fblissjr-ComfyUI-QwenImageWanBridge/Documentation/SIMPLE_DECISION_GUIDE.md

Qwen Image Edit - Simple Decision Guide

Quick Setup Checklist

For ANY Qwen Image Workflow:

1. Load Qwen model → QwenVLCLIPLoader
2. Connect VAE to QwenVLTextEncoder (ALWAYS!)
3. Choose mode: text_to_image or image_edit
4. Pick your latent input method (see below)

---

THE MOST IMPORTANT DECISION

What connects to KSampler's latent_image input?

┌─────────────────────────────────────────────────────────┐
│  Want to KEEP the original structure?                    │
│  (Change colors, minor edits, style tweaks)              │
└─────────────────────────────────────────────────────────┘
                          ↓ YES
         LoadImage → VAEEncode → KSampler.latent_image
         Denoise: 0.3-0.7 (lower = more preservation)
         Template: "minimal_edit"


┌─────────────────────────────────────────────────────────┐
│  Want to TRANSFORM completely?                           │
│  (New style, major changes, creative freedom)            │
└─────────────────────────────────────────────────────────┘
                          ↓ YES
         EmptyLatentImage → KSampler.latent_image
         Denoise: 0.9-1.0 (MUST be high!)
         Template: "creative" or "default"

---

Parameter Quick Reference

QwenVLTextEncoder Settings

Parameter	What to Choose	Why
mode	image_edit if you have an input image	Tells model what to expect
template_style
→ For small changes	minimal_edit	Preserves most of original
→ For creative freedom	creative	Allows artistic interpretation
→ For realistic output	photorealistic	Maintains real-world logic
token_removal	Keep at auto	Unless you're debugging
optimize_resolution	Turn ON	Better quality, uses Qwen resolutions
apply_template	Keep ON	Unless you're an expert
debug_mode	Turn ON if confused	Shows what's happening

Denoise Settings

With VAE Encode (structure preservation):

• 0.3 = Tiny changes only
• 0.5 = Moderate changes
• 0.7 = Significant changes

With Empty Latent (full generation):

• Always use 0.9-1.0
• Lower values will look terrible

---

Common Workflows

1. "Change the car color to blue"

Mode: image_edit
Latent: VAEEncode → KSampler
Denoise: 0.4

2. "Make it cyberpunk style"

Mode: image_edit
Template: creative
Latent: EmptyLatent → KSampler
Denoise: 1.0

3. "Fix the lighting"

Mode: image_edit
Latent: VAEEncode → KSampler
Denoise: 0.5

4. "Generate a landscape"

Mode: text_to_image
Latent: EmptyLatent → KSampler
Denoise: 1.0

---

Still Confused?

Turn on debug_mode and check console for:

• Image dimensions at each step
• Token counts
• Reference latent status
• Template being applied

Remember:

• VAE must ALWAYS be connected to encoder for reference latents
• Reference latents ≠ Structure preservation
• Latent input choice determines structure vs. vision guidance
• Template style affects how much creative freedom

---

Advanced: Two-Stage Refinement

Want even better quality? Add QwenLowresFixNode:

KSampler → QwenLowresFixNode → VAEDecode
         ↑                    ↑
    (your output)     (refined output)

Settings:

• Upscale: 1.5x (default is good)
• Denoise: 0.5 (polish without changing)