Qwen Edit

Qwen-Image-Edit Skill

AI-powered image editing using Qwen-Image-Edit-2511 via RunPod serverless.

Status: Evolving - learnings being captured as we experiment

When to Use This Skill

Use when the user wants to:

• Edit/transform photos while preserving identity
• Reframe cropped images (fix cut-off heads, etc.)
• Change clothing, add accessories
• Change pose (arm positions, hand placement)
• Apply style transfers (cyberpunk, anime, oil painting)
• Adjust lighting/color grading
• Add/remove objects
• Character transformations (Bond, Neo, etc.)

When NOT to Use

• Background replacement (single image) - creates cut-out artifacts, halos
• Face swapping - cannot preserve identity from reference
• Outpainting - can't extend canvas reliably

Use With Care

• Multi-image compositing - CAN work with explicit identity anchors (see examples.md for prompt patterns). Requires describing distinctive features (hair texture/color, ethnicity, outfit) and using guidance ~2.0
• Camera angle changes - Inconsistent results. Vertical angles (low/high) work better than rotational (three-quarter view)

Quick Reference

# Basic edit
python tools/image_edit.py --input photo.jpg --prompt "Add sunglasses"

# With negative prompt (recommended)
python tools/image_edit.py --input photo.jpg \
  --prompt "Reframe as portrait with full head visible" \
  --negative "blur, distortion, artifacts"

# Style transfer
python tools/image_edit.py --input photo.jpg --style cyberpunk

# Background (use cautiously - often fails)
python tools/image_edit.py --input photo.jpg --background office

# Higher quality
python tools/image_edit.py --input photo.jpg --prompt "..." --steps 16 --guidance 3.0

# Multi-image composite (identity-preserving)
python tools/image_edit.py --input person.jpg background.jpg \
  --prompt "The [ethnicity] [gender] with [hair description] from first image is now in [scene] from second image. Same [features], [outfit]." \
  --negative "different ethnicity, different hair color, different face shape, generic stock photo" \
  --steps 16 --guidance 2.0

Key Files

• prompting.md - Prompt patterns and structure
• examples.md - Good/bad examples from experiments
• parameters.md - Tuning steps, guidance, negative prompts

Tool Location

tools/image_edit.py - CLI wrapper for RunPod endpoint

Related Docs

• docs/qwen-edit-patterns.md - Character transformation patterns
• .ai_dev/qwen-edit-research.md - Research notes

Qwen-Edit Examples

Real results from experiments. This document will grow as we learn.

---

Reframing Cropped Images (2026-01-04)

Problem: Portrait photo with head cropped at top (missing hair/forehead)

What Failed

What Worked

Prompt:

Reframe this photo as a portrait composition with the woman's full head visible, positioned in the lower portion of the frame with sky above

Negative:

blur, blurry, ghostly, distortion, abstract, haze, fog, out of focus

Result: Clean portrait with full head, nice hair, clear face, professional look

Variations Tested

Conclusion: The negative prompt was the key differentiator. Parameter tuning gave variations but all were usable once negative prompt was added.

---

Character Transformations (2026-01-01)

From docs/qwen-edit-patterns.md:

Key insight: Dark scenes lose identity. Strong visual references (iconic costumes) work best.

---

Background Replacement (2026-01-03)

Status: Generally fails with single-image approach - use multi-image compositing instead

Update (2026-01-04): Multi-image compositing CAN work - see "Identity-Preserving Composites" below.

---

Identity-Preserving Composites (2026-01-04)

Problem: Place a person into a new corporate/office scene while preserving their identity

What Failed

What Worked

Prompt:

The Irish woman with shoulder-length wavy dark auburn hair from first image is now standing confidently in the open plan office from second image. Same warm friendly smile, same facial features, black blazer, professional.

Negative:

different ethnicity, different hair color, straight hair, different face shape, generic stock photo

Settings: steps=16, guidance=2.0

Result: Good identity preservation - hair, smile, pose, outfit all recognizable. Open plan office background integrated naturally.

Prompt Pattern for Composites

The [ethnicity/nationality] [gender] with [specific hair: length, texture, color] from first image
is now [pose/action] in the [scene type] from second image.
Same [distinctive feature 1], same [distinctive feature 2], [outfit description].

Key Insights

1. Explicit identity anchors - "Irish woman with shoulder-length wavy dark auburn hair" beats vague "same woman" 2. Describe distinctive features - Hair texture, color, specific outfit items help anchor identity 3. Negative prompts for drift - "different ethnicity, different hair color, straight hair" prevents common failures 4. Mid guidance (2.0) - Sweet spot for composites. Too low = artifacts, too high = identity loss 5. Pose consistency - Matching the original pose (e.g., arms crossed) preserves identity better than new poses 6. Avoid mood words - "Serene", "wellness", "coach" trigger generic stock-photo aesthetics

Guidance Scale for Composites

---

Pose and Camera Angle Changes (2026-01-04)

Problem: Change a person's pose or camera angle while preserving identity and outfit

Pose Changes - What Worked

Key: Use negative prompt to exclude original pose: "arms crossed, defensive pose"

Camera Angles - Mixed Results

Settings

• steps: 14
• guidance: 2.0 (same as composites)

Key Insights

1. Pose changes work - Arm positions, hand placement can be modified successfully 2. Camera angles are inconsistent - Vertical angle changes (low/high) work better than rotation (three-quarter) 3. Be explicit about what to remove - Negative prompt should exclude the original pose 4. Identity preserved well - Face and outfit remain consistent through pose changes

Prompt Pattern for Pose Changes

Same woman, same face, same [outfit description], [new pose description], [mood/context]

Negative:

different person, different face, [original pose to remove]

---

Template for Adding New Examples

## [Use Case Name] (YYYY-MM-DD)

**Problem:** [What were you trying to achieve]

### What Failed
| Approach | Result |
|----------|--------|
| ... | ... |

### What Worked
**Prompt:** ...
**Negative:** ...
**Settings:** steps=X, guidance=Y
**Result:** [Description]

### Key Insight
[What we learned]

Qwen-Edit Parameters

Available Parameters

Steps (--steps)

Controls number of denoising passes.

Recommendation: Start with 8. Use 16 for final renders if detail matters.

Guidance Scale (--guidance)

Controls how strictly the model follows your prompt.

Note: Qwen-Edit uses lower guidance than typical diffusion models. Don't go above 5-6.

Recommendation: Start with 1.0. Try 3.0 for more literal interpretation.

Negative Prompt (--negative)

Critical for avoiding artifacts. Always consider using one.

Common Negative Prompts

For portraits/reframing:

blur, blurry, ghostly, distortion, abstract, haze, fog, out of focus

For zoom/scale changes:

black borders, letterbox, vignette, dark edges, frames, cropped

For viewpoint changes:

silhouette, dark figure, no face, shadow, faceless, artifacts, distortion, black

General quality:

deformed, disfigured, bad anatomy, extra limbs, blurry, low quality

Seed (--seed)

Use for reproducibility when iterating on prompts.

# First attempt
python tools/image_edit.py --input photo.jpg --prompt "..." --seed 12345

# Same seed, different prompt - compare results
python tools/image_edit.py --input photo.jpg --prompt "..." --seed 12345

Cost vs Quality Tradeoffs

(Costs based on L4 GPU @ $0.34/hr)

Experimental Parameters

These are supported by the handler but not yet exposed in the CLI:

To expose these, edit tools/image_edit.py.

Qwen-Edit Prompting Guide

Prompt Structure

Optimal length: 50-200 characters (short and specific beats long and detailed)

Formula: [Action], [details], [constraints]

Always Include Constraints

Preservation constraints significantly improve results:

• "Keep face unchanged"
• "Maintain original pose"
• "Preserve facial features"
• "Keep expression the same"
• "Preserve the person's identity"

Prompt Patterns

Reframing (Fix Cropped Images)

When head/body is cropped out of frame:

Reframe this photo as a portrait composition with the woman's full head visible, positioned in the lower portion of the frame with sky above

Critical: Pair with negative prompt to avoid artifacts:

--negative "blur, blurry, ghostly, distortion, abstract, haze, fog, out of focus"

What doesn't work for cropped images:

• "Zoom out" - adds letterboxing/black borders
• "Viewpoint change" - destroys identity, creates silhouettes
• "Extend the image" - not an outpainting model

Character Transformations

"{Character name}, wearing {costume details}, {pose/action}, {background}, {lighting style}, {aesthetic keywords}"

Examples:

• "James Bond 007, wearing a tuxedo with bow tie, holding a martini glass, casino background, suave and sophisticated, cinematic lighting, spy thriller aesthetic"
• "Neo from The Matrix, wearing long black leather trench coat, dark sunglasses, green digital rain code background, cyberpunk aesthetic"

Style Transfer

"Apply {style} style" or "Restyle as {style}"

Examples:

• "Apply Monet impressionist style, like Water Lilies"
• "Restyle as cyberpunk with neon lighting"
• "Make it look like a vintage 1970s photograph"

Clothing/Accessory Changes

"Change {item} to {new item}, maintain pose"

Examples:

• "Change blue t-shirt to red hoodie, maintain pose"
• "Add sunglasses and a warm smile"
• "Replace jacket with leather bomber jacket"

Lighting/Color

"{Lighting effect}, {mood keywords}"

Examples:

• "Add warm sunset tones, golden hour lighting"
• "Cinematic dramatic lighting with shadows"
• "Enhance lighting, studio portrait quality"

What Makes Prompts Fail

Negative Prompts

Always consider what to avoid: