OpenRouter 上线 Visual Image Benchmarks，横向对比 39 个图像模型能力
Image Benchmarks: See the Capabilities of Every Model
- 1OpenRouter 上线 Visual Image Benchmarks，可统一对比其平台上全部 39 个图像模型
- 2七类挑战提示词覆盖低概率场景、计数、文字、空间关系、否定、图像编辑与一致性
- 3同一对比格式已扩展到视频和语音模型，结果网格支持按价格和生成时间排序
- 4新图像模型上线即加入基准网格，评估后可在图像生成 API 或 Chat 中用自己的提示词测试
图像模型选型缺少类似文本模型的成熟基准，OpenRouter 把选型对比做成了按价格与生成时间排序的可视化工具。
- 影响：企业挑选图像模型时有了统一参照，可先看计数、文字、编辑这些高失败率场景的差异再下单；价格与生成时间可直接排序，逐家试错的评估成本能压下来；新模型上线即入网格，比选参照更新更快。
- 参考：可考虑按自身高频场景挑选对应基准提示词，如海报长文字、产品一致性，先跑一遍候选模型再比价；值得把价格与生成时间的排序结果纳入模型选型清单。
- 依据：全部信息来自 OpenRouter 官方博客，39 个模型的覆盖数与七类提示词设计均为官方口径，暂无独立验证。
Unlike choosing a text model, where we have a vast array of LLM benchmarks, picking an image model can feel arbitrary. Output samples tend to be curated eye candy and LLM-as-a-judge evals can’t yet capture the details a human would notice instantly. While arena scores help, they evaluate which outputs people prefer rather than what a model can actually do.
Today we’re launching Visual Image Benchmarks to help you quickly evaluate the capabilities of all the image models we offer (39 as of August ‘26). We’ve selected a set of challenging prompts designed to differentiate the capabilities of models, and show every result in a grid with sorting for both price and generation time.
Prompts designed to test the boundaries of image model capabilities
Each challenge is designed to differentiate the capabilities of image models. We’ve initially grouped them into seven families:
- Improbable scenes. A wine glass filled level with the rim, umbrellas that are closed. Training data is full of the ordinary version of both.
- Counting. Three fingers, specific numbers of cards and dice.
- Text. One long exact string on a poster, and several languages in the same frame.
- Spatial relations. Occlusion and mirror reflections.
- Negation. A zebra with no stripes, a Times Square with no advertising.
- Editing. Minimal diffs, object removal, person removal, all from a reference image.
- Consistency. Holding a product or four reference subjects steady across a new scene.
The prompts are written so you can instantly evaluate them visually. For example, can a model follow the instruction to fully fill a wine glass?
Benchmarks for video and speech models
Video and audio models are just as hard to evaluate, so the same side-by-side format now covers video and speech models too. We add new image models to the grid as they launch.
Check out our image benchmarks today! If you have any prompts that could challenge the next round of models, share it with us in #feedback on our Discord.
Generating images via the OpenRouter API and Chat
Once you’ve evaluated models via these benchmarks, try them on your own prompts through the image generation API or Chat to see how they perform on your own content.
来源：OpenRouter：Blog（Announcements） · openrouter.ai