Twitter Prompt by @94vanAI

在AI角色设计中，**“多图人物一致性”**一直是个痛点。很多创作者发现，生成单张图完美无缺，一旦尝试多宫格（Grid View），角色特征（Identity）就会发生严重的漂移。 经过大量测试，我总结了一套标准化的工作流（Workflow），能够稳定输出高质量的15宫格角色卡。以下是核心逻辑： 目标设定单图内包含15个独立格，人物ID高度统一，光影（Lighting）与构图（Composition）连贯，达到专业Casting Card标准。 核心步骤详解 1. Character Seed（特征锚点）不要依赖抽象形容词。你需要对基准图进行像素级拆解： 骨相：oval face, angular jawline 皮相：warm skin tone, light freckles 特征：almond eyes, sharp nose 建议：利用LLM（如ChatGPT）反推Prompt，生成标准化的面部描述符。 2. Grid Prompt（结构化提示词）建立稳定的画面结构，防止AI自由发挥。 必须包含的Token： 15-panel grid photo sheet, same character, consistent appearance, uniform lighting, plain background 3. Action Matrix（动作矩阵控制）为了避免表情重复或崩坏，需引入“分层控制”概念： Layer 1 (Pos 1-3): Neutral/Basic (定基调) Layer 2 (Pos 4-6): Angles (侧面/回眸) Layer 3 (Pos 7-9): High Emotion (大笑/惊恐) Layer 4 (Pos 10-15): Interaction (手部动作/特殊微表情) Tech Stack: 使用 panel [x]: [description] 语法进行精确定位。 4. Consistency Lock（一致性死锁）如果以上步骤完成后仍有漂移，请执行**“白T恤策略”**： wearing simple white t-shirt in all panels 原理：降低服装纹理对模型注意力的干扰，强制模型将权重分配给面部特征。这是目前纯Prompt流派中最有效的Hack技巧。 结构化模板 (JSON)为了方便集成到工作流中，我封装了一个JSON配置对象，可直接用于自动化脚本或作为Prompt参考： 使用InstantID / IP-Adapter插件：上传一张参考脸，AI会严格复刻 这两个方法能让15宫格的一致性达到98%以上，出图直接能当演员资料卡、模特卡用。 一键复制：结构化模板 最后送大家一个我自己在用的JSON格式模板，适合喜欢结构化思维的朋友。直接改参数就能用： { "moodboard": { "layout": "15 panels vertical collage, 3×5 grid", "overall_aspect": "9:16", "each_panel_aspect": "3:4 or 4:5 (portrait)", "common_base": { "subject": "real diverse fashion model, close-up head-shoulders portrait, neutral/calm expression, professional studio photo", "lighting": "soft even studio lighting, ring light skin highlights, photorealistic texture", "background": "minimal plain or very soft radial/vertical gradient, low to medium saturation", "negative": "cartoon, illustration, 3d render, cgi, painting, blurry, deformed, plastic look" }, "panels": [ { "id": 1, "effect": "horizontal RGB glitch slices, channel offset" }, { "id": 2, "effect": "vertical gradient color bars covering eyes" }, { "id": 3, "effect": "full face holographic/iridescent gradient overlay" }, { "id": 4, "effect": "floating neon triangles + circles around head" }, { "id": 5, "effect": "digital melting drips from hair and chin downward" }, { "id": 6, "effect": "strong horizontal RGB displacement split" }, { "id": 7, "effect": "rectangular gradient blocks mosaic on lower face/mouth" }, { "id": 8, "effect": "cyber grid lines + glowing intersection points" }, { "id": 9, "effect": "diagonal wide holographic stripe across face" }, { "id": 10, "effect": "multiple floating circles forming halo/ring pattern" }, { "id": 11, "effect": "pixel sort vertical distortion on right half" }, { "id": 12, "effect": "fragmented color blocks partially covering face" }, { "id": 13, "effect": "VHS tracking lines + light chromatic aberration" }, { "id": 14, "effect": "horizontal glitch bars with color bleed" }, { "id": 15, "effect": "large centered floating geometric frame (square/circle)" } ], "optional_per_panel_elements": [ "reflective sunglasses / visor / partial chrome helmet", "small floating number (100–999 range)", "minimal single word / short phrase (Dream, Void, Error, Pulse, 404, Awake, Neon, etc.)", "subtle scanline / vhs noise overlay" ] } } 总结一句话： 锁脸 → 搭框架 → 派任务 → 上保险。四步走完，15宫格稳了。 有问题评论区见，祝大家出图顺利