Seedance 2.5:更长、参考很多的创作简报
在列出的 Topview 工作流中,每条片段最长 30 秒,参考素材合计最多 50 个。


该选哪个 AI 视频模型?
Seedance 2.5 和 MiniMax H3 都支持带同步声音的多模态 AI 视频生成,但适合的创作工作流不同。
简报需要单次生成最长 30 秒、更大的参考集,或更细的场面时间控制时,选 Seedance 2.5。需要更短的 5–15 秒片段、一套可用的 2K 输出工作流,或在符合条件的定制部署中使用可下载模型权重时,可以考虑 MiniMax H3。
生成前,先对比支持的设置和原始视频示例。两个模型都不会自动更适合每一个镜头。
Seedance 2.5:更长、参考很多的创作简报
在列出的 Topview 工作流中,每条片段最长 30 秒,参考素材合计最多 50 个。
MiniMax H3:短视频与 2K 重新生成
每条片段最长 15 秒,带立体声音频,并有一条最高 2K 的高分辨率重新生成路径。
主要取舍是片段时长和参考容量,相对的是 H3 更短的工作流,以及它的高分辨率重新生成选项。表格把已公布的模型能力,和 Topview 列出的设置区分开。
| 你在对比什么 | Seedance 2.5 | MiniMax H3 |
|---|---|---|
| 模型开发方 | ByteDance | MiniMax |
| 主要模型类别 | 多模态 AI 视频模型(托管) | 全模态 AI 视频模型(开放权重) |
| Topview 上可用的时长 | 4–30 秒 | 5–15 秒 |
| 已公布的上游时长 | 最长 30 秒 | 4–15 秒 |
| Topview 列出的输出选项 | 480p / 720p / 1080p | 480p / 768p / 2K |
| 2K 输出如何实现 | 未列为 Topview 的原生输出档位 | 先输出 768p 基底,再经 H3-Regenerate-2K,且仅在支持该流程时 |
| 文生视频 | 支持 | 支持 |
| 图生视频 | 支持 | 支持 |
| 参考图片 | 最多 30 | 最多 9 |
| 参考视频片段 | 最多 10 | 最多 3 |
| 参考音频片段 | 最多 10 | 最多 3 |
| 混合参考文件总数 | 最多 50 | 最多 12(各参考合计最多 12) |
| 参考视频时长 | 合计时长受 30 秒上限约束 | 合计时长最长 15 秒 |
| 参考音频时长 | 合计时长受 30 秒上限约束 | 合计时长最长 15 秒 |
| 生成的音频 | 在支持的工作流中生成对白、音乐和音效 | 原生立体声音视频生成 |
| 时间控制 / 创作方向 | 带时间码的镜头与声音指令 | 多模态指令,以及由参考引导的方向 |
| 编辑工作流 | 在 Topview 上可用时,可做选定片段编辑 | 能力取决于 H3 变体和工作流 |
| 在 Topview 上的访问方式 | 托管生成器,以及受支持的 API 访问 | 托管生成器,以及受支持的 API 访问 |
| 可下载的模型权重 | 不通过此 Topview 工作流提供 | 已公布,受模型许可和地区限制约束 |
| 适合作为起点 | 更长的视频、更大的参考包、复杂的连续镜头 | 更短的片段、面向 2K 的交付,以及符合条件时的定制部署 |
MiniMax 的官方模型文档规定,参考文件合计上限为 12 个,尽管各类别上限相加为 15。音频参考不能作为 H3 的唯一参考输入。Topview 网页端和 API 的可选项可能因模式、账号和模型变体而不同。
按视频要完成的任务来选,而不是按模型名称。
更长、参考很多的场面,Seedance 2.5 是自然的起点。更短的片段、经其支持的重新生成路径做 2K 交付,以及获许可的模型定制,则值得考虑 MiniMax H3。
在一次生成里安排开头、发展和结尾。广告、讲解和多节拍叙事,会比单条 15 秒片段有更多空间。
在列出的 Topview 工作流中,最多使用 50 个受支持的参考素材:30 张图片、10 段视频和 10 个音频文件,并受文件与总时长限制。用来引导角色、产品、地点、动作和声音。
用带时间码的指令指导运镜、动作、对白、音乐和音效。特定节拍必须按计划顺序出现时,这很有用。
所选 Topview 工作流支持局部编辑时,可以修改目标片段,并尽量保留周围内容。用于正式制作前,先确认编辑模式。
使用受支持的 H3 高分辨率重新生成工作流,产出最高 2K 的结果。产品细节、电影感构图,以及其他画面细节重要的短片交付,值得拿它来测试。
在同一工作流中生成带声音的音视频片段。H3 的文档写明立体声输出和多种对白语言;请始终在实际导出文件里核对语音准确度和同步。
若预期用途符合其社区许可、地区要求和硬件需求,可以探索 H3 已公布的权重,用于定制基础设施和集成。通过 Topview 使用 H3,本身并不授予自托管许可。
想法适合较短时长和较小参考集时,先对比 H3,再决定是否进入更复杂的制作工作流。单次生成要 15–30 秒时,改为查看 Seedance 2.5。
六条创作简报。两套模型输出。对比真正要紧的细节。
以下是 Topview 上现有的六组对照示例。它们覆盖视觉特效、口语对白、多角色动作、舞蹈、文学叙事,以及由参考引导的游戏战斗。
每项测试都保留原始输入简报和参考素材。两段视频都看,核对提示词里要求的动作,再判断每条输出是否完成创作任务。这些展示的对照是说明性对比,不是对任一模型有统计代表性的排名。
文生视频 · 15s · 16:9
一位带北欧风格的英雄悬浮在破裂的黑曜石神殿上方。金色闪电、漂浮的石块碎片和发光符文逐渐汇聚,形成一把战锤。原始简报要求连续的环绕运镜、一道放射状冲击波,以及最后的广角英雄镜头。


特效很重的视频可能看起来很抢眼,却漏掉让镜头连贯的关键动作。这个场景测试每个模型能否随时间协调运镜、角色表演和特效。
这个镜头必须在同一次观看中逐步升级,并以广角结束时,Seedance 2.5 更合适。画面保持同一位留胡须的英雄,经过闪电和飞散的碎片,举起锤子,并以该人物立于峰顶的广角结束。镜头距离有变化,所以环绕并非完全不中断,但动作顺序仍然可读。MiniMax H3 也生成了闪电和锤子,但景别从行走的广角跳到特写、漩涡,再回到行走,而且英雄的服装在不同画面之间发生变化。要运镜连贯和计划中的广角结尾,选 Seedance 2.5。只有较短的特效蒙太奇就够,并且可以接受切换时,才选 MiniMax H3。
A high-concept mythological action VFX cinematic spectacle shot on Sony Venice 2 with Cine Alta 8K sensor and Leica Summilux-C lenses, featuring photorealistic golden lightning arc flashes, dynamic ray-traced particle physics, floating terrestrial debris, and glowing ancient divine runes illuminated by hyper-detailed volumetric atmosphere. The continuous, seamless 360-degree whip-pan camera begins in a low-angle orbital sweep around a muscular, bearded Norse god suspended in mid-air above a fractured obsidian mountain shrine under a tempestuous crimson sky. His initial solemn and stoic composure dramatically shifts into furious rage as his eyes burst into blinding blue plasma energy, casting vibrant specular highlights across his weathered skin and braided hair. As the camera accelerates its continuous spiral rotation around the god, he lets out a guttural, thunderous roar, raising his right hand to summon violent streaks of celestial lightning that erupt from swirling storm clouds. The raw electrical arcs violently pull floating stone fragments and divine liquid metal together into a swirling vortex, instantly fusing them into a massive, glowing, runic-engraved battle hammer in his fist. Upon gripping the weapon, he detonates a massive radial gravity shockwave that instantly pulverizes the surrounding floating boulders into fine stardust particles and ripples through the thick atmosphere, leaving the ancient fractured shrine beneath him pulsing with deep blue subterranean magic. The unbroken dynamic shot completes its orbital loop to settle into a dramatic wide hero framing of the god hovering amidst the dissipating energy dust, legendary, majestic, thunderous, and escalating from tense stillness to an overwhelming explosive climax.
3 张图片参考 · 15s · 16:9
两位风格化的动画角色 HAORAN 和 XIAOYU 在咖啡馆外见面,交换四句指定的英文对白,然后进入建筑。三张图片定义角色和环境。简报还指定了镜头的推进方式,以及向外开的咖啡馆门。


对白视频需要的不只是看起来自然的脸。模型还要协调具体台词、说话角色、肢体动作,以及基于参考的身份。
两段视频都保留了咖啡馆外、穿蓝色毛衣的风格化男性,以及穿白色上衣、粉色裙子的女性,并且都从街道走向门口。Seedance 2.5 展示他们见面、交谈、走到台阶,以及男性到达门口。MiniMax H3 展示同样的接近和面对面交流,然后女性转向门口,此时男性仍在她身边。在已核对的画面里,两段都没有清楚地向外开门,并让 XIAOYU 先进入。Seedance 2.5 更接近这个入场动作。门的动作属于简报的一部分时,用它。目标是一段短的咖啡馆对话时,用 MiniMax H3。在依赖准确的台词顺序之前,先带声音播放两段;这些画面展示的是见面和接近,不是逐字稿。
文生视频 · 15s · 16:9
一群毛绒毡偶动物角色坐在车里。提示词给柴犬司机和乘客分配了不同动作,包括唱歌、自拍、鸭子抢镜、移动帽子、飞起的物体,以及对车辆运动的反应。镜头应固定在车内。


拥挤场面要求模型同时维持多个身份和重叠动作。开头一帧很强,并不保证同一批角色能撑完整段镜头。
Seedance 2.5 是更稳的拥挤镜头。机位留在车内,柴犬司机、戴西瓜帽的猫、鸭子、浣熊和篮子里的动物保持可辨认,帽子在片段中段脱落。并非每个指定的笑点都彼此分开,但角色阵容没有被替换。MiniMax H3 保留了车内和其中几只相同的动物,但鸭子、帽子和乘客发生替换,还出现了新的动物,所以身份和道具连贯更弱。同一组角色必须撑满 15 秒时,选 Seedance 2.5。只有热闹的车内比每只动物各司其职更重要时,才选 MiniMax H3。
1 张图片 + 音频参考 · 15s · 9:16
五名舞者跟随上传的音乐,并保持群像参考图片中的脸、发型和服装。编舞从 V 字形开始,并要求在一个连续镜头里变成横排。


舞蹈片段同时考验对身份、编舞、站位和音乐时间的控制。一帧好看的静帧,不能证明动作跟上了节拍。
两段都无法可靠地锁定五名舞者。Seedance 2.5 开头是涂鸦墙前的一小群人,后来展开成一排,但以两人特写结束,所以五名舞者的要求并不成立。MiniMax H3 开头就多出了人,后来又换成另一排舞者,脸和服装没有始终对应同一份参考。如果需要开头的人群保持更少,Seedance 2.5 是更好的起点。对照参考音轨播放之前,不要把任一结果当成已经对齐节拍;画面显示的是姿势变化,不是步伐是否落在节拍上。
文生视频 · 15s · 16:9
Kenneth Grahame 所著 The Wind in the Willows(《柳林风声》)中河船场面的水彩风格动画改编。水鼠划着一艘蓝白相间的船,鼹鼠坐在船尾;船撞上河岸,水鼠向后摔倒,然后恢复过来。


文学改编测试的是模型把书面事件转成视觉动作的准确程度。好看的插画不能代替按描述实际发生的故事。
Seedance 2.5 更贴近这段原文。画面保持水彩河流、一艘蓝色的船、一只在划船的水鼠,以及船尾的鼹鼠,然后显示水鼠仰躺在船里,再坐起来。这是划船、摔倒和恢复,并且是同一种绘画风格。MiniMax H3 也保持水彩和蓝色的船,但划船者被画成穿条纹外套、戴帽子的人,摔倒也更不清楚是否由船撞上河岸引起。角色分工和故事顺序重要时,选 Seedance 2.5。就这段原文而言,MiniMax H3 是更弱的改编。
2 张图片参考 · 15s · 16:9
TOPVIEW BOT 是一位小巧敏捷的女主角,在色彩丰富的游戏世界里与巨型玩具机甲战斗。两张参考图片定义这两个角色。简报要求一连串衔接的攻击、投射物、攀爬移动和最后一击,同时保持角色设计和相对比例。


复杂战斗测试角色一致性、动作依赖和空间推理。如果底层运动已经无法跟上,一次壮观的爆炸并不够。
两段都保留了矮小的黄色女主角,以及大得多的蓝黄机甲,所以比例对比在各自画面里都可读。Seedance 2.5 在近身战斗、爆炸和广角中的最后一击里,都停留在这一对角色上,身份变化更少。MiniMax H3 也展示了炮管、爬上机甲和一次击中,但较早的一帧先出现了另一个黄色角色,然后女主角才回来,所以参考设计没那么稳定。提供的角色必须撑过整场战斗时,选 Seedance 2.5。如果优先的是比例对比,并且你会核对女主角是否仍匹配参考,则选 MiniMax H3。
我们把创作者要求的任务,和每个模型实际生成的结果放在一起比较。
每个场景都有自己的原始提示词,并在适用时带有原始图片或音频参考。两个模型应使用同一条预期简报。模型专用的输入标签可能不同;这些差异已经记录。
记录每条输出的实际时长、画幅比例、分辨率、模型 ID、生成模式和音频设置。要求的时长相同,并不保证导出配置相同,尤其是两个模型提供不同分辨率档位时。
检查指定的事件、行动者、物体、运镜和结尾是否按正确顺序出现。
对由参考引导的案例,把可辨认的脸、服装、产品、道具和环境细节与源素材核对。
评估镜头变化、身体动作、物体轨迹、空间关系,以及场面是否意外重置。
音频属于任务的一部分时,检查口语准确度、说话角色、口型同步、音乐连贯,以及与动作的同步。
把技术上完成的渲染,和满足预期创作要求的视频区分开。失败的任务、不完整的输出和质量否决应分开记录。
现有片段按成对示例展示。我们事先把生成总次数定为 52 次,并事先定好筛选规则,同时保留全部尝试。展示的视频是得分处于中位数的那条输出。
先用同一条预期任务,再比较积分报价和可用结果。
| 已公布的输出档位 | Topview 上的 Seedance 2.5 | Topview 上的 MiniMax H3 |
|---|---|---|
| 480p | 0.7 积分/秒 | 0.15 积分/秒 |
| 720p(Seedance)/ 768p(H3) | 1.5 积分/秒 | 0.3 积分/秒 |
| 1080p(Seedance)/ 2K(H3) | 3.7 积分/秒 | 0.6 积分/秒 |
中间档和更高档并不是相同的分辨率,而且 H3 的 2K 交付涉及重新生成过程。这些示例比较的是已公布的积分收费,不是同一像素数量下的画面质量,也不是相同技术条件下的制作成本。
单次生成的最低积分,不一定是完成整个项目的最低积分。记录每一次消耗了积分的尝试,以及有多少结果达到你的质量要求。
每条可用输出的成本 = 消耗的总积分 ÷ 可用视频条数。
公平的成本对比还需要工作流、账号套餐、任务时长、输出设置和尝试次数。
当你的输出目标比更长的片段或更大的参考包更重要时,MiniMax H3 值得测试。
如果镜头能放在 15 秒以内,并且你想评估 2K 重新生成工作流,就把 H3 的输出和最终交付里实际需要的细节做对比。
可下载权重可能会吸引搭建内部系统的技术团队。在部署或分发输出之前,先核对当前的社区许可、排除地区、额外许可和硬件要求。
在其限制内,H3 可以组合受支持的文本、图片、视频和音频上下文。更小、结构清楚的参考集,对短角色场面、产品概念或音视频实验可能就够了。
你需要一次连续的 16–30 秒生成、超过 H3 所支持数量的参考素材,或 Topview 上可用的选定片段编辑工作流。
一个工作区。两种不同的创作方式。
更长的场面和参考很多的制作,选 Seedance 2.5。更短的音视频、一条可用的 2K 输出路径,或符合条件的开放权重集成,可以了解 MiniMax H3。
核对设置,检查积分,再按你的实际创作任务对比结果。



Pixar-style 3D animation, style locked to <<<Image3>>> , <<<Image1>>> and <<<Image2>>> — render everything exactly in their look. HAORAN, the young man from <<<Image1>>> , and XIAOYU, the young woman from <<<Image2>>> , run into each other on the sidewalk in front of the café from <<<Image3>>> , both breaking into surprised, warm smiles. XIAOYU says: "Haoran? What a coincidence!" He laughs: "First time seeing you since graduation!" She replies: "Come on in, coffee's on me." He answers: "Let's go." They walk up the steps together; HAORAN pulls the right door leaf outward by its brass handle — hinges on the outer edge — and holds it open as XIAOYU steps inside, then follows her in as the door swings shut behind them. Soft rounded stylized character proportions, subsurface scattering on skin, large expressive animated eyes, warm golden-hour cinematic lighting, gentle squash-and-stretch motion. The camera tracks them from across the street with a slow push-in. Dialogue in English with accurate lip sync, no subtitles. Avoid photorealistic faces, live-action look, 2D cel-shaded style, any style shift mid-clip, doors opening inward, sliding doors, or extra background characters entering frame.
Hyper-realistic wool-felt plush texture, cottagecore pastel palette, matching the reference image exactly. Fixed camera locked at this exact interior angle, no cuts, no camera movement. A hilarious and heartwarming 15-second road trip scene bursting with individual personality. [0-4s] The shiba inu driver is having the time of her life — head bobbing to music, singing loudly with mouth wide open, one paw drumming the steering wheel. The watermelon-hat cat films a selfie video, the duck photobombs by suddenly stretching its neck into frame with a loud honk, the cat rolls its eyes dramatically. The raccoon shakes its iced drink like a cocktail, ice cubes rattling. [4-8s] Sharp left turn! A comedic chain reaction: the croissant hat slides off the grey cat's head and lands on the bonnet kitten's face, covering its eyes — the kitten flails its tiny paws in confusion. The poodle slides across the back seat and squishes the bichon against the window, both bursting into laughter. The duck slides sideways with a deadpan expression, completely unbothered, like it has seen everything. [8-12s] Bumpy road! Everything bounces — a grape pops out of the picnic basket, flies through the air, and the raccoon catches it in its mouth mid-bounce, looking extremely proud of itself. One kitten in the pink basket gets bounced awake, grumpily grabs the edge of the basket, gives everyone an annoyed squint, then dramatically flops back down into the basket. The goldfish leaps out of the sloshing water and splashes back in. [12-15s] The road smooths out into warm golden light. All animals perk up, bright-eyed and alert, turning toward the camera with big open-eyed smiles — the shiba driver glances back with a wide grin and a thumbs-up paw, eyes sparkling; the croissant cat retrieves its hat, puts it on backwards without noticing, and blinks curiously at the camera; the bonnet kitten hiccups a tiny flower petal and giggles with wide shining eyes. Everyone stays awake, lively and energetic, eyes wide open, looking around cheerfully as sunset light glows on their fur. Every animal has exaggerated theatrical acting with natural fur physics — wide eyes, flattened ears, wagging tails, puffed cheeks, dramatic reactions, while rendering stays fully realistic plush-felt style throughout. All characters keep their eyes open and alert throughout, except the basket kittens. Blurred mountain scenery streams past the windows, tree shadows flicker across the interior.

音频节拍参考
The five dancers from <<<Image1>>> , keeping each dancer's exact face, hairstyle and outfit, perform street dance in the graffiti plaza, choreographed to the music in <<<Audio1>>> , every move landing on the beat. Formation change is required: starting in the V-formation from <<<Image1>>> , midway through the clip all five dancers travel to new positions — the center dancer moves to the back as the other four sweep forward into one horizontal line, finishing the routine in this new formation. Camera movement mirrors the music's rhythm and energy. Continuous single take. Use the ORIGINAL audio from <<<Audio1>>> unchanged; audio reference only, not visual reference. Avoid generated music, added sound effects, cuts, face distortion, identity swapping, subtitles, dancers standing in place for the whole clip.
A fifteen-second continuous shot, Edwardian pastoral watercolor style, soft riverside light. A water rat sculls a small rowing boat, painted blue outside and white within, along a gentle river, with a mole seated in the stern, exactly as the original text describes: "Believe me, my young friend, there is nothing—absolute nothing—half so much worth doing as simply messing about in boats. Simply messing," he went on dreamily: "messing—about—in—boats; messing—" "Look ahead, Rat!" cried the Mole suddenly. It was too late. The boat struck the bank full tilt. The dreamer, the joyous oarsman, lay on his back at the bottom of the boat, his heels in the air. "—about in boats—or with boats," the Rat went on composedly, picking himself up with a pleasant laugh." — Kenneth Grahame, The Wind in the Willows (1908), Chapter I, Project Gutenberg #27805 (Scribner 1913 ed.)


Toy-world giant boss battle CG, 15 seconds, one continuous take. Input references: <<<Image1>>> for character appearance — TOPVIEW BOT PRO exact costume and palette <<<Image2>>> for boss appearance — TITAN UNIT-01 final giant form, exact design and palette References are design sheets only: do NOT render the sheet layout, labels, swatches, scale panel or grey background. Extract character design only. Character lock (match <<<Image1>>> exactly, no changes): A small agile girl TOPVIEW BOT PRO with long silver-white hair in loose low twin tails, pale mauve eyes, calm focused expression. Yellow cat-ear cap with black inner ears and black cat face emblem, black and hot pink headphones with heart marks, heart earrings. Oversized yellow bomber jacket with black cat patches, PRO badge, X badges and pink hearts, white-black cropped top, pink belt with heart buckle, black shorts, pink heart mark on her thigh, mismatched socks, chunky yellow platform sneakers with pink soles. She carries a huge yellow and black toy cannon with pink handle and X markings, connected by pink cables to a back tank. Color palette locked: yellow #FFD400, hot pink #FF4D79, black #1A1A1A, white, grey. Boss lock (match <<<Image2>>> exactly): A hundred-meter toy titan mech in its tall heroic giant form, glossy injection-molded plastic and diecast materials, dense overlapping armor plates with oversized screws and molded seams. Deep grey and matte blue armor, blue helmet with hot pink crest raised above broad flat shoulder decks, glowing yellow eye visor, shoulder missile pods with yellow-black warning stripes and pink missile tips, pink-yellow energy lines, glowing chest core with pink X emblem set into layered chest armor plates forming flat ledges at varying heights — standable plates, deep seams, no man-made stairs or ladders. The titan is an active aggressive fighter: its head tracks her constantly, its fists strike and grab at her throughout the fight, servos whining with every motion. Environment: giant toy city battlefield — building blocks, gift boxes, toy trains, a Ferris wheel, teddy bears, plastic debris; realistic grey-white skyscrapers behind, bright blue sky, floating confetti. Timeline: 0.0–3.0s | Face-off Opening frame: seen from a low angle behind her — TOPVIEW BOT PRO stands with her back to the camera on a wide flat armor ledge beside the glowing X core, tiny against the titan's mass, facing up toward the helmet far above. Wind blasts through her twin tails and yellow jacket. She levels her cannon at the distant glowing visor. The titan's head tilts down, visor locking onto her and narrowing to a bright slit; both shoulder missile pods snap open with mechanical clicks. Camera: low-angle over-the-shoulder push-in past her toward the titan's head — her back to camera in the near foreground, the helmet towering above, shallow depth of field. 3.0–8.0s | Homing missile gauntlet The shoulder pods fire: missiles arc high into the sky trailing white smoke, then curve sharply downward and home in on her, chasing her path. She sprints up the titan's chest across overlapping armor plates, vaulting ledge to ledge as the homing missiles slam into the armor just behind her footsteps, each impact exploding into colorful plastic debris. The titan's left fist sweeps across its own chest to swat her — she slides under the passing fist without breaking stride, then fires once on the run, destroying the last missile behind her. Camera: fast side-tracking shot rising with her, wide lens, heavy motion blur, the sweeping fist and missile impacts crossing the foreground. 8.0–15.0s | Face-to-face point-blank finish As she reaches the shoulder deck, the titan's right hand rushes in to grab her — she leaps off the closing fingers, using the hand as a springboard, arcs through the air IN FRONT of the titan's face, and lands on the narrow brow ledge directly above the glowing yellow visor. She faces the titan's face head-on — her body turned OPPOSITE to the direction the titan faces, her back to the open sky, the huge visor glowing right in front of her. She presses the cannon muzzle horizontally against the center of the visor. The weapon charges yellow-pink energy for one beat while the titan thrashes — then she fires the beam horizontally straight INTO the face. The visor shatters inward, colorful energy bursts through every helmet seam, and the titan breaks apart into thousands of flying toy blocks and plastic pieces. She kicks off the collapsing face through the falling debris, landing on a giant tumbling block, still running forward. Camera: following her leap around to the FRONT of the titan's face, holding a frontal close shot — girl and giant face in the same frame, facing each other — during the charge and firing, then pulling wide as the mech shatters; follow her momentum through the debris. End frame: cut during her run across the falling blocks — no victory pose, no freeze frame, no fade. Style: AAA game boss battle trailer, stylized anime-inspired 3D CG, glossy toy materials, Ratchet & Clank + Astro Bot energy, extreme scale contrast between the small girl and the colossal mech, high saturation, cinematic rim light, physical debris simulation, motion blur, 24fps anamorphic widescreen. Audio: fast electronic rock, heavy drums, toy-box chimes, servo sounds, missile launch whooshes, homing missile whistles, sliding metal scrapes, cannon charge whine, massive beam impact, cascading plastic destruction; music cuts off mid-action. Avoid static camera, slow pacing, scene reset, character facing the camera in the opening shot, girl facing the same direction as the titan's head, girl standing on top of the head firing outward into empty sky, beam firing away from the titan, missiles flying away in straight lines, missiles appearing from off-screen without being launched, passive motionless boss, staircases, man-made steps, ladders, handrails, character redesign, hair or costume color change, blue jacket, bunny ears, weapon redesign, chibi proportions on the mech, rendering the reference sheet layout or text, extra limbs, distorted face, bad hands, subtitles, watermark, UI, gore, desaturated colors, slow-motion ending.