A / INPUT
02
人物身份照片
Identity portraits
耶稣作为幸存者；女撒旦作为变成僵尸的爱人。白底照片只提供脸、头发和服装信息。
Jesus as the survivor; Female Satan as the zombified beloved. White-background portraits provide identity, hair, and wardrobe only.
PROMPT ENGINEERING · VIDEO EXPERIMENT
让镜头Directing 15 seconds
变得可读every second
一次关于 Zombie Love 生成的可复现实验：从宽泛的剧情描述，逐步改成有明确人物身份、画面轴线、僵尸妆效和逐秒摄影指令的镜头表。我们不只比较 prompt 字数，而是看模型是否真的拍出想要的镜头。
A reproducible Zombie Love generation experiment: we moved from broad story beats to explicit identity, screen direction, makeup requirements, and second-by-second cinematography. The question is not how long the prompt is, but whether the model produces the intended shots.
9:16 verticalSeedance 2.0 MiniDreamina · photo-only v515 秒完整版：已完成15s full cut: complete
01 / QUESTION
我们在验证什么？
What are we testing?
当只提供人物照片、不给动作参考视频时，prompt 能否控制剪辑节奏、镜头角度、人物位置，以及妆效从首次出场就清晰可见？
With character photos but no motion-reference video, can the prompt control pacing, camera language, screen position, and readable zombie makeup from first appearance?
B / METHOD
15×1s
按秒写镜头
One shot beat per second
把景别、机位、动作、焦点与运镜放到时间轴上，减少模型自行补剧情的空间。
Place framing, angle, action, focus, and camera movement on a timeline to reduce room for invented beats.
C / CONTROL
NO VIDEO
隔离文字影响
Isolate prompt effects
完整版只输入两张人物照，不输入动作参考视频；因此它测试文字引导，不承诺逐帧复刻参考片。
The full cut uses only two portraits, no motion reference. It tests text direction, not frame-accurate imitation.
02 / PROMPT EVOLUTION
改动在哪里？
What changed?
v1 → v4 → v5
V1 · Broad story beats
“Handheld camera, shallow depth of field…”
剧情按 0–4 / 4–6 / 6–10 / 10–15 秒分段。
“Handheld camera, shallow depth of field…”
Four broad segments: 0–4 / 4–6 / 6–10 / 10–15s.
→
V4 · Shot-list test
新增：耶稣固定左、女撒旦固定右
新增：逐秒景别、方向、焦点与运镜
新增：忽略白底；首次出场即僵尸妆
Added: fixed left/right screen positions
Added: framing, direction, focus, movement per second
Added: ignore white background; zombie makeup from first appearance
SeeAPI v3 · Identity leakage
“Replace the woman … with the person in the first reference image.” 结果：僵尸换成功，幸存者的原演员脸仍保留。
“Replace the woman … with the person in the first reference image.” Result: the zombie changed, but the survivor retained the source actor.
→
SeeAPI v4 draft · Not generated
写明东亚女性、黑色长直发、棕色眼睛
要求每个镜头都完全替换原演员，特写也覆盖
仅完成 prompt 草稿，未提交生成
Specify an East Asian woman with long straight black hair and brown eyes
Require complete replacement in every shot, including close-ups
Prompt draft only; never submitted
V4 · Limitation observed
6 秒测试中，门口与首次近景的僵尸妆较弱；裂纹皮肤和乳白眼睛不够明确。
In the 6s test, makeup read weakly in the doorway and first close-up; cracked skin and milky eyes were not clear enough.
→
V5 · Full 15-second plan
15 个连续 1 秒镜头，并把僵尸妆效写成可检查的视觉条件；10 秒硬切至麦田，逐秒描述结尾拥抱与拉远。
15 consecutive one-second shots, with zombie makeup expressed as observable visual requirements; a hard cut to the field at 10s and shot-by-shot ending coverage.
HIGHLIGHT KEY
绿色高亮 本轮新增或强化的指令new or strengthened instructions
红色文字 旧版的模糊点或已观察到的失败vague wording or observed failure
IMPORTANT
每个版本都保留原始 prompt 文件；高亮是在语义层面对照，不是逐字 diff。完整 v5 的真实输出出来后，结果区与评估会更新。
Original prompt files remain unchanged. Highlights are semantic comparisons, not a character-level diff. The results section will be updated when the full v5 render completes.
03 / SHOT DESIGN
15 秒镜头地图
15-second shot map
这张表把“故事应该发生什么”改成“每一秒镜头看见什么”。
This turns “what should happen” into “what the camera sees each second.”
耶稣含泪极近特写，枪管虚化在右下前景Jesus tearful extreme close-up; blurred barrel in lower-right foreground慢推；荆棘冠留在画内Slow push-in; keep crown in frame
左肩后越肩镜头；女撒旦在右侧逆光门口出现Over-left-shoulder; Female Satan appears in the backlit right doorway竖屏中远景，缓慢前移Vertical medium-wide, slow dolly forward
女撒旦向屋内慢走一步Female Satan takes one slow step inside从枪拉焦到她的脸；不跨轴Rack focus gun-to-face; hold axis
耶稣双手持枪，泪流满面但不开枪Jesus aims with trembling hands, crying but never firing稳定正面中近景Steady frontal medium close-up
女撒旦极近特写，直视画面左侧Female Satan extreme close-up, facing screen-left停留在乳白眼、裂纹皮肤和角Hold on milky eyes, cracked skin, horns
她微张嘴，露出克制无声的低吼She parts her lips in a restrained silent snarl紧特写，不攻击Tight close-up; no attack
耶稣视线从她移向枪，决定不开枪Jesus looks from her to the gun and decides not to fire含泪特写，轻推Tearful close-up, gentle push-in
耶稣放低手枪，让它安全落到地板Jesus lowers and safely releases the gun to the floor中景下摇跟枪，再回到脸Medium shot; tilt down with gun, return to face
耶稣张开双臂，女撒旦迟疑后靠近Jesus opens his arms; she hesitates, then steps closer保持轴线，缓慢推近Hold axis, slow dolly-in
女撒旦投入他怀中，他温柔接住She moves into his arms; he catches her gently木屋稍宽镜头，短横摇Wider cabin shot, short follow pan
拥抱动作中硬切至金色麦田；两人已恢复健康Hard match cut during embrace to golden field; both are healthy暖逆光中景双人镜头Warm backlit medium two-shot
两人相拥，轻轻转身They stay embraced and gently turn together夕阳下缓慢横向跟拍Slow lateral tracking at sunset
侧面中近景，悲伤渐渐转为释然Side medium close-up; grief softens into relief稳定机位，暖光Stable camera, warm light
脸贴着脸的紧特写，两人温柔对视Tight cheek-to-cheek close-up; tender eye contact微弱阳光光晕Subtle sun flare
慢慢拉远至麦田中景，夕阳收尾Slow pull-back to medium-wide; settle on sunset稳定停留结束Settle and hold to end
04 / GENERATED OUTPUTS
模型实际生成了什么？
What did the model actually make?
四条视频均可直接播放，包含失败样片和成功样片。SeeAPI 失败版用参考视频换身份；Dreamina 完整版只用人物照片。
All four videos play here, including failed and successful attempts. The SeeAPI failure used a motion reference; the Dreamina full render used portraits only.
05 / PROMPT EXCERPTS
指令文本对比
Prompt excerpts
可滚动查看关键差异；全部完整 prompt 和中文翻译均保存在实验记录中。
Scroll through the key differences. Full prompts and Chinese translations are preserved in the experiment logs.
V1 / ORIGINAL
Create one continuous 15-second cinematic zombie love story with sound, in four timed segments…
Segment 1 (0–4s): close-up, over-the-shoulder, medium shot.
Segment 2 (4–6s): two quick zombie close-ups.
Segment 3 (6–10s): lower the pistol, embrace.
Segment 4 (10–15s): wheat field memory.
Handheld camera, shallow depth of field…
Segment 1 (0–4s): close-up, over-the-shoulder, medium shot.
Segment 2 (4–6s): two quick zombie close-ups.
Segment 3 (6–10s): lower the pistol, embrace.
Segment 4 (10–15s): wheat field memory.
问题：镜头动作与切点仍留给模型自行解释。
Gap: camera moves and cut points remain open to interpretation.
V4 / SHOT LIST
CAST / AXIS:
In shared shots Jesus stays screen-left; Female Satan stays screen-right.
1–2s: over-the-left-shoulder medium-wide; slow forward dolly.
2–3s: rack focus from gun to her face; no side swap.
4–5s: extreme close-up; milky eyes, cracked skin, horns.
Ignore white portrait background.
In shared shots Jesus stays screen-left; Female Satan stays screen-right.
1–2s: over-the-left-shoulder medium-wide; slow forward dolly.
2–3s: rack focus from gun to her face; no side swap.
4–5s: extreme close-up; milky eyes, cracked skin, horns.
Ignore white portrait background.
变化：身份、左右关系、焦点和妆效变成硬约束。
Change: identity, screen direction, focus, and makeup become explicit constraints.
V5 / 15-SECOND PLAN
SHOT LIST — exactly one described shot per second:
0–1s extreme CU → 1–2s OTS dolly → 2–3s rack focus…
6–7s decision close-up → 7–8s tilt down with gun → 8–9s arms open…
10–11s hard match cut to field → 14–15s pull back and hold.
0–1s extreme CU → 1–2s OTS dolly → 2–3s rack focus…
6–7s decision close-up → 7–8s tilt down with gun → 8–9s arms open…
10–11s hard match cut to field → 14–15s pull back and hold.
变化：15 个镜头拍点都有景别、角度或运镜提示。
Change: all 15 beats specify framing, angle, or camera movement.
06 / FINDINGS
目前结论与限制
Findings & limitations
参考视频能锁住动作，也可能带入原演员
Motion references can lock action and leak source actors
SeeAPI v3 中僵尸身份替换成功，但幸存者在多个近景和麦田回忆里仍保留原片金发女演员；前置妆效测试也出现过原演员身份泄漏。v4 的身份覆盖改写只有草稿，尚无生成结果。
In SeeAPI v3, zombie replacement worked, but the survivor kept the source blonde actress across close-ups and the wheat-field memory. The earlier makeup-control test also leaked source identities. The stronger v4 identity-lock prompt remains an unsubmitted draft.
逐秒文字能控制剧情大节拍，不能锁定每一帧
Shot-by-shot text guides beats, not every frame
完整输出呈现了对峙、僵尸特写、放枪、拥抱和麦田回忆。各段顺序大体符合设计，具体镜头长度与机位仍有模型自由度。
The full output includes the standoff, zombie close-ups, gun drop, embrace, and wheat-field memory. The overall order follows the plan, while exact shot lengths and camera placement still vary.
音频说明：Audio note: 15 秒故事参考没有音轨。v4、妆效测试和完整版都有 AAC 音轨；本次只确认音轨存在，未核验声音内容或音乐匹配。The 15-second story reference is silent. The v4, makeup-control, and full outputs have AAC tracks. We confirmed stream presence only; audio content and music matching were not checked.
费用与状态：Credits & status: 6 秒任务报价不是最终净结算。本次完整任务已成功，resource 返回的 75 credits 是授权额度快照，不代表独立核实后的净扣费/退款。Dreamina 样片最长保存 7 天。Quotes for the 6s jobs are not final settlement. The full render succeeded; its 75-credit resource field is the pre-authorized snapshot, not an independently verified net charge/refund. Dreamina samples are stored for up to 7 days.
07 / REPRODUCIBILITY
复现实验
Reproducibility
FULL PROMPTS
EXPERIMENT SETTINGS
模型：照片输入实验使用 Dreamina Seedance 2.0 Mini · 720p · 9:16。这里不是只改变 prompt 的严格 A/B：SeeAPI v3 用 MiniMax H3、768P、参考视频和两张照片；Dreamina v4 与完整 15 秒版只用人物照片。已生成版本与未提交草稿分别标明。
Model: Dreamina Seedance 2.0 Mini · 720p · 9:16 for the photo-only tests. This is not a strict prompt-only A/B: the SeeAPI v3 clip used MiniMax H3, a 768P output, a motion-reference video and two photos; the Dreamina v4 and full 15s render used portraits only. The tested and unsubmitted drafts are labeled separately.