多图合成问题

1. 我们在用 Qwen-Image-Edit-2511 做多图车辆合成。两张各 1 辆车合成 2 辆可以；但把第 3 辆往已经完成的 2 车图上加，模型会按 2 辆重画，丢掉新增那辆。换顺序、放宽布局、换成差异很大的红车都不行。先 PIL 贴成 3 车再让模型融合，辆数能保住，但很像贴纸。想确认：往已完成多主体照片里再插入一个物体，是不是 2511 的设计能力？有没有推荐的 mask/inpaint 或 prompt 写法？
2. 我们尝试把3张车图合并到一张上去，背景不变，多次尝试，要不就是 3张汽车 在一起，但是背景丢失了（空白背景），要不就是会少一辆车。

但把第 3 辆往已经完成的 2 车图上加的 Prompt如下： “ 将 Image2 中的汽车主体提取出来，合成进 Image1（第一张图）里。背景不需要变化——以第一张图的场景为准。

Hard constraints:

• This step pastes 1 extra car onto Image1.
• Keep every vehicle already visible in Image1, then add Image2's car. Do NOT omit, merge, replace, or invent cars.
• Image1 is the BASE PHOTO / canvas: keep Image1's original background exactly (river, rocks, forest, lighting, weather). Do NOT replace, wipe, or invent a new scene. Do NOT use empty studio / plain white/gray / cutout plate.
• Fill the entire Image1 frame edge-to-edge. River, rocks, and forest must reach the top and bottom edges — no white/gray bar, letterbox, pillarbox, or empty padding at the bottom.
• From Image2: keep ONLY the vehicle subject (body color, shape, trim, logos). Discard Image2's background entirely.
• Place Image2's car naturally into Image1's scene alongside the vehicles already in Image1 (same ground plane).
• Keep every vehicle already in Image1 as-is (body color, shape, trim, accessories).
• Image2 subject: keep that reference vehicle as-is (body color, shape, trim, accessories).
• Arrange cars side-by-side or in a natural staggered parking/driving layout on the same ground plane. NEVER stack one car on top of another; no overlapping chassis / floating cars.
• Brand logos: if a reference shows a xxx logo, the output MUST keep the xxx logo on that car — do NOT replace with Mercedes-Benz or any other brand emblem.
• Reproduce brand logos and model names EXACTLY — do NOT alter any lettering.
• REMOVE any 'AI生成', 'AI Generated', 'AI creation', or similar AI-tool watermark badges from the output — the final image must be clean. ”