PROMPT RECORD图像记录

对比 Omni 与 Seedance 2 我想用一套复杂的提示词结构同时丢给两个模型,看看在榨干每一秒性能时它们各自表现如何。不废话,不打广告,纯粹一对一、像素级公平比较图像与提示词注入。 起...

中文说明

对比 Omni 与 Seedance 2 我想用一套复杂的提示词结构同时丢给两个模型,看看在榨干每一秒性能时它们各自表现如何。不废话,不打广告,纯粹一对一、像素级公平比较图像与提示词注入。 起始图像: 两边一致,见下方。 起始提示词 (在 Omni 上做了长度精简,而 SD2 最多可以跑到 15 秒): [CINEMATIC SETUP] 变形宽银幕 35mm,顶部荧光实用光 — 会议中心或社区礼堂那种平板无情感的灯光,柯达克罗姆胶片色调配以低饱和的机构色 — 米色地毯、折叠桌、姓名牌、1980 年代公共电视制作美学、背景中古董展会的环境人群低语声、机载麦克风音质 [@IMG_1 LEGEND] @IMG_1 = the artifact: 表面光滑的灰色乳胶皮肤的人形,头顶带猩红色尖刺王冠,额间镶嵌华丽第三眼...

原始 Prompt

Comparing Omni to Seedance 2

I wanted to throw a complex prompt structure at both models to see how they perform when pushed for every second. No bullshit, no paid ads, just raw apples to apples image and prompt injection.

Starting image: Same for both and included below. 

Starting Prompt (edited for length on Omni where as with SD2 I can run up to 15 seconds): 

[CINEMATIC SETUP]
Anamorphic 35mm, fluorescent overhead practical lighting — the flat affectless light of a convention center or community hall, Kodachrome with muted institutional colors — beige carpet, folding tables, name badges, 1980s public television production aesthetic, ambient crowd murmur of an antique fair in the background, on-camera microphone audio quality

[@IMG_1 LEGEND]
@IMG_1 = the artifact: glossy grey latex-skinned figure with crimson spiked crown, ornate third eye medallion at forehead, wide blue-irised bulging eyes with copper pupils, oversized red lips open in frozen expression, pointed ears, jeweled collar, decorative metalwork throughout — displayed on a folding table under fluorescent light

[SECOND-BY-SECOND TIMELINE]
MODE: SINGLE TAKE
0–2s | locked medium on host and artifact | host — middle-aged, cardigan, reading glasses, name badge — holds a small reference card and speaks directly to camera in complete deadpan: "What we have here is a mid-period Threshold Vessel, probably 1200 to 1400 interdimensional, the craftsmanship on the crown detail is really exceptional" — artifact completely still | convention hall ambience, PA system murmur, host's flat measured voice
2–5s | locked medium | host sets down reference card, grips the artifact's face with both hands at the jaw and temples, and PEELS IT OFF in one clean motion — the face comes away like a rubber mask, clean at the edges — host holds it up briefly, examines it, sets it aside on the table | wet suction release sound, host's expression unchanged, murmur continues
5–8s | slow zoom out begins | the faceless artifact begins glowing at the eye sockets — deep amber light pulsing from inside — host still speaking: "The interior cavity confirms it's a functional vessel rather than decorative — you can see the original resonance chambers are still intact" — artifact's neck elongating slowly, extra joints forming | glow hum begins under convention ambience, wet organic stretching sound, host's voice unchanged
8–11s | zoom out continues, artifact now mid-transformation | artifact has grown two additional arms from the collar area, jeweled metalwork spreading across the table surface like roots, the spiked crown extending into the air above, third eye medallion now orbiting the head slowly — host tilts his reading glasses up, leans slightly to examine a detail: "The medallion drift is a little unusual, that could affect value, but overall condition is quite good" | host's voice steady and academic, transformation sounds layering — metal growth, organic expansion, amber light intensifying
11–14s | zoom out wide, full convention hall visible | artifact now filling the frame above the folding table — eight feet tall, arms spanning the aisle, surrounding antique fair patrons visible in background continuing to browse completely unbothered — host now looking up at it, nodding: "Comparable pieces went at auction last cycle for somewhere in the range of forty to sixty thousand interdimensional units, though the provenance documentation you've provided is a bit thin" | ambient fair noise unchanged, one patron in background briefly glances up then returns to browsing, transformation sounds now architectural — deep resonance
14–15s | wide locked | host removes his reading glasses, folds them, slides them into his cardigan pocket, looks at camera: "I'd get a second opinion before you sell" | artifact fully transformed behind him, pulsing, orbital elements still moving — host completely still and satisfied | ambient fair noise, single PA chime, cut

[STYLE & QUALITY BOOSTERS + NEGATIVES]
BOOSTERS: Host must read as completely genuine public television antique appraiser — no camp, no winking at the camera, the comedy is entirely in his absolute sincerity, convention hall background activity must continue normally throughout — no one reacting, folding table must stay in frame as anchor of normalcy throughout zoom, artifact transformation must escalate in scale and strangeness proportionally to host's increasing calm, fluorescent lighting must persist — no dramatic lighting shift onto artifact, VHS grain and flat TV color throughout
NEGATIVES: No host reacting with fear or surprise at any point, no background crowd panicking, no dramatic music — convention hall ambience only, no artifact making sound directly — all transformation sounds are ambient and physical, no zoom moving faster than the transformation warrants

Video #1 is the Google Omni version (and this was the best of the 4 gens)

Video #2 is the Seedance 2 version and while not perfect, it included EVERYTHING in the prompt where Google's Omni omitted a lot. 

SD2 is still leading the pack currently.