qwen

ComfyUI Qwen-Image-Edit Nodes

Custom nodes for Qwen-Image-Edit with multi-image support, more flexibility around the vision transformer (qwen2.5-vl), custom system prompts, and some other experimental things to come.

BREAKING CHANGE (v2.7.0+)

Existing workflows will break. You must:

1. Delete and re-add, or Right Click -> Recreate Template Builder, Encoder, and Encoder Advanced nodes from your workflow
2. Connect: Template Builder template_output → Encoder template_output (single connection only)

Old multi-connection system (mode + system_prompt) no longer works as it was getting convoluted and confusing. One connection from template builder now handles everything (prompt, mode, system_prompt).

Updated Workflows Available

1. Multi image edit workflow
2. Single image edit workflow

---

Required Connections

When using Template Builder:

• Connect: Template Builder template_output → Encoder template_output
• That's it. One connection handles everything (prompt, mode, system_prompt)

Without Template Builder:

• Use encoder's dropdown for mode selection
• Type text and system_prompt manually

---

Documentation:

• CHANGELOG.md - Full changelog history
• nodes/docs/ - Detailed node documentation
• nodes/docs/resolution_tradeoffs.md - Resolution and scaling guide
• nodes/docs/QwenImageBatch.md - Batch node documentation

Features

Core Capabilities

• Qwen-Image-Edit-2509: Multi-image editing (1-3 optimal, up to 10 max because I had to pick something)
• QwenImageBatch: Smart batching with auto-detection, aspect ratio preservation, scaling, batching strategy
• Resolution Control & Power User Mode: Per-image resolution control
• Template Builder Auto-Sync: Automatic mode matching between template and encoder
• System Prompt Control: Customizable system prompts via the template builder
• Automatic Double-Scaling Prevention: Batch node and better encoder intelligence
• Full Debug Output: Complete prompts, character counts, aspect ratio tracking

Still Experimental or Not Working Well (or at all)

• Mask-Based Inpainting: Selective editing with diffusers blending, including Eligen entity control
• All wrapper nodes, probably more stuff I'm omitting

v2.7.0 Highlights

File-Based Template System - Single source of truth

• 9 templates in nodes/templates/*.md files (was 22 hardcoded)
• YAML frontmatter for metadata (mode, vision, experimental flags)
• Easy to add/edit without code changes
• JavaScript UI auto-fills custom_system field for editing

Template Builder → Encoder Connection

• mode is now STRING input (accepts connections or manual typing)
• REQUIRED: Connect BOTH mode and system_prompt to encoder
• Missing either = broken vision token formatting

QwenImageBatch - No more KJNodes dependency (docs)

• Auto-detects up to 10 images (no inputcount parameter)
• Skips empty inputs (no black images)
• max_dimensions strategy (minimal distortion) or first_image (hero-driven)
• Prevents double-scaling with metadata propagation
• See resolution_tradeoffs.md for detailed scaling guide

Nodes

QwenVLTextEncoder

Standard encoder with automatic labeling.

• Token dropping: 34 (text), 64 (image_edit)
• Multi-image support via Image Batch
• auto_label parameter for "Picture X:" control
• System prompt from Template Builder
• Full debug output with character counts
• verbose_log for console tracing of model passes

Resolution Scaling Modes:

• preserve_resolution (default): Keeps original size with 32px alignment
  • Best quality, no zoom-out effect
  • Recommended for typical images (512px-2048px)
  • May use more VRAM on very large images
• max_dimension_1024: Scales largest side to 1024px
  • Reduces VRAM on large images
  • Good for 4K images or VRAM constraints
  • Some zoom-out on large images
• area_1024: Scales to ~1024×1024 area (legacy)
  • Consistent output size but poor scaling behavior
  • Not recommended for general use

QwenVLTextEncoderAdvanced

Power user encoder with resolution control.

• All standard features plus:
• scaling_mode - Same three modes as standard encoder (preserve/max_dimension/area)
• resolution_mode - Applies weights to scaling_mode base (balanced/hero_first/hero_last/progressive)
• Per-image resolution weighting
• Memory budget management (max_memory_mb)
• Hero/reference modes for importance
• Custom resolution targets (vision & VAE separate)
• Choice of "Picture" vs "Image" labels
• Simplified interface (no validation_mode)
• verbose_log for console tracing of model passes

QwenTemplateBuilder

File-based system prompt templates (9 templates).

• Templates loaded from nodes/templates/*.md files
• YAML frontmatter for metadata
• JavaScript UI auto-fills custom_system field
• REQUIRED: Connect BOTH mode and system_prompt to encoder
• Available templates: default_t2i, default_edit, multi_image_edit, artistic, photorealistic, minimal_edit, technical, inpainting, raw

Helper Nodes

• QwenVLEmptyLatent: Creates 16-channel empty latents
• QwenVLImageToLatent: Converts images to 16-channel latents
• QwenTemplateConnector: Connects template builder to encoder (optional)
• QwenDebugController: Comprehensive debugging and profiling system

Inpainting Nodes

• QwenMaskProcessor: Mask preprocessing with blur, expand, feather controls
• QwenInpaintSampler: Diffusers-pattern inpainting with strength control

Experimental Nodes (Available but Low Priority)

• QwenSmartCrop: Automated face isolation with VLM detection (via Qwen3-VL) - docs
• QwenSpatialTokenGenerator: Visual editor for spatial tokens that don't seem to do much of anything right now
• QwenEliGenEntityControl: Entity-level mask control
• QwenEliGenMaskPainter: Simple mask creation
• QwenTokenDebugger: Debug token processing
• QwenTokenAnalyzer: Analyze tokenization

Workflows

Text-to-Image

QwenTemplateBuilder → QwenVLTextEncoder → KSampler
                  ↘ (system_prompt)

Single Image Edit

LoadImage → QwenVLTextEncoder (edit_image) → KSampler
QwenTemplateBuilder → QwenVLTextEncoder (system_prompt)

Multi-Image Edit (2509)

LoadImage (×N) → Image Batch → QwenVLTextEncoder → KSampler
QwenTemplateBuilder → QwenVLTextEncoder (system_prompt)

Use prompts like: "Combine the person from Picture 1 with the background from Picture 2"

Face Swap (Experimental)

LoadImage (portrait) → QwenSmartCrop (auto_fallback, face_headshot) → tight face crop
LoadImage (scene) ────────────────────────────────────┬────────────→ QwenImageBatch
                                                                          ↓
                                                               QwenVLTextEncoder (multi_image_edit)
                                                               "Replace face... while keeping pose..."
                                                                          ↓
                                                               QwenVLEmptyLatent → KSampler

See Prompt Cookbook for tested face swap prompts.

Example Workflows

Located in example_workflows/

Key Settings

Multi-Image Requirements

• Connect multiple images to Image Batch node (from KJNodes pack)
• Reference images with "Picture 1", "Picture 2" in prompts
• Images should ideally have similar dimensions for best results

Image Ordering

Images are numbered based on connection order to Image Batch:

• image_1 input → "Picture 1" in prompts
• image_2 input → "Picture 2" in prompts
• image_3 input → "Picture 3" in prompts

Example: "Take the old man from Picture 1 and place him in Picture 2"

See MULTI_IMAGE_ORDERING.md for detailed guide.

Template System

• IMPORTANT: Always connect BOTH outputs from Template Builder to encoder:
  • Template Builder mode → Encoder mode (ensures correct vision token formatting)
  • Template Builder system_prompt → Encoder system_prompt (provides instruction text)
• Template Builder prompt → Encoder text (your actual prompt text)
• Use custom_system field to edit any template's system prompt before using it
• JavaScript UI auto-fills custom_system when you select a preset

Why connect both?

• mode controls how vision tokens are formatted (labels, placement, token dropping)
• system_prompt provides the instruction text for the model
• Using mismatched mode/system_prompt causes incorrect token formatting
• Example: inpainting system prompt with image_edit mode = broken workflow

Troubleshooting

Subjects appearing zoomed out or too small?

Problem: Characters or objects in edited images appear smaller than expected.

Solution: Use preserve_resolution scaling mode (default in v2.6.1+)

• The old area_1024 mode aggressively downscaled large images
• Example: 1477×2056 was scaled to 864×1216 (0.58x), causing zoom-out
• New default preserves original size: 1477×2056 → 1472×2048 (1.00x)

Alternative: If hitting VRAM limits, try max_dimension_1024 mode instead

System prompt appearing in images?

• Verify Template Builder system_prompt output connects to Encoder system_prompt input
• Check workflow shows proper connection (should be Link 22 in example workflows)
• Enable debug_mode to see what's being tokenized

Multi-image not working?

• Use "Picture 1", "Picture 2" references in prompts
• Verify Image Batch node connection to edit_image input
• Enable debug_mode to see Picture formatting

Dimension mismatch errors?

• Ensure all images have similar dimensions before batching
• Use ComfyUI's Image Scale node if needed to match dimensions

Out of memory (OOM) errors with large images?

• Switch to max_dimension_1024 scaling mode
• This reduces VRAM usage while maintaining reasonable quality
• For 4K images: max_dimension_1024 is recommended over preserve_resolution

Status

Working/Tested:

• Text-to-image generation
• Single and multi-image editing
• Resolution control
• Template system with custom overrides
• Dimension-aware processing

Experimental (Low Priority):

• Spatial token generation (QwenSpatialTokenGenerator - effectiveness unclear)
• Entity control nodes (QwenEliGenEntityControl - untested with current models)
• Token debugging tools (QwenTokenDebugger, QwenTokenAnalyzer)

Latest Updates:

• v2.6.1: Resolution scaling fix with three modes (preserve/max_dimension/area)
• v2.6: Mask-based inpainting system with QwenMaskProcessor and QwenInpaintSampler
• v2.5: Advanced encoder for power users with resolution control
• Configurable auto-labeling (can now be disabled)
• Memory optimization for limited VRAM
• Hero/reference image weighting

License

MIT - See LICENSE file for details.