Microsoft AIImage modelPreview
MAI-Image-2.5-Pro
PopularityRank 670 of 875
0out of 100same as last month
Summary
MAI-Image-2.5-Pro is Microsoft’s proprietary image model for high-fidelity images, detailed editing, and precise text rendering. Microsoft placed it in public preview on July 23, 2026, with access through Azure Foundry and the Microsoft Playground.
Its pricing separates text input, image input, and image output. Public model weights have not been announced.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 12 published benchmark results
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
VGI-Bench image-generation adapted subset, Avg. success rate (Show measurement, test conditions, and source)
- Source value
- 36.8
- Score
- 36.8
- Metric
- Success Rate (%) against the ground-truth image, Avg.
- Unit
- %
- Benchmark version
- VGI-Bench arXiv:2608.19583v3, Table 4 image-generation adapted subset
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- VGI-Bench image-generation adapted subset
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Independent evaluation. The VGI-Bench image-generation auxiliary evaluation adapts 16 tasks to single-image output. Commercial models are accessed through APIs, image outputs are 1280 x 720, half of the instances in each task are selected with a fixed random seed, and Gemini-3-Flash is the automatic evaluator. The metric is static target-state success against the ground-truth image, not the video Final Score.
- Context
- The paper does not give per-model or per-difficulty sample counts or confidence intervals, and it does not identify exact commercial API endpoints or model revisions. This is a static target-state subset, not evidence of video process validity. The author list includes a Microsoft Research-affiliated coauthor, and the table includes Microsoft AI's MAI-Image-2.5-Pro; this affiliation is disclosed without asserting bias.
VGI-Bench image-generation adapted subset, Easy success rate (Show measurement, test conditions, and source)
- Source value
- 44.9
- Score
- 44.9
- Metric
- Success Rate (%) against the ground-truth image, Easy difficulty
- Unit
- %
- Benchmark version
- VGI-Bench arXiv:2608.19583v3, Table 4 image-generation adapted subset
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- VGI-Bench image-generation adapted subset
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Independent evaluation. The VGI-Bench image-generation auxiliary evaluation adapts 16 tasks to single-image output. Commercial models are accessed through APIs, image outputs are 1280 x 720, half of the instances in each task are selected with a fixed random seed, and Gemini-3-Flash is the automatic evaluator. The metric is static target-state success against the ground-truth image, not the video Final Score.
- Context
- The paper does not give per-model or per-difficulty sample counts or confidence intervals, and it does not identify exact commercial API endpoints or model revisions. This is a static target-state subset, not evidence of video process validity. The author list includes a Microsoft Research-affiliated coauthor, and the table includes Microsoft AI's MAI-Image-2.5-Pro; this affiliation is disclosed without asserting bias.
VGI-Bench image-generation adapted subset, Mid success rate (Show measurement, test conditions, and source)
- Source value
- 35.1
- Score
- 35.1
- Metric
- Success Rate (%) against the ground-truth image, Mid difficulty
- Unit
- %
- Benchmark version
- VGI-Bench arXiv:2608.19583v3, Table 4 image-generation adapted subset
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- VGI-Bench image-generation adapted subset
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Independent evaluation. The VGI-Bench image-generation auxiliary evaluation adapts 16 tasks to single-image output. Commercial models are accessed through APIs, image outputs are 1280 x 720, half of the instances in each task are selected with a fixed random seed, and Gemini-3-Flash is the automatic evaluator. The metric is static target-state success against the ground-truth image, not the video Final Score.
- Context
- The paper does not give per-model or per-difficulty sample counts or confidence intervals, and it does not identify exact commercial API endpoints or model revisions. This is a static target-state subset, not evidence of video process validity. The author list includes a Microsoft Research-affiliated coauthor, and the table includes Microsoft AI's MAI-Image-2.5-Pro; this affiliation is disclosed without asserting bias.
VGI-Bench image-generation adapted subset, Hard success rate (Show measurement, test conditions, and source)
- Source value
- 30.3
- Score
- 30.3
- Metric
- Success Rate (%) against the ground-truth image, Hard difficulty
- Unit
- %
- Benchmark version
- VGI-Bench arXiv:2608.19583v3, Table 4 image-generation adapted subset
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- VGI-Bench image-generation adapted subset
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Independent evaluation. The VGI-Bench image-generation auxiliary evaluation adapts 16 tasks to single-image output. Commercial models are accessed through APIs, image outputs are 1280 x 720, half of the instances in each task are selected with a fixed random seed, and Gemini-3-Flash is the automatic evaluator. The metric is static target-state success against the ground-truth image, not the video Final Score.
- Context
- The paper does not give per-model or per-difficulty sample counts or confidence intervals, and it does not identify exact commercial API endpoints or model revisions. This is a static target-state subset, not evidence of video process validity. The author list includes a Microsoft Research-affiliated coauthor, and the table includes Microsoft AI's MAI-Image-2.5-Pro; this affiliation is disclosed without asserting bias.
AA Image T2I (Show measurement, test conditions, and source)
- Source value
- 1100
- Score
- 1100
- Metric
- elo
- Unit
- elo
- Benchmark version
- AA Image T2I v2.0
- Category
- image-quality
- Direction
- Higher is better
- Sample
- 14804
- Confidence interval
- 7
- Lower confidence bound
- 1093
- Upper confidence bound
- 1107
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Independent evaluation. Text-to-image pairwise human-preference Arena; exact MAI-Image-2.5-Pro row; rank 12 with range 9-13.
- Context
- Independent Arena evidence. MAI-Image-2.5 base-model results are excluded because they do not identify Pro.
AA Image Arena Elo (Show measurement, test conditions, and source)
- Source value
- 1112
- Score
- 1112
- Metric
- pairwise human-preference Elo
- Unit
- elo
- Benchmark version
- AA-Image-T2I v2.1
- Category
- text-to-image
- Direction
- Higher is better
- Sample
- 11945
- Confidence interval
- 10
- Lower confidence bound
- 1102
- Upper confidence bound
- 1122
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-Image-T2I v2.1
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Overall text-to-image; exact leaderboard row MAI-Image-2.5-Pro.
AA Image Arena Elo (Show measurement, test conditions, and source)
- Source value
- 1122
- Score
- 1122
- Metric
- pairwise human-preference Elo
- Unit
- elo
- Benchmark version
- AA-Image-Editing v2.1
- Category
- image-editing
- Direction
- Higher is better
- Sample
- 15452
- Confidence interval
- 12
- Lower confidence bound
- 1110
- Upper confidence bound
- 1134
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis AA-Image-Editing v2.1
- Status
- active
- Retrieved at
- 2026-10-16
- Methodology
- Independent evaluation. Overall image editing; exact leaderboard row MAI-Image-2.5-Pro.
AA Image Arena Elo (Show measurement, test conditions, and source)
- Source value
- 1113
- Score
- 1113
- Metric
- pairwise human-preference Elo
- Unit
- elo
- Benchmark version
- AA-Image-T2I v2.1
- Category
- text-to-image
- Direction
- Higher is better
- Sample
- 11954
- Confidence interval
- 10
- Lower confidence bound
- 1103
- Upper confidence bound
- 1123
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Overall text-to-image; exact leaderboard row MAI-Image-2.5-Pro.
- Context
- This is a dynamic leaderboard snapshot retrieved on 2026-10-17, not a published vote-collection date. The page reports Samples but does not disclose a unique-prompt or unique-rater basis, exact serving revision, or whether votes overlap earlier snapshots. Do not treat this as independent replication.
AA Image Arena Elo (Show measurement, test conditions, and source)
- Source value
- 1123
- Score
- 1123
- Metric
- pairwise human-preference Elo
- Unit
- elo
- Benchmark version
- AA-Image-Editing v2.1
- Category
- image-editing
- Direction
- Higher is better
- Sample
- 15510
- Confidence interval
- 12
- Lower confidence bound
- 1111
- Upper confidence bound
- 1135
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Overall image editing; exact leaderboard row MAI-Image-2.5-Pro.
- Context
- This is a dynamic leaderboard snapshot retrieved on 2026-10-17, not a published vote-collection date. The page reports Samples but does not disclose a unique-prompt or unique-rater basis, exact serving revision, or whether votes overlap earlier snapshots. Do not treat this as independent replication.
OpenRouter Media full wine-glass prompt, displayed checklist result (Show measurement, test conditions, and source)
- Source value
- 3
- Score
- 3
- Metric
- Displayed four-point prompt score; rubric/check definitions not disclosed
- Unit
- count
- Benchmark version
- OpenRouter Media image challenge, live board; evaluation date and version unspecified
- Category
- text-to-image, single-prompt instruction following
- Direction
- Higher is better
- Source type
- official-leaderboard
- Evaluator
- OpenRouter Media image benchmark, full wine-glass prompt
- Status
- active
- Retrieved at
- 2026-10-19
- Methodology
- Official benchmark leaderboard. Single curated text-to-image prompt: a wine glass filled to the brim with white wine, surface level with the rim, upright on a plain wooden table, straight-on eye-level, no other objects. Displayed result is 3/4. Rubric, judge, repeats, exact evaluation date/version, and serving revision are not disclosed.
- Context
- One curated prompt, not a general image-quality scale. The board does not disclose the four-check rubric, judge, repeat count, evaluation date/version, or exact serving revision.
OpenRouter Media, full wine glass displayed generation time (Show measurement, test conditions, and source)
- Source value
- 30.3
- Score
- 30.3
- Metric
- Time displayed for the exact image result
- Unit
- seconds
- Category
- image-generation
- Direction
- Lower is better
- Source type
- official-leaderboard
- Evaluator
- OpenRouter Media image benchmark
- Status
- active
- Retrieved at
- 2026-10-21
- Methodology
- Official benchmark leaderboard. OpenRouter's Full wine glass image prompt; the exact result card is MAI-Image-2.5 Pro and displays $0.11, 30.3s, and checklist result 3/4. The page does not disclose a repeat count, aggregation rule, serving revision, or generation-region setting.
- Context
- This is the displayed time for one prompt-specific result, not a general model latency estimate. Replication count and timing aggregation are not reported. The 3/4 quality result is already present in the frozen benchmark signatures; only the time is a distinct cell.
AA-Image-T2I (Show measurement, test conditions, and source)
- Source value
- 1114
- Score
- 1114
- Metric
- Elo from pairwise human preference votes
- Unit
- elo
- Benchmark version
- AA-Image-T2I v2.1
- Category
- text-to-image, Overall
- Direction
- Higher is better
- Confidence interval
- 10
- Lower confidence bound
- 1104
- Upper confidence bound
- 1124
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- active
- Retrieved at
- 2026-10-23
- Methodology
- Independent evaluation. Overall category. Text-to-image generation from text prompts, rated by Elo from pairwise human-preference votes. Further per-model generation settings are not established.
- Context
- The number of independent test cases is not established.
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
The 100 most popular models of this category, 0 to 100, higher is more popular
20 of 101 models
MAI-Image-2.5-Pro0
Stable Diffusion XL 1.0, Stability AI: 72
Anima Base v1.0, CircleStone Labs and Comfy Org: 68
SDXL Turbo, Stability AI: 63
Stable Diffusion Turbo, Stability AI: 63
Nano Banana 2, Google: 61
FLUX.2 Max, Black Forest Labs: 55
Krea 2 Raw, Krea.ai: 54
Nano Banana Pro, Google: 50
FLUX.2 Pro, Black Forest Labs: 48
MOSS-Image-SFT, OpenMOSS: 46
FLUX.2 Flex, Black Forest Labs: 46
GPT Image 2, OpenAI: 46
LongCat-Image, Meituan: 45
MAI-Image-2.6, Microsoft AI: 45
FLUX.2 Klein 9B, Black Forest Labs: 45
PixArt-Sigma-XL-2-1024-MS, PixArt-Alpha: 44
Janus-Pro-7B, DeepSeek: 44
HunyuanImage 3.0 Instruct, Tencent: 44
MOSS-Image-Base, OpenMOSS: 44
MAI-Image-2.5-Pro, Microsoft AI: 0
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 5, 2026
MAI-Image-2.5-ProFLUX 3 Image KleinHy Image 3.5 PreviewMAI-Image-2.6 FlashIdeogram 4.5
All specifications
Architecture
Input and output
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
Sources and data date
Show all 13 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.