Skip to content

digitalsamba/claude-code-video-toolkit / .claude/skills/qwen-edit

qwen-edit

Unverified

AI image editing prompting patterns for Qwen-Image (Qwen-Image-Edit-2511 by default, Qwen-Image-2.1 as a non-commercial Modal opt-in). Use when editing photos while preserving identity, reframing cropped images, changing clothing or accessories, adjusting poses, applying style transfers, or character transformations. Provides prompt patterns, parameter tuning, and examples.

FileKindSize
SKILL.mdskill md7.0 KB
examples.mdmarkdown6.6 KB
parameters.mdmarkdown4.4 KB
prompting.mdmarkdown2.9 KB

parameters.md

View on GitHub
# Qwen-Edit Parameters

> **Which model?** Modal runs **Qwen-Image-Edit-2511** unless the app was deployed with
> `IMAGE_EDIT_MODEL=qwen-image-2.1`; `--cloud runpod` runs **Qwen-Image-Edit-2511 +
> Lightning LoRA**. The tables below the first one were measured on 2511. See
> "Qwen-Image-2.1 (Modal opt-in)" at the end for how 2.1 differs.

## Available Parameters

| Parameter | CLI Flag | Default | Range | Notes |
|-----------|----------|---------|-------|-------|
| `num_inference_steps` | `--steps` | 8 (Qwen-Image-Edit-2511, Modal and RunPod); 40 only on a Modal app deployed with `IMAGE_EDIT_MODEL=qwen-image-2.1` | 4-50 | More = higher quality, slower |
| `guidance_scale` | `--guidance` | 1.0 | 1.0-7.0 | Higher = follows prompt more strictly. On 2.1 this is true CFG, off at 1.0 |
| `negative_prompt` | `--negative` | "" | text | Things to avoid |
| `seed` | `--seed` | random | int | For reproducibility |

## Steps (`--steps`)

Controls number of denoising passes. (2511 / RunPod figures -- 2.1 defaults to 40.)

| Value | Use Case | Inference Time |
|-------|----------|----------------|
| 4 | Quick preview, Lightning LoRA | ~10s |
| 8 | Default, good balance | ~18s |
| 16 | Higher quality, finer detail | ~34s |
| 25+ | Diminishing returns | 45s+ |

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

## Guidance Scale (`--guidance`)

Controls how strictly the model follows your prompt.

| Value | Behavior |
|-------|----------|
| 1.0 | Default, more creative freedom |
| 3.0 | Stronger adherence, more dramatic |
| 4-5 | Recommended max for Qwen-Edit |
| 7.0+ | Risk of artifacts, over-saturation |

**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.

```bash
# First attempt
uv run tools/image_edit.py --input photo.jpg --prompt "..." --seed 12345

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

## Cost vs Quality Tradeoffs

| Profile | Steps | Guidance | Time | Cost |
|---------|-------|----------|------|------|
| Quick preview | 4 | 1.0 | ~10s | ~$0.001 |
| Default | 8 | 1.0 | ~18s | ~$0.002 |
| Quality | 16 | 1.0 | ~34s | ~$0.004 |
| Dramatic | 8 | 3.0 | ~18s | ~$0.002 |
| Best | 16 | 3.0 | ~34s | ~$0.004 |

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

## Experimental Parameters

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

| Parameter | Default | Notes |
|-----------|---------|-------|
| `use_fp8` | true | FP8 quantization (false = BF16 full quality) |
| `auto_resize` | true | Auto-resize input for optimal processing |
| `true_cfg_scale` | 4.0 | Internal CFG scale |

To expose these, edit `tools/image_edit.py`.

## Qwen-Image-2.1 (Modal opt-in)

| Parameter | Default | Notes |
|-----------|---------|-------|
| `--steps` | 40 | Model card default, used whenever `--steps` is not given. 2.1 is not step-distilled and there is no Lightning LoRA for it here, so low step counts degrade faster than on 2511 |
| `--guidance` | 1.0 | Maps to the pipeline's `true_cfg_scale`. 2.1 is meant to be sampled **without** guidance; `>1` enables true CFG (the endpoint supplies a blank negative prompt if `--negative` is missing) and doubles per-step cost |
| `--negative` | none | Only used when `--guidance > 1` |
| inputs | 1-10 | First is edited, the rest are references or masks |

Endpoint-only extras (not exposed in the CLI): `output_resolution` (default 1024,
output area ~= its square; the model is native up to 2K) and explicit `width`/`height`.

Measured on the Modal A100-80GB ($0.000694/s), 640x640 input -> 1024x1024 output:
27s inference at 40 steps (~$0.02 warm). A cold call took 111s end to end (~85s of it
loading ~33GB of weights), plus the 60s scaledown window -- ~$0.12 for a one-off edit.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexUnverifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall