# H3 提示词修正：白发白衣山景（参考图保持不变） > 素材：竖构图 1152×2048，白发高盘髻 + 粉色莲花发饰 + 长白发带、白色汉服（袖口祥云刺绣、白+青腰带）、 > 双手合十闭目祈祷；环境为**室外山谷**：前景成片紫色风铃草、背景绿丛与青灰岩峰、柔和阴天散射光。 > 原提示词（出问题的版本）： > ``` > integrated_multimodal_description: @图片1 Live-action, cinematic, a medium-wide shot frames a person > dancing energetically in a spacious room. The camera tracks the dancer's movements ... > ``` ## 一、为什么会"白发变黑发、室外变室内" | # | 原因 | 说明 | |---|---|---| | 1 | **没有锚定外观** | 全篇只有 "a person"，没写发色/发型/服装。H3 的参考机制要求**逐项声明要保留什么**，不写就等于不保留 | | 2 | **主动写了一个与原图冲突的场景** | `in a spacious room` = 室内空房间。模型优先服从文字描述，于是把山景整个换掉 | | 3 | **动作幅度与原图姿态不兼容** | `dancing energetically` 与"静立合十祈祷"差异极大，模型必须重构画布 → 顺手把人也换成了黑发 | | 4 | 参考图只提了一句 `@图片1` | 没有 `retention` 声明（保留/部分保留/属性迁移）。H3 需要显式写 `: identity_fully_preserved.` | **一句话**：这不像"参考图没生效"，而像"提示词明确要求换一个人、换一个房间"——模型照做了。 --- ## 二、方案 A（推荐）：参考生视频 Ref2VA 六段式 **节点设置**：模式 = `参考生视频 / reference`，参考图尺寸 `1k`，引用方式 `按序号(index)`，比例 `9:16`（原图是竖构图），时长 5s。 ```text subject_definitions: is the young woman in : silver-white hair combed into a high elegant bun, a pink lotus hair ornament on each side with long white hair ribbons hanging past her shoulders, fair skin, soft pink eye makeup, and a white hanfu with subtle cloud embroidery on the wide sleeves, a white inner collar and a white-and-teal waist sash; her eyes are closed and her hands are pressed together in prayer in front of her chest. is the outdoor environment in : a mountain valley with tall purple bellflowers in the foreground, blurred green shrubs and blue-grey rocky peaks behind, and soft overcast daylight with light mist in the distance. summary: [reference generation] The target video is one continuous shot of praying quietly in . A light breeze moves her sleeves, her hair ribbons and the purple flowers around her while the camera slowly pushes in. Her identity, her silver-white hair, her white hanfu and the outdoor mountain valley are fully preserved; the scene stays outdoors for the entire shot. retention_analysis: (appears in [Shot 1]): fully_preserved - identity_fully_preserved; the silver-white hair, high bun, pink lotus hair ornaments, long white hair ribbons, white hanfu with cloud embroidery and the white-and-teal sash are all retained, her eyes stay closed and her hands stay pressed together in prayer. (appears in [Shot 1]): fully_preserved - the outdoor mountain valley, the purple bellflowers in the foreground, the green shrubs and blue-grey peaks behind, the overcast sky and the light mist are retained; the scene remains outdoors and never moves into an interior space. detailed_description: The target video is in a gentle Chinese-fantasy film style with soft overcast daylight, cool green-and-violet tones, light valley mist and a shallow depth of field. [Shot 1] The shot begins exactly from : , the young woman with silver-white hair in a high bun and pink lotus hair ornaments, stands upright among the tall purple bellflowers with her eyes closed and her hands pressed together in prayer in front of her chest. She occupies the centre of the frame in the same pose, hairstyle, clothing and framing as . Behind her, is fully visible: the blurred green shrubs and the blue-grey rocky mountain peaks under the overcast sky. The camera pushes in with small amplitude at slow speed. A light breeze lifts the long white hair ribbons on both sides of her head and makes the wide embroidered sleeves sway; the tall purple bellflowers in the foreground rock gently and a few petals drift slowly past the lens. Her expression stays peaceful, her eyes remain closed, her hands do not separate and her feet do not leave the ground. She takes one slow, quiet breath and the white fabric over her shoulders rises and settles. The camera settles on her face at the end of the push-in, with the same mountain valley still visible behind her. overall_soundscape: A soft mountain breeze moves through leaves and tall grass around her. The wide white sleeves and the long hair ribbons rustle lightly with each gust. A few distant birds call once or twice, and one slow, quiet breath is audible near the microphone. non_diegetic_music: A sustained low string pad at a slow tempo with a single soft flute line above it. The flute enters as the camera begins to push in, holds a long note through the breeze, and fades out over the final two seconds. ``` --- ## 三、方案 B：图生或首尾帧 I2VA 三段式（你当前模式的最小改动版） **节点设置**：模式 = `图生或首尾帧 / image`，关键帧角色 `首帧（first）`，比例 `9:16`，时长 5s。 **注意**：第一行那句固定指令必须放在最前面，不能删。 ```text For the target video, at 0.00 seconds into the target video, (from [Shot 1]) is fully referenced. integrated_multimodal_description: [Shot 1] Live-action, cinematic, the young woman shown in stays exactly as she is — silver-white hair in a high bun with pink lotus hair ornaments and long white ribbons, a white hanfu with cloud embroidery, a white-and-teal waist sash, her eyes closed and her hands pressed together in prayer at her chest — and remains standing in the same outdoor mountain valley, with tall purple bellflowers in the foreground and blurred green shrubs and blue-grey peaks behind her under soft overcast daylight. Her appearance, hairstyle, clothing, pose, position in the frame and the entire background are preserved without change; no new location is introduced and the scene never moves indoors. The camera pushes in with small amplitude at slow speed as a light breeze lifts her white hair ribbons and sways her wide sleeves; the purple bellflowers rock gently and a few petals drift across the frame. She keeps her eyes closed and her hands together, takes one slow quiet breath, and at the end of the shot is still standing on the same spot with the same mountain valley behind her as the camera settles on her face. overall_soundscape: A soft mountain breeze moves through the leaves and tall grass around her, and the wide white sleeves and hair ribbons rustle lightly with each gust. A few distant birds call once or twice, with one slow quiet breath audible. non_diegetic_music: A sustained low string pad at a slow tempo with a single soft flute line above it, entering as the camera pushes in and fading out over the final two seconds. ``` --- ## 四、方案 C：确实要"跳舞"的写法 在方案 A 的 `detailed_description` 里，把动作句换成下面这段即可（**其余锚定句一句都不能删**）： ```text She opens her eyes, slowly lowers her hands and begins a restrained, graceful dance on the spot among the purple bellflowers: her arms sweep in wide, slow arcs so the white sleeves and the long hair ribbons trail through the air, and she turns her upper body from side to side while her feet stay on the same patch of grass. The camera pulls out with small amplitude at slow speed to keep her full figure and the mountain valley in frame, then arcs gently to her left. The dance stays in the same outdoor valley — the background never changes to a room, and the purple bellflowers, green shrubs and blue-grey peaks stay behind her throughout. ``` > 提醒：参考图是"静立合十"的姿态，写"大幅跳舞"意味着模型必须重构画面，**白发/服装/环境漂移的风险显著上升**。 > 若要跳舞，建议把时长放宽（8–10s）让它有铺垫，或先用方案 A 出一版稳定的，再拿那一版当参考续接。 --- ## 五、复用这个修正的通用公式 写参考类提示词时，本图的三个锚点缺一不可： 1. **人物外观逐项写**：发色/发型/发饰/服装/配饰/姿态 —— 少写一项就少保一项 2. **环境逐项写 + 显式禁止漂移**：`the scene remains outdoors and never moves into an interior space` 3. **动作幅度与参考图姿态兼容**：参考图是静姿就别要求剧烈运动，否则模型重建画面的同时会重建人物 配合 `retention_analysis` 里的固定标记（`fully_preserved` / `partially_preserved` / `attribute_transfer` / `weak_reference`）使用。