AIToday
Image GenerationZenn AI/MLPublished: Nov 18, 2026, 04:00 JST

Qwen-Image-Edit: 20B MMDiT picks for text-only fixes

Send on LINE
Qwen-Image-Edit: 20B MMDiT picks for text-only fixes

3 Key Points

  1. What happened

    Qwen-Image-Edit, published Aug 18, 2025 under Apache-2.0 at 20.43B parameters with a 20B MMDiT backbone, is positioned as the go-to engine for correcting text inside an existing image.

  2. Why it matters

    Splitting an edit into Qwen2.5-VL semantic control and a VAE appearance anchor is presented as the reason local fixes break down less than coarse inpainting, where small text outside the mask tends to collapse.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The problem is structural rather than prompt-related. Conventional text-driven regeneration and coarse inpainting push the input image deep into noise space before denoising again, which tends to leave unnatural mask edges and to crush small characters outside the mask. That failure mode is why teams retrying prompts on posters, packaging and menus end up inflating their generation count without gaining stability.

The proposed fix in Qwen-Image-Edit is to route the same input image down two paths. Qwen2.5-VL interprets the instruction and the image together to decide what should change and what should not, while a VAE encoder keeps color, texture, lighting and text layout as an anchor. Only after that separation does the 20B MMDiT receive the combined signal.

The tool also has boundaries. It is an edit model that requires an input image, so generating a base image from scratch is better left to Z-Image Turbo or GPT-Image-class models. Even with 2511's improved drift resistance, it recommends a single pass from the original image plus one instruction over four or five stacked overwrites. No recommended output resolution is stated, so callers are expected to plan sizes themselves.

FAQ
How large is Qwen-Image-Edit, and under what license is it released?
It is 20.43B parameters in BF16 with a 20B MMDiT backbone, released under Apache-2.0 on August 18, 2025.
What changed in the later 2509 and 2511 versions?
2509 added 1-to-3 image multi-image input and a native ControlNet for depth, edge and keypoint maps. 2511 suppresses quality and color drift from repeated edits, improves person consistency, and bundles community LoRA.
Are the recommended inference settings the same as the first release?
No. The first release was 50 steps at true_cfg_scale 4.0. The 2509 and 2511 versions are listed at 40 steps with true_cfg_scale 4.0 and guidance_scale 1.0.
Does it handle Japanese text well?
The officially stated strengths are English and Chinese text rendering. For posters with complex Japanese kanji, zoom in to check results.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAIエージェント guide: delegate email and research tasks, not whole jobs