Qwen Image Edit Plus AI Photo Editor

Every result below was edited here on 19 August 2026 with the exact instruction shown, on input images this site generated with other models. Seven of eight edits completed on the first run; one was retried once after a transient upstream error. Median time: 15 seconds.

Edits that touch only what you named — down to a point

The claim this model is bought for is that everything you did not mention survives. We ran it twice, once to pass and once to break. Asked to remove a plate and fork from a food photo, it erased them and extended the table naturally — galette, thyme and window light intact. Then we asked it to change one word on a supplement box and nothing else: the brand name came back cleanly re-set, the mid-size lines and even the batch code survived, but the 8pt ingredients panel it was explicitly told to keep was re-rendered into words that only look like words. Both results are below, input above output, unretouched.

Input above, output below · 11s

Input above, output below · 12s

The right-hand test is the one we designed to fail, and it did — partially. RENOVA, the dosage line and LOT 4471-B are perfect; the ingredients panel is not, because the model re-renders the whole frame rather than masking your edit. Keep must-survive fine print out of its hands: crop it away first, or add it back afterwards.

Rewrite the text inside the picture

This is the capability reviewers rank it best-in-class on, and the one that separates it from general editors: it re-sets type instead of painting over it. We asked for one line on a coffee tumbler's label to change — SINGLE ORIGIN to DARK ROAST — naming everything else as untouchable. The new line came back in the same typeface and spacing, and the caption-size line under it (COLOMBIA • HUILA | EST. 2018 | 12 OZ) stayed correct and legible. The Chinese half of the claim is tested in the set further down: a four-character calligraphy title rewritten correctly, with one honest caveat noted there.

Input — generated here with Nano Banana Pro

Output · 15s

Zoom into the output: the em dash, the letter spacing and the caption line all held. The scene is re-rendered rather than copied — the scattered beans shift slightly — so treat "unchanged" as visually faithful, not pixel-identical.

Three images in, one scene out

The 2509 update's headline feature is multi-image composition: person + product, person + scene, up to three inputs. We gave it the portrait, the tumbler and the galette from this page's other tests and asked for one cafe scene. The woman is recognisably the same person, the tumbler is in her hands, the galette is on the table, and the window light is consistent across all three — genuinely one scene, not a collage. The cost: the tumbler's label, which survived a single-image edit above, degraded into near-words through the composite. Plan multi-image work around that.

3 input images · 25s

All three inputs are images you have already seen on this page: the portrait from the identity test below, the tumbler from the text test above, the galette from the removal test. The instruction asked it to keep the label accurate; it kept the label's look, not its words.

No API key, flat pricing, and honest routing

Using this model directly means an Alibaba Cloud Model Studio account, an API key and per-image billing, or self-hosting a 20B model that wants serious GPU memory. Here it is signed-in-and-go: upload, type the instruction, and the flat credit cost shows before you run. Our Ava agent also knows this model's real profile — tightly-scoped text and object surgery — and routes away from it when the brief calls for something else: whole-scene creative restyling goes to Nano Banana 2, text-heavy artwork built from scratch goes to GPT Image 2. Every edit keeps its instruction and inputs in history, so a working recipe is reusable next week.

• E-commerce and product teams

  Label variants and background swaps without a reshoot: our tumbler kept its caption line through a one-word label change, and the perfume bottle moved from dark marble to daylight with its label intact.

• Marketers localizing creatives

  The same poster, retitled in Chinese or English by instruction. Our four-character calligraphy title came back correctly written — budget for the typeface drifting, and proofread anything you cannot read.

• Photographers cleaning up shots

  Remove the object that ruins the frame and let the model extend the background. The rest of the photo is re-rendered faithfully, not masked — so check fine detail before you ship.

• Anyone compositing people and products

  Three inputs, one scene, one light. Faces held up well in our composite; printed labels did not. Put the person in the scene first, then fix the label in a second, single-image edit.

• Not for: fine print and fresh canvases

  It cannot generate from nothing — no input image, no output. And whatever must stay pixel-perfect at 8pt should not pass through it at all: crop the panel out first or re-set it afterwards.

Credits, reference-image limits and size options are read from this workspace's own model registry on 19 August 2026. Elo scores are from Artificial Analysis' image-editing leaderboard on the same date — one source, one date, all four models. Speed rows are our own measurements with sample sizes stated; note the tasks differ (edits here, generations for the other three), so compare them as workflows, not as a benchmark.

1. 1

   Upload 1 to 3 images

   Create a free account — no API key, no cloud console. Drop in the photo to edit, plus up to two more if you are compositing a person, product or scene.

2. 2

   Describe one change

   Name what changes and, just as explicitly, what must stay. Tightly-scoped instructions are where this model shines; stacking several changes into one prompt is where it slips.

3. 3

   Generate

   The flat credit cost shows before you run. Our eight test edits took 11 to 25 seconds each, and the result lands in your history with the instruction attached.

4. 4

   Chain edits one at a time

   For a second change, feed the result back in as a new input rather than rewriting one mega-prompt. Each pass re-renders the frame, so keep chains short and check fine detail between steps.