Back to Models

Qwen Image Edit

qwenqwen/qwen-image-edit

Alibaba Qwen team's image-edit model built on the 20B Qwen-Image. Excels at text rendering inside images and semantic + appearance edits.

Qwen Image Edit is Alibaba Qwen's image-to-image editing model derived from the 20B Qwen-Image foundation model. It is built for precise visual edits rather than open-ended chat: provide an input image plus a natural-language instruction, and the model returns a revised image. Its main differentiator is text-aware editing β€” especially Chinese and English text inside posters, signs, calligraphy, UI screenshots, and product graphics β€” where many image editors either blur glyphs or redraw surrounding layout unnecessarily.

Under the hood, Qwen describes a dual-control editing path: the input image is fed to Qwen2.5-VL for visual semantic control and to a VAE encoder for visual appearance control. That lets the model cover both semantic edits (rotate an object, restyle a portrait, preserve a character/IP identity while changing the scene) and appearance edits (add, remove, recolor, or replace a local object while leaving the rest of the image stable). Through TheRouter, it is routed as qwen/qwen-image-edit on the OpenAI-compatible image endpoint at a flat $0.048 per output image.

Best for
  • β€’ Text edits inside existing visuals β€” replace Chinese or English copy on posters, packaging, screenshots, signage, and calligraphy while preserving font feel, size, layout, and surrounding pixels.
  • β€’ Localized marketing asset iteration β€” adapt campaign creatives for Chinese/English markets without rebuilding the whole image in a design tool.
  • β€’ Product and scene retouching β€” add or remove a sign, recolor a specific object, change a background, or adjust clothing while preserving the unchanged regions.
  • β€’ Character/IP variation β€” rotate a mascot, create emoji-pack variants, or transfer a portrait into a new style while keeping semantic identity consistent.
Reach for something else if
  • β€’ Pure text-to-image generation β€” use qwen/qwen-image, Doubao Seedream, or GPT Image models when there is no source image to preserve.
  • β€’ High-volume multi-output batches β€” the base qwen-image-edit variant returns one image; use Plus/Max/2.0 variants where you need 1–6 outputs or custom resolution controls.
  • β€’ Deterministic pixel-perfect design QA β€” seed helps relative stability, but image editing remains probabilistic and can vary across calls.
  • β€’ Video, audio, or editable layered-file output β€” the model returns rendered PNG-style image results, not video clips, audio, PSD/Figma layers, masks, or vector files.
Context Length
--
Max Output
--
Image Priceper request
$0.0486/ image
Input Priceper request
$0.0432/ 1M tokens

Modalities

textimage→image

Capabilities

VisionImage GenerationImage Edit

Media Generation Capabilities

image_edit
sizes
  • 1328x1328
  • 1664x928
  • 1472x1140
  • 1140x1472
  • 928x1664
max_reference_images
1
defaults
size
1328x1328

Pricing Breakdown

TypeRate
Image$0.0486 / image
Input$0.0432 / 1M tokens
Per image$0.0486 / image

Per-image flat fee. Alibaba Model Studio international list price $0.045/image (alibabacloud.com/help/en/model-studio/model-pricing, read 2026-07-29).

Supported Parameters

promptsizeimageseedcfgnum_inference_steps

Specifications

Release date2025-08-18github.com β†—verified
Base modelQwen-Image, a 20B MMDiT image foundation modelqwenlm.github.io β†—verified
Editing architectureDual control: Qwen2.5-VL semantic representation plus VAE appearance/reconstruction representationarxiv.org β†—verified
Input modalitiesText instruction plus one or more images; TheRouter catalog exposes text+image inputwww.alibabacloud.com β†—verified
OutputOne edited PNG image for the base qwen-image-edit variantwww.alibabacloud.com β†—verified
Text editing languagesChinese and English text editing are explicitly highlighted by Qwenqwenlm.github.io β†—verified
Input image requirementsJPG/JPEG, PNG, BMP, TIFF, WEBP, GIF first frame; recommended width/height 384–3072 px; max 10 MB in Alibaba Model Studio docswww.alibabacloud.com β†—verified
TheRouter price$0.048 per output imagetherouter.ai β†—verified
Training cutoffNot publicly disclosedunknown

Benchmarks

BenchmarkDistributionScoreSource
Public image-editing benchmarks
The launch blog says Qwen-Image-Edit reaches state-of-the-art performance across multiple public image-editing benchmarks, but the page does not publish a numeric table. Treat this as a qualitative vendor claim unless you run your own eval set.
SOTA claimed by Qwen; numeric scores not disclosed in launch blogqwenlm.github.io β†—
Qwen-Image technical report
The arXiv report attributes editing consistency to multi-task TI2I/I2I training and the Qwen2.5-VL + VAE dual-encoding mechanism.
State-of-the-art editing performance reportedarxiv.org β†—

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

Recommended: use the async API

Image generation typically takes 30–180s, beyond the edge sync timeout. The examples below use the ?async=true submit + poll pattern. Read the full async image generation & edit guide β†’

cURL
# 1) Submit job (returns 202 immediately with a polling URL).
# Image generation takes 30-180s β€” always use the async path in production.
JOB=$(curl -s -X POST "https://api.therouter.ai/v1/images/generations?async=true"   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "qwen/qwen-image-edit",
    "prompt": "A cinematic product render with soft studio lighting"
  }' | python3 -c "import sys,json;print(json.load(sys.stdin)['id'])")
echo "submitted: $JOB"

# 2) Poll until terminal (succeeded / failed / cancelled / expired).
while :; do
  R=$(curl -s "https://api.therouter.ai/v1/jobs/$JOB"     -H "Authorization: Bearer $THE_ROUTER_API_KEY")
  S=$(echo "$R" | python3 -c "import sys,json;print(json.load(sys.stdin)['status'])")
  echo "status: $S"
  case "$S" in
    succeeded) echo "$R" | python3 -c "import sys,json;print(json.load(sys.stdin)['unsigned_urls'][0])"; break ;;
    failed|cancelled|expired) echo "$R"; exit 1 ;;
  esac
  sleep 5
done

API guide

Image editing

Send a source image and an edit instruction to TheRouter's OpenAI-compatible image endpoint. Use base64 for local files or a URL when your application already hosts the input asset.

cURL
curl https://api.therouter.ai/v1/images/generations \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen-image-edit",
    "prompt": "Replace the headline with: TheRouter AI Gateway. Keep the same font style, layout, and background.",
    "image": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUg...",
    "seed": 12345,
    "cfg": 4.0,
    "num_inference_steps": 50
  }'

Image Editing Examples

Upload an image and describe the edit you want with a text prompt; the model returns the edited image as base64.

cURL
# Same async submit + poll pattern as /v1/images/generations.
JOB=$(curl -s -X POST "https://api.therouter.ai/v1/images/edits?async=true"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -F "model=qwen/qwen-image-edit"   -F "prompt=Turn this scene into a watercolor painting"   -F "size=1024x1024"   -F "image=@input.png" | python3 -c "import sys,json;print(json.load(sys.stdin)['id'])")
echo "submitted: $JOB"

while :; do
  R=$(curl -s "https://api.therouter.ai/v1/jobs/$JOB"     -H "Authorization: Bearer $THE_ROUTER_API_KEY")
  S=$(echo "$R" | python3 -c "import sys,json;print(json.load(sys.stdin)['status'])")
  echo "status: $S"
  case "$S" in
    succeeded) echo "$R" | python3 -c "import sys,json;print(json.load(sys.stdin)['unsigned_urls'][0])"; break ;;
    failed|cancelled|expired) echo "$R"; exit 1 ;;
  esac
  sleep 5
done

More from qwen

Similar models

Cross-provider sibling models

News & changes

2025-12-23

Qwen releases Qwen-Image-Edit-2511 weights

The Qwen-Image repository lists Edit-2511 as a later image-editing release, with Hugging Face and ModelScope weights plus day-0 acceleration and inference-stack support. For TheRouter users, that is a lifecycle signal: the base qwen/qwen-image-edit route is useful today, but regression tests should compare it with newer Qwen edit variants before locking long-lived production workflows.

re-authored by TheRoutergithub.com/QwenLM β†—
2025-09-22

Qwen introduces Qwen-Image-Edit-2509 monthly iteration

Qwen's public changelog shows that image editing quality moved quickly after the August base release. Treat prompts, seeds, and acceptance tests as version-sensitive: a creative QA set that passes on one Qwen edit route may need re-baselining on later versions.

re-authored by TheRoutergithub.com/QwenLM β†—
2025-08-18

Qwen open-sources Qwen-Image-Edit

Alibaba Qwen released the editing version of Qwen-Image, extending the 20B model's text-rendering strengths into image editing. The launch emphasized semantic editing, appearance-preserving local edits, and direct Chinese/English text replacement inside images.

re-authored by TheRouterqwenlm.github.io β†—

Frequently asked

Is Qwen Image Edit a text-to-image model or an image editor?

It is primarily an image editor: send an existing image plus an instruction and receive an edited image. If you need to create an image from scratch, choose a text-to-image model instead.

re-authored by TheRouterqwenlm.github.io β†—
Why choose Qwen Image Edit over FLUX Kontext?

Choose Qwen when the hard part is editing Chinese/English text inside an existing image or preserving poster typography. Choose FLUX Kontext when you need very fast iterative edits, character consistency, or a Kontext-specific workflow already in your stack. For high-value assets, run both on a small QA set.

re-authored by TheRouterqwenlm.github.io β†—
Can it edit multiple input images?

Alibaba's Model Studio reference says qwen-image-edit supports single-image editing and multi-image fusion, while the base variant outputs one image. Through TheRouter, confirm the accepted image payload shape for your client before relying on multi-image production flows.

re-authored by TheRouterwww.alibabacloud.com β†—
Does seed make edits deterministic?

No. Seed can improve relative stability, but Alibaba's API reference explicitly warns that image generation is probabilistic and results may vary even with the same seed. Use visual regression checks for production asset pipelines.

re-authored by TheRouterwww.alibabacloud.com β†—
What should I log for a production Qwen Image Edit workflow?

Log model id, source-image hash (not raw private image unless policy allows), prompt version, seed, cfg/step settings, output URL expiry handling, and human approval status. This makes creative QA reproducible without storing sensitive assets unnecessarily.

re-authored by TheRouterwww.alibabacloud.com β†—
Fact ledger β€” every claim on this page traces here
sourceURLretrieved
Release dategithub.com β†—2026-07-22verified
Base modelqwenlm.github.io β†—2026-07-22verified
Editing architecturearxiv.org β†—2026-07-22verified
Input modalitieswww.alibabacloud.com β†—2026-07-22verified
Outputwww.alibabacloud.com β†—2026-07-22verified
Text editing languagesqwenlm.github.io β†—2026-07-22verified
Input image requirementswww.alibabacloud.com β†—2026-07-22verified
TheRouter pricetherouter.ai β†—2026-07-22verified
Training cutoffβ€”β€”unknown
Public image-editing benchmarksqwenlm.github.io β†—2026-07-22to verify
Qwen-Image technical reportarxiv.org β†—2026-07-22to verify
Qwen releases Qwen-Image-Edit-2511 weightsgithub.com/QwenLM β†—2026-07-22verified
Qwen introduces Qwen-Image-Edit-2509 monthly iterationgithub.com/QwenLM β†—2026-07-22verified
Qwen open-sources Qwen-Image-Editqwenlm.github.io β†—2026-07-22verified
Is Qwen Image Edit a text-to-image model or an image editor?qwenlm.github.io β†—2026-07-22to verify
Why choose Qwen Image Edit over FLUX Kontext?qwenlm.github.io β†—2026-07-22to verify
Can it edit multiple input images?www.alibabacloud.com β†—2026-07-22to verify
Does seed make edits deterministic?www.alibabacloud.com β†—2026-07-22to verify
What should I log for a production Qwen Image Edit workflow?www.alibabacloud.com β†—2026-07-22to verify
Help & contact