fblissjr-ComfyUI-QwenImageWanBridge/SIMPLE_GUIDE.md

Simple Usage Guide - Qwen to WAN I2V Bridge

Production Nodes Summary

Main Nodes:

1. QwenWANUnifiedI2V - Complete I2V solution (USE THIS ONE!)
2. QwenWANUnifiedT2V - T2V/V2V with optional Qwen guidance

Helper Nodes: 3. QwenToImage - Converts Qwen latent to image 4. QwenWANI2VBridge - Basic I2V conditioning 5. QwenWANI2VDirect - Direct latent experiments

Basic Setup (90% of use cases)

[Load Image]
    ↓
[Qwen2VLFlux Encode]
    ↓
[Text Encode] → positive/negative
    ↓
[QwenWANUnifiedI2V]
    Connect:
    - qwen_latent (from Qwen encoder)
    - positive (from text encode)
    - negative (from text encode)

    Settings:
    - i2v_mode: "standard"
    - noise_mode: "no_noise"
    - width: 832, height: 480
    - num_frames: 81
    ↓
[KSampler]
    Connect ALL outputs:
    - positive (from unified node)
    - negative (from unified node)
    - latent_image (from unified node)
    - model: Your WAN model
    ↓
[VAE Decode]
    ↓
[Save Video]

Key Settings Explained

i2v_mode (What strategy to use)

• "standard" - Best for most cases, proper I2V conditioning
• "direct_latent" - Experimental, injects Qwen directly
• "reference" - Uses Qwen as reference/guide only
• "hybrid" - Blends approaches
• "vace_style" - Keyframe-based conditioning

noise_mode (How much variation)

• "no_noise" - Exact Qwen latent (highest fidelity)
• "add_noise" - Adds variation on top
• "mix_noise" - Blends with noise
• "scaled_noise" - Increases over time
• "decay_noise" - Decreases over time

Other Important Settings

• noise_strength: 0.0-0.2 for subtle, 0.3-0.5 for moderate
• start_frames: How many frames to condition (1-3 typical)
• frame_blend: How much influence decays (1.0 = full)
• wan_version: "auto" detects, or force "wan21"/"wan22"

Common Scenarios

"I want the video to look exactly like my input image"

i2v_mode: "standard"
noise_mode: "no_noise"
start_frames: 3
frame_blend: 1.0

"I want creative variation based on my image"

i2v_mode: "hybrid"
noise_mode: "mix_noise"
noise_strength: 0.3
frame_blend: 0.7

"I want to use multiple keyframes"

i2v_mode: "vace_style"
start_frames: 4 (number of keyframes)
noise_strength: 0.2

"I'm getting pixelated output"

wan_version: "wan21" (force WAN 2.1)

Input Options

The node accepts multiple input types:

• qwen_latent: From Qwen2VLFlux Encode (recommended)
• bridge_latent: From our other bridge nodes
• start_image + vae: Direct image input

Core Concept Reminder

This bridge attempts to use Qwen-Image's latent representation directly with WAN video generation, avoiding the quality loss from VAE decode/encode cycles.

Why it's tricky: Despite both being from Alibaba and WAN being trained on Qwen embeddings, the latent spaces aren't perfectly aligned. Small differences cause quality issues.

Best results:

• Use WAN 2.1 (16 channels match)
• Start with "standard" mode
• Adjust noise for variation
• Test different settings

Troubleshooting

Nothing connects?

• Restart ComfyUI after installing

Error about channels?

• Set wan_version to "wan21"

Too much/little motion?

• Adjust noise_strength

Doesn't look like input?

• Reduce noise_strength
• Increase start_frames
• Try "standard" mode