FP16 by default, efficient batched processing, zero-copy state sharing
• Error Handling: Fail-fast validation with clear error messages
• Modular Configuration: Centralized config system enabling easy hyperparameter tuning

Performance Optimizations

• Efficient Attention: Uses F.scaled_dot_product_attention for hardware-accelerated attention
• Memory-Efficient Reshaping: Fused tensor operations to minimize memory copies
• Lazy Evaluation: IP conditioning computed only when active (based on timestep schedule)

Developer Experience

• Comprehensive Documentation: Detailed docstrings with computational complexity analysis
• Extensible Design: Easy to add new vision encoder types or modify architecture
• Runtime Validation: Input validation prevents common errors before execution

---

🔧 TODOs

• Allow multiple adapters without conflict

• Auto-detect model parameters (hidden size, num layers)

• Convert .bin to safetensors format

• Add more workflows for different models

• Support additional vision encoder types beyond SigLIP2

---

🧑‍💻 Credits

• Adapted from: InstantX IPAdapter for SD3.5

• Version 2.0 enhancements include:

  • Type-safe architecture with comprehensive validation
  • Time-conditioned resampler with adaLN
  • Timestep-based scheduling for fine-grained control
  • Enhanced memory management and performance optimizations

---

📝 Changelog

Version 2.0

• ✨ Added timestep-based scheduling (start_percent/end_percent)
• ✨ Implemented time-conditioned resampler with Adaptive Layer Normalization
• ✨ Enhanced type safety with dataclasses, enums, and protocols
• ✨ Improved memory management (FP16 default, optimized batching)
• ✨ Added comprehensive documentation and error handling
• ✨ Structured configuration system for easier customization

---

Feel free to contribute or suggest improvements via GitHub Issues or Pull Requests.