AI场景生成绘画:一键创建完整环境与背景设定的高效方法
我做AI绘画最常接的私活类型不是角色设计——是场景。电商产品需要场景图把白底产品放进去、游戏项目需要大量环境概念图来定世界观、自媒体需要不同主题的背景来做封面。场景的需求量是角色的大概五倍——但大部分AI绘画教程都在讲怎么画人,很少讲怎么画"没有人但有很多东西的地方"。我刚开始画场景的时候犯过所有新手都会犯的错误:把Prompt写成了一个"物体列表"——"a forest with trees, a river, mountains in the background, a wooden bridge"——出来的图就像把所有素材从PS里拖到一个画布上,每个东西都"有",但整个画面就是没有"空间感"。后来我发现画场景的核心不是"画什么"——是"怎么让这些东西在空间里有正确的前后关系和统一的光环境"。这篇就专门拆场景生成的三个核心问题:空间深度怎么造、世界观怎么统一、角色怎么融进去。
场景的空间层次——前景框+中景主体+远景大气透视
好的场景图让人感觉"可以走进去"——这个"可进入感"来自明确的前景-中景-远景三层结构。前景不是必须的——但没有前景的场景就像透过一扇擦得太干净的窗户看出去:视野很清晰,但没有"身在镜头后面"的感觉。
三层空间Prompt:"the scene is composed in three distinct depth layers creating a strong sense of spatial depth——foreground layer: an out-of-focus stone archway frames the left side of the image, the stone is dark and textured with patches of moss visible even in the blur, this foreground element is so close to the viewer that it is partially cut off by the edge of the frame——it tells the viewer 'you are standing just behind this archway', middle ground layer: at the center of focus approximately 20 meters ahead, a small stone cottage with a smoking chimney sits in a clearing, every detail of the cottage is sharp——individual stones, the thatched roof texture, the curl of smoke, this is where the eye rests, background layer: beyond the cottage the land rises into forested hills that fade progressively into a blue-grey haze, the farthest ridge is barely distinguishable from the sky, atmospheric perspective has reduced its contrast to almost nothing"。在AI冰元素魔法绘画中冰封场景的空间构建用的也是这个三层法——冰雾取代了普通雾气但在空间中的角色是一样的。参考国家地理的风景摄影——专业风光摄影师对前景-中景-远景的运用是场景Prompt最好的老师。
世界观的"视觉宪章"——让十个不同地点看起来属于同一个世界
游戏和影视项目最需要AI做的不是"一张好场景图"——是"五十张风格统一的场景图"。保持统一的方法是在所有Prompt前面加一个固定不变的"视觉宪章"段落——定义这个世界的色调、光照、建筑、材质、植被的"规则"。规则定死之后,所有的变化都在规则范围内发生。
视觉宪章示例(废土世界):"[World Charter: Post-Collapse Earth——Color Palette: the world has forgotten vivid colors, everything exists in a spectrum of rust orange, concrete grey, and the pale green of plants that refuse to die, the sky is a permanent overcast——not dark storm clouds but a flat white-grey ceiling that has not broken in decades, Lighting: the sun is a rumor——all light is diffused through the permanent cloud layer creating shadowless ambient illumination, the only directional light comes from human sources——oil drum fires, neon signs salvaged from the old world, flickering fluorescent tubes running on jury-rigged generators, Architecture: no new buildings have been constructed in living memory——all structures are pre-collapse buildings that have been patched and modified, corrugated metal sheets cover shattered windows, satellite dishes are repurposed as water collectors, every structure looks like it could collapse but hasn't yet, Vegetation: nature is taking back the cities but slowly——moss and ivy grow in the cracks but they are grey-green and stunted, nothing grows tall or lush, the plants that survived the collapse were the ones that learned to live on concrete dust and acid rain]"。这个宪章写好后,任何地点追加描述——"a marketplace built in the ruins of a shopping mall / a solitary watchtower overlooking a dead highway / an underground bunker converted into a bar"——生成结果全部属于同一个废土世界。更多世界观构建在AI机械壁纸绘画的科幻场景设定中也有类似的"世界规则"描述方法。
氛围光的情绪控制——不同场景类型的"视觉温度"
场景的氛围不是靠"加滤镜"做出来的——是靠"光的方向+颜色+软硬"决定的。同样的森林,正午的硬顶光让它看起来像徒步路线、清晨的斜射逆光让它看起来像童话、阴天的漫反射光让它看起来像恐怖片。
三种氛围光对比。温馨治愈风:"the scene is bathed in warm morning sunlight streaming through a canopy of leaves——the light is golden and diffused by the morning mist, every surface is touched with a gentle amber glow, shadows are soft-edged and warm grey not black, the light rays are visible as distinct beams where they pass through gaps in the foliage——dust motes dance in these beams"。神秘探索风:"the scene is lit by an unseen light source from above——a shaft of pale silver light cuts through the darkness like a spotlight, the edges of the light beam are sharp and defined, everything outside the beam falls into deep indigo shadow, the light reveals only fragments of the scene——a stone staircase, a carved door, a single statue——the rest remains unknown"。末日荒凉风:"the light is the sickly yellow-green of a sky before a tornado——the entire scene is lit from a low angle by a sun that is barely visible through particulate haze, shadows are long and distorted stretching toward the viewer, the light flattens textures and drains colors of their warmth, everything looks like it was photographed through a dirty car windshield"。在AI堕天使绘画中哥特场景的氛围控制也用了同样的"光即情绪"逻辑——神圣场所和堕落场所的区分不在建筑样式,在光的颜色。
场景与角色的合成——一体生成vs分层合成的工作流选择
一体生成(角色和场景在同一张图里生成)的优点是光影天然统一——AI会自动让角色的皮肤色调和场景的环境光一致。但缺点也很致命:角色可能崩、场景局部可能崩、而且你无法单独修改其中一个。分层合成(分别生成角色和场景然后合成)的优缺点正好相反。
推荐的混合工作流:"Step 1: generate a combined scene+character image at low resolution to establish the lighting scheme——this serves as a 'lighting reference' only, don't worry about details, Step 2: generate the scene alone at full resolution using the lighting scheme from Step 1 as reference, iterating until the background is perfect, Step 3: generate the character alone at full resolution with matching lighting Prompt——'the character is lit by the same warm golden afternoon light as a forest clearing scene, rim light on the right shoulder, soft fill light from the left', Step 4: composite in Photoshop——place the character layer over the scene layer, use Curves and Color Balance adjustment layers clipped to the character to match the scene's exact color temperature, add a unified grain layer over everything to bind the two layers"。这个工作流在AI绘画师求职应聘中提到的商业项目交付流程里也是标准操作——专业交付和爱好者作品的区别就在于有没有走完这四步。
常见问题
AI生成的场景里总出现"不该有的东西"——比如森林里飘着一个路灯、沙漠里站着个邮筒?
这是AI的"语义污染"——AI对"forest"的理解混入了"公园"的训练数据(公园里有路灯),对"desert"的理解混入了"公路休息站"的数据(有邮筒)。消除方法:在Prompt里加"排除清单"——"the forest is wild and untouched by human infrastructure——no paths, no benches, no signs, no streetlights, no power lines, no evidence of human presence whatsoever, this is a forest that has never been managed or visited"。排除清单告诉AI"这个场景里不该出现什么"——对于降低语义污染非常有效。
怎么快速测试一个场景Prompt好不好——有没有比跑全分辨率图更快的验证方法?
快速验证法:用小尺寸(512×512或更小)+低步数(20步)快速出图——只看构图和光影对不对,不看细节质量。构图OK了再上全分辨率。快测一轮只要30秒——全分辨率一轮要3分钟——效率差6倍。在Midjourney中用"--style raw"参数可以去掉默认的美化滤镜,更快看到Prompt的真实构图效果。
室内场景和室外场景在Prompt写法上有什么根本不同?
室外场景的Prompt重点是"天空+大气+自然光"——光从天上某个位置来,受天气影响。室内场景的Prompt重点是"光源定位+空间围合+材质反射"——光从窗户、灯、火炉来,光在房间内弹跳。室内Prompt关键句:"the room is lit by a single window on the left wall——the light is soft northern exposure creating gentle shadows, the light falls across a wooden floor and illuminates dust particles in the air, the corners of the room farthest from the window fall into warm darkness——not pitch black but the deep amber brown of old wood in shadow, a secondary warm light from a fireplace on the right wall adds orange highlights to the opposite side of objects"。室内和室外写法的核心差异:室外写"天空",室内写"光源的数量和位置"。