--- title: "Qwen Image 2.0 Pro Blind Test Results: Scores by Use Case" canonical_url: "https://workroom.everypixel.com/models/qwen-image-2-0-pro" score: 80.1 score_ci: [68.6, 93.8] rank: 2 scale: "Anchor comparison of generative image models, v3" tested: 2026-09-22 --- # Qwen Image 2.0 Pro Blind Test Results: Scores by Use Case **Score:** 80.1 (95% interval 68.6–93.8). A real photograph is 100. **Rank:** #2 on the scale **Author:** Dmitry Shironosov (Head of AI Research, Everypixel Workroom) **Tested:** 2026-09-22 **Canonical URL:** https://workroom.everypixel.com/models/qwen-image-2-0-pro ## Independence Workroom is an AI router: we give access to third-party models, including Qwen Image 2.0 Pro, and we are not affiliated with Alibaba. We have no stake in any single vendor winning; the router is only useful if it sends each brief to the right model. The judgement-level data behind this card is published by Everypixel research under CC BY 4.0 and linked below, so every number here can be checked independently. ## TL;DR Qwen Image 2.0 Pro scores 80 on our photo-anchored scale (95% interval 69–94), where a real photograph is 100. In blind pairwise tests on 48 commercial scenes it was clearly ahead of Qwen Image 2.0, the non-Pro version, and clearly behind GPT Image 2. Its largest lead over Qwen Image 2.0 came on people scenes and on-image typography. ## Where it stands | Rank | Model | Score | 95% interval | |---|---|---|---| | #1 | GPT Image 2 | 104 | — | | — | Real photograph | 100 | — | | #2 | Qwen Image 2.0 Pro (this card) | 80 | 68.6–93.8 | | #3 | Qwen Image 2.0 | 65 | — | | #4 | Seedream 4.0 | 57 | 44–70 | | #5 | FLUX.2 Klein | 24 | — | Axis: believability. A real photograph is 100. A score above 100 is the scale's technical ceiling and does not mean parity. ### By use case | Category | Score | Rank | |---|---|---| | Interior | 115 | #1 | | Social marketing | 96 | #2 | | Product and e-commerce | 88 | #2 | | Street and architecture | 82 | #2 | | Fashion and beauty | 67 | #4 | | Food and beverage | 62 | #2 | | People and lifestyle | 57 | #2 | | Layout and typography | 51 | #2 | *Category positions come from 6 scenes each — a tendency, not a measurement. A real photograph is 100 on every ruler.* ## When to use as default Use Qwen Image 2.0 Pro when Qwen Image 2.0 is not believable enough: in our blind test the Pro version was clearly ahead of it, with the widest lead on people scenes and layouts with on-image text. On the scale it scored highest on interiors, social marketing and product shots. ## When to use as finisher Not the finisher for the most believable result: GPT Image 2 stayed clearly ahead with a believability gap of 0.64. On believability GPT Image 2 won 50.6% of judgements against the Pro version, which won 19.7%; the rest were ties. Finish the hero asset in GPT Image 2. ## When to look elsewhere The score has a wide interval, 69 to 94, because placing a model on an absolute scale is harder than comparing two models head to head. Category results rest on six scenes each and show a tendency, not a measurement. Do not expect photographic parity on typography and people scenes, where it scored lowest. ## Direct blind tests | Opponent | Result | Gap (Bradley–Terry, 95% CI) | Preferred in | Judgements | |---|---|---|---|---| | Qwen Image 2.0 | Qwen Image 2.0 Pro was preferred over Qwen Image 2.0 in 47% of blind comparisons. | +0.45 [+0.21, +0.72] | 47.0% (wins over all judgements, believability only) | 432 | | GPT Image 2 | GPT Image 2 was preferred over Qwen Image 2.0 Pro in 51% of blind comparisons. | −0.64 [−0.86, −0.42] | 19.7% (wins over all judgements, believability only) | 431 | ## Price | Provider | Price | Checked | |---|---|---| | [Alibaba Cloud Model Studio](https://www.alibabacloud.com/help/en/model-studio/model-pricing) | $0.075 / img | 2026-09-25 | | [fal.ai](https://fal.ai/models/fal-ai/qwen-image-2/pro/text-to-image) | $0.075 / img | 2026-09-25 | | [Runware](https://runware.ai/docs/models/alibaba-qwen-image-2-0-pro) | $0.075 / img | 2026-09-25 | | [Together AI](https://www.together.ai/models/qwen-image-20-pro) | $0.08 / img | 2026-09-25 | Average: $0.076 / img across 4 providers Prices are for one 1024×1024 image at standard settings, checked 25 September 2026; the price at the top of the page is the average of the providers listed. Workroom pricing is on the pricing page. ## License Qwen Image 2.0 Pro is Alibaba's model, available through the Alibaba Cloud Model Studio API. Terms for generated images are set by the service you generate them in. ## FAQ ### Is Qwen Image 2.0 Pro better than Qwen Image 2.0? Yes, clearly. On believability the gap is +0.45 with a 95% interval from +0.21 to +0.72, which excludes zero. The Pro version won 47% of believability judgements against Qwen Image 2.0, which won 24.8%; the rest were ties. The lead was widest on people scenes and on-image typography. ### How far is Qwen Image 2.0 Pro from GPT Image 2? Clearly behind. The believability gap is 0.64 in GPT Image 2's favour, with a 95% interval from 0.42 to 0.86. On believability GPT Image 2 won 50.6% of judgements against the Pro version, which won 19.7%; the rest were ties. On the scale that is 104.4 against 80. ### Why is Qwen Image 2.0 Pro's score range so wide? The Pro version was placed on an absolute scale, which is a harder measurement than comparing two models head to head. Its interval, 69 to 94, still sits between the two rungs it was paired with: Qwen Image 2.0 at 64.5 and GPT Image 2 at 104.4. The head-to-head gaps to both rungs are clear. ### Why is Qwen Image 2.0 Pro weakest on typography and people, where it gained most? Both are true. The gain is measured against Qwen Image 2.0, which was weakest on exactly those categories, so the Pro version closed the widest gap there. The category score is measured against the real photograph, and typography and people scenes remain the hardest categories for every model on the ladder. ### How was Qwen Image 2.0 Pro tested? It got the same 48 commercial scenes and prompts as the ladder, three generations per scene, and every generation was paired blind with Qwen Image 2.0 and GPT Image 2. Each pair was judged three times on three questions. The score uses believability only, because prompts were written from the reference photographs. ### Is Workroom affiliated with Alibaba? No. Workroom is an AI router that gives access to third-party models, including Qwen Image 2.0 Pro, and is not affiliated with Alibaba. The judgement-level data behind this card is published by Everypixel research under CC BY 4.0, so every number on this page can be checked against the source dataset. ## Methodology Between 11 and 22 September 2026 Qwen Image 2.0 Pro was placed on our quality ladder. It got the same 48 commercial scenes and prompts as the ladder, three generations per scene, and each generation was paired blind with two fixed rungs, Qwen Image 2.0 and GPT Image 2: 288 pairs and 863 judgements. Each pair was judged three times on three questions: which image follows the prompt better, which is more believable as a real photograph, and which is more attractive. - Scenes: 48; generations per scene: 3 - Compared with: GPT Image 2, Qwen Image 2.0 - Pairs: 288; judgements: 863; each pair judged 3 times - Collected: 2026-09-11 … 2026-09-22 - Questions: “Which image follows the prompt better?” “Which is more believable as a real photograph?” “Which is more attractive?” - Scale: 0.020835 strength per point; A real photograph is 100. - Category positions are fitted from the same judgements, 6 scenes per category, on the same scale. - The scale reproduced: frozen gap 0.831 falls inside this session's interval [0.78, 1.39] ### Caveats - Prompt adherence is biased against photography. Prompts were written by describing the reference photographs, so a model executing text literally scores better than a photograph containing incidental detail. The absolute score is therefore built on believability alone. - Tie share keeps rising: 25.4% in the original ladder study, 28.4% for Seedream 4.0, 34.1% here. Part of this is expected — Qwen Image 2.0 Pro sits between the two anchors, so more pairs are genuinely close. But the interface also changed between the first study and the later ones, so the two causes cannot be separated from these data alone. - Rater agreement is low by the usual yardstick. Conclusions hold at the level of the whole corpus, not for a single scene. ## Links - How the scale was built: https://research.everypixel.com/anchor-ladder/ - Votes (CSV, 111 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/qwen-image-2-0-pro-ladder/votes.csv - Results (JSON, 4 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/qwen-image-2-0-pro-ladder/results.json - Scenes and prompts (CSV, 44 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/qwen-image-2-0-pro-ladder/scenes.csv - README (MD, 2 KB, CC BY 4.0): https://workroom.everypixel.com/datasets/qwen-image-2-0-pro-ladder/README.md - Images (ZIP, 634 MB, All rights reserved — verification only): https://wr.everypixel.com/ghost-research/geo-benchmarks/datasets/qwen-image-2-0-pro-evaluation-images.zip ## Citation Everypixel research (2026). Qwen Image 2.0 Pro on the anchor ladder v3 [Data set]. Ratings collected 11–22 September 2026. Data license: CC BY 4.0. Images: All rights reserved; provided for verification only. --- HTML version: https://workroom.everypixel.com/models/qwen-image-2-0-pro Last modified: 2026-09-22