Fred Bliss
4a091e9298
- Fixed generateTokens to properly handle different region types (bounding_box, polygon, object_reference) - Added checkbox UI to control includeObjectRef for boxes/polygons - Object references now correctly handle 2 coordinates instead of expecting 4 - Updated Python code to respect includeObjectRef flag from JavaScript - Fixed coordinate format issues (x1,y1,x2,y2 for boxes, x,y pairs for polygons) - Restored updateBasePromptWithTokens function for editable prompts - Default behavior keeps object_ref (checkbox checked) for backward compatibility - Without object_ref, generates pure spatial tokens like <|box_start|>0.1,0.2,0.3,0.4<|box_end|>
Qwen-Image-Edit Workflow Examples
This directory contains example workflows demonstrating the Qwen-Image-Edit nodes.
🔴 CRITICAL: Understanding Latent Input Choice
The Most Important Decision in Image Editing
When doing image editing, you have TWO fundamentally different approaches based on what you connect to KSampler's latent_image input:
Option 1: VAE Encode Input Image → KSampler (Structure Preservation)
LoadImage → VAEEncode → KSampler.latent_image
- Purpose: Keep composition, change details
- Denoise: 0.3-0.7 (lower = more preservation)
- Template: Use
minimal_edit - Example: "Change car color to blue" - keeps exact car shape/position
- When to use: Small edits, color changes, style tweaks
Option 2: Empty Latent → KSampler (Vision-Guided Generation)
EmptyLatentImage → KSampler.latent_image
- Purpose: Complete transformation using vision understanding
- Denoise: 0.9-1.0 (must be high - generating from scratch!)
- Template: Use
creativeordefault - Example: "Transform into cyberpunk style" - full reimagining
- When to use: Major transformations, style changes, creative edits
KEY INSIGHT: The reference latents (passed via VAE in QwenVLTextEncoder) provide vision guidance through conditioning REGARDLESS of which method you choose. They don't provide structure - the latent input does!
Workflow Categories
0. Enhanced Workflow (NEW!)
qwen_enhanced_workflow.json- RECOMMENDED STARTING POINT- All fixes from DiffSynth-Studio/Engine
- RoPE batch fix applied
- Template styles (creative/minimal_edit/photorealistic)
- Optimal resolution support
- Lightning LoRA 4-step generation
- Clear notes on VAE Encode vs Empty Latent choice
1. Basic Workflows
qwen_t2i_basic.json- Simple text-to-image generationqwen_i2i_edit_basic.json- Basic image editing with vision tokensqwen_simple_node.json- Using the simplified all-in-one node
2. Advanced Editing
qwen_autoregressive_edit.json- Sequential editing with contextqwen_entity_control.json- Regional generation with EliGenqwen_prompt_interpolation.json- Semantic blending between prompts
3. Sampler Integration
qwen_sampler_wrapper.json- Using proper reference latentsqwen_sampler_advanced.json- Advanced sampling with entity masksqwen_iterative_sampler.json- Progressive refinement
4. Helper Workflows
qwen_resolution_helper.json- Finding optimal dimensionsqwen_vae_normalizer.json- Applying proper VAE normalizationqwen_model_validator.json- Validating model configuration
5. Experimental
qwen_wan_bridge.json- Attempting Qwen→WAN (experimental)qwen_controlnet.json- BlockWise ControlNet integration
Node Descriptions
Core Nodes
QwenImageEditTextEncoder
- Replaces ComfyUI's broken Qwen text encoder
- Handles vision tokens properly
- Supports T2I, Edit, and Autoregressive modes
- Template dropping for better quality
QwenImageEditUnified
- All-in-one workflow node
- Combines all features in single interface
- Auto-detects mode based on inputs
- Includes ControlNet support
QwenImageEditSimple
- Simplified version for basic use
- Auto mode detection
- Minimal inputs required
Feature Nodes
QwenEliGenController
- Entity-level generation control
- Up to 4 entities with masks
- Regional attention control
QwenPromptInterpolator
- Learned semantic blending
- Smooth transitions between prompts
- Training capability for custom interpolations
QwenAutoregressiveEditor
- Sequential editing chains
- Context-aware generation
- Edit history tracking
Helper Nodes
QwenImageResolutionHelper
- Calculates optimal dimensions
- Based on target pixel count
- Ensures 32-pixel alignment
QwenVAENormalizer
- Applies official normalization
- 16-channel mean/std values
- Required for proper generation
QwenModelValidator
- Verifies token IDs
- Checks VAE configuration
- Validates model dimensions
Quick Start
- Load the Qwen-Image-Edit model
- Use
QwenImageEditSimplefor basic generation - For editing: provide an image and use edit mode
- For advanced features: use the unified node
Important Notes
- Always use 16-channel VAE (not SD VAE)
- Vision tokens only work in edit mode
- Autoregressive editing builds on previous edits
- Entity masks should match entity prompt count
- Low denoise (0.3-0.5) preserves structure better