开发者总结 Jev 图片任务的五条实战经验
一位开发者给出用 Jev 处理图片的建议：结构化上下文加图片、1080px 以上分辨率、低置信度时升级到更强模型、边界框只当信号之一、用编码 agent 做评测拆解。
i don't know anything about it! if i had to guess based on what i know about jev so far - 1. provide as much structured context as possible, rather than either alone 2. model does better with hi res images rn so probably at least 1080px (~512kb-2mb jpegs is sweet spot) 3. escalate to smarter models in cases where jev is not confident, they are slower, but better than getting stuck 4. bounding boxes are still kinda experimental so i would use them as one of several signals what to do rather than go all in on them 5. think a lot and do evals with a coding agent to try and decompose the goal into multiple typesafe calls and programmatic steps rather than shove too much into one step