Showcase案例展示

All three are official Mini output: authentic-feeling talking heads, handheld camera moves, and ambient room tone are all present. A 15-second 720p vertical clip costs roughly $1.2, meaning a $20 monthly budget can produce 16 TikTok/Reels-ready clips. Worth noting: the handheld shake and the presenter's tone in the real-estate clip feel exactly like a genuine vlog. 这三支全是 Mini 官方直出：真人感口播、手持运镜、环境收音全都在。一支 15 秒 720p 竖屏 ≈ $1.2，也就是说 $20 的月额度能出 16 条 TikTok/Reels 素材。看点：第一支的手持晃动感和地产讲解的口播情绪，完全是真人 vlog 的质感。

Upload a classical oil painting as @Image1, and Mini turns the whole painting into a breathing, moving shot — the figure's movement and the flickering candlelight are all filled in by the model. This verified third-party reference-to-video example runs on the Mini tier: even the smaller model can revive an asset or animate an old photo. 上传一张古典油画当 @Image1，Mini 就把整幅画变成一段有呼吸感的动态镜头——人物动作、烛光摇曳全是模型补出来的。这是经过验证的第三方 reference-to-video 案例，使用 Mini 档：小模型也能做素材复活、老照片动化。

For the same prompt, Mini's composition, camera work, and pacing are essentially identical to 2.0 — the only differences are resolution and detail fidelity. So the right move is: use Mini at 480p ($0.04/sec) to quickly generate 3–5 versions and pick a direction, then switch the winning version to Seedance 2.0 at 4K for the final render. The trial-and-error cost for a 15-second final cut drops from $11.7 (testing directly at 4K) to $0.6 per attempt. 同一句 prompt，Mini 的构图、运镜、节奏和 2.0 基本一致，差的只是分辨率和细节精度。所以正确姿势是：用 Mini 480p（$0.04/秒）快速生 3–5 个版本挑方向，选中的那版换 Seedance 2.0 拉 4K 出正片。一条 15 秒正片的试错成本从 $11.7（4K 直接试）降到 $0.6 一次。

An official 8-second battlefield FPV shot: the explosions, debris, and the rush of air from the low-altitude flyby are all generated by the model alongside the visuals. Physical accuracy and camera stability under fast, complex motion is where Veo pulls ahead of other models. 官方 8 秒战场 FPV 镜头：爆炸、碎屑、低空掠过的气流声全部是模型和画面一起生成的。复杂快速运动下的物理与运镜稳定性，是 Veo 和其他模型拉开差距的地方。

An official example: a static image of "a monkey and a polar bear recording a podcast" → the two characters start talking, with distinct voices, synced lip movement, and even the room tone from the mic sounding right. This is the path for having a mascot or IP character speak your brand's story. Playground also lets you add a last-frame image to control how the shot ends. 官方示例：一张「猴子和北极熊做播客」的静态图 → 两个角色开口对谈，声线不同、口型同步、连麦克风的房间声都对。吉祥物/IP 角色开口讲品牌故事，用的就是这条路。Playground 里还支持再加一张尾帧，控制镜头结束的画面。

Watch the input box in this official screen recording closely: 8 reference images are dragged straight in, and the prompt is written along a timeline covering 30 seconds of blocking, dialogue ("Everyone, enjoy this party!"), and camera moves — generating the entire party scene in one pass, with no cut points in between. That's what "no more stitching clips together" means. 注意看这支官方录屏里的输入框：8 张参考图直接拖进去，prompt 按时间轴写满 30 秒的走位、台词（"Everyone, enjoy this party!"）和运镜，一次生成整段宴会戏——中途没有任何剪辑点。这就是「不用再拼片段」的意思。

Same bedroom, same camera angle, same lipstick — on the left, an English-speaking creator says "I'm trying shade eleven today"; feeding in just one new model reference image, the right side becomes a Japanese-speaking creator saying the same line (with the Japanese subtitles updated to match). Not a single prop in the room changes. For cross-border e-commerce: one video turns directly into N market-localized versions. 同一间卧室、同一个机位、同一支口红——左边英文博主说 "I'm trying shade eleven today"，官方只喂了一张新模特参考图，右边就变成日语博主说同一句话（日本語字幕同步变）。房间里每一件道具都没变。对电商出海：一条片直接翻成 N 个市场版本。

This 30-second "Tianwen-1 Mars mission" explainer is the official flagship pipeline demo: Seedream 5.0 Pro first generates the infographics and keyframes, which are then fed into Seedance 2.5 to stitch together a continuous film — combining image-layout precision with video motion. You can already run this pipeline today using Seedream + Seedance 2.0's image-to-video, then upgrade seamlessly once 2.5 opens up. 这支「天问一号火星探测」30 秒科普片是官方主推的 pipeline demo：Seedream 5.0 Pro 先生成信息图与关键帧，再喂给 Seedance 2.5 串成连续影片——图的排版精度 + 视频的运动能力。这条路线现在就能用 Seedream + Seedance 2.0 图生视频先跑通，2.5 开放后无缝升级。

The generated skateboard magazine cover (left) can be split apart on demand: the headline text, figure, stickers, and background each become independent layers (right) — like getting the source file instead of a flattened image. Changing the tagline doesn't require regenerating the whole image. The official demo video is above. 生成的滑板杂志封面（左）能按需求拆开：标题字、人物、贴纸、背景各自独立成层（右）——像拿到源文件而不是一张死图。改标语不用重生成整张图。官方演示视频在上方。

An official example: circle a squirrel and draw an arrow dragging it to a tree stump, highlight a bride's dress hem and annotate a color change, box an empty area and write "add a painting" — the model executes every annotation in one pass, with the perspective, lighting, and style of each change automatically matching the original image. Zoom in on the right image to see the squirrel's nest on the stump and the newly added wedding portrait. 官方案例：在原图上圈住松鼠画箭头拖到树桩、高亮新娘裙摆标注换色、框选空白处写「加一幅画」——模型一次执行所有标注，每处修改的透视、光影、风格都自动贴合原图。放大看右图树桩上的松鼠窝和新增的婚礼画像。

Drop in the sketch layout plus photos of the sofa, rug, wall art, and lighting fixtures, and Seedream understands exactly where each item belongs, producing a photorealistic image with completely unified lighting, color, and perspective — the sofa is still that sofa, the rug is still that rug. A concept image for a pitch goes from half a day to one minute. 把线稿构图、沙发、地毯、挂画、灯具的素材照全部丢进去，Seedream 读懂每件东西该在哪，输出一张光线、色调、透视完全统一的实景图——沙发还是那张沙发，地毯还是那块地毯。提案概念图从半天缩到一分钟。

Give a HEX color code directly in the prompt, and the specified object changes to exactly that color — the fabric texture, lighting, and surrounding environment all stay untouched. Brand-color mockups, multi-color e-commerce SKU shots — one sentence, one image. 直接在 prompt 里给 HEX 色号，指定对象就换成精确的那个颜色——绒布纹理、光影、周围环境全部不动。品牌色打样、电商 SKU 多色图，一句话一张。

Left: a complete infographic — data visualization, flowcharts, mixed text and image — generated in one pass, with small text still legible. Right: a 16-panel ad storyboard straight from one prompt — unboxing, close-up, street shot, water splash, closing slogan, the full shot vocabulary — usable directly as a shooting script or fed into Seedance as a reference. 左：数据可视化、流程图、图文混排的完整信息图一次生成，小字依然清晰。右：产品广告的 16 格分镜 storyboard 一句话直出——开箱、特写、街拍、水花、收尾 slogan 全套镜头语言，直接当拍摄脚本或喂给 Seedance 当参考。

Same layout, but the brush-lettering headline is redrawn to fit each language — the Chinese calligraphy feel and Arabic script connections are both correct. Beyond Chinese and English, 14 more languages are supported (Malay, Thai, Japanese, Korean, Spanish...). For us: one key visual directly produces localized versions for the entire Southeast Asian market. 同一个版式，标题笔刷字效跟着语言重新绘制——中文书法感、阿拉伯文连笔都对。中英之外新增 14 种语言（马来文、泰文、日韩、西语…）。对我们：一套 KV 直接出全东南亚市场的本地化版本。

Google's official example shows all 14 supplied characters and objects across the top, then the generated farm scene below. The hard part is not merely placing many items: each identity remains recognizable while scale, pose, lighting and interaction are rebuilt into one playful composition. Playground now accepts up to 14 references for this model and routes multi-reference requests through its editing mode. Google 官方案例把 14 个输入角色与物件全部列在上方，下面是合成后的农场场景。难点不只是把很多东西塞进去，而是每个身份仍可辨认，同时把比例、姿势、光线与互动重新组织成统一画面。Playground 现在为这个模型开放最多 14 张参考，并让多参考请求走编辑模式。

An official edit example: two independent reference images → one composited image, with both subjects' features and materials fully preserved. This is how it handles product + scene, person + outfit composites. 官方 edit 示例：两张独立参考图 → 一张合成图，两个主体的特征、材质都被完整保留。这就是它做产品 + 场景、人物 + 服装组合图的方式。

A verified third-party headline example showing extremely fine-grained imagery and typography. Zoom in on every line of text: this is exactly where other models most often slip up, and where GPT Image 2 is especially reliable. 经过验证的第三方 headline 示例，展示极度精细的图像与排版。放大看图里的每一行字：这是其他模型最容易露馅、GPT Image 2 特别稳定的地方。

An official example: a bare-vehicle photo + a brand design file → one finished photo of a fully wrapped bus ad. The graphics correctly warp along the curve of the vehicle body, and the text stays crisp. For client pitches: mockups of a logo/key visual applied directly to a vehicle, storefront, packaging, or uniform — one image per minute. 官方案例：素车照 + 品牌设计稿 → 一张涂装完成的巴士广告实拍。图形沿车身曲面正确变形、字没有糊。对客户提案：logo/KV 直接「上」到车身、店面、包装、制服的效果图，一分钟一张。

An official example: a late-night convenience-store mystery radio drama — multi-character dialogue, footsteps, a door chime, and background sound are all generated in one pass, not stitched together from separate tracks. 65 seconds costs $0.16. Listen once and you'll understand what "cinematic audio generation" means. 官方示例：深夜便利店悬疑广播剧——多角色对白、脚步声、门铃、背景音全部一次生成，不是分轨拼的。65 秒成本 $0.16。听一遍就明白「电影级音频生成」是什么意思。

An official example: first upload two reference voices separately (Dex / Priya), then in the same prompt use @Audio1 / @Audio2 to name them and have the two voices have a conversation — the clip playing at the bottom is dialogue generated with these two cloned voices. This is exactly the same workflow as "upload reference audio → reference it with @" in Playground. Apply the same trick with Chinese/English/Malay copy, and your brand voice stays the same person across markets, with no need to re-hire a voice actor. 官方示例：先分别上传两段参考人声（Dex / Priya），然后在同一段 prompt 里用 @Audio1 / @Audio2 指名让两个音色对话——最下面播放的就是用这两个克隆音色生成的对白。这跟 Playground 里「上传参考音频 → 用 @ 引用」的操作方式完全一样。同一招换成中/英/马来文文案，品牌声音跨市场保持同一个人，不用重新请配音。

This capability is documented in the official API docs (1 reference image, mutually exclusive with voice cloning), but we haven't been able to find an official playable example, so we're not putting up a fake demo here — the result is unverified, and we'd suggest logging into the dashboard and trying it yourself. Once you get a result you're happy with, we'll add the real audio to this card. 这项能力在官方 API 文档里有写（1 张参考图，与音色克隆二选一），但我们没能找到官方发布的可播放案例，所以这里先不放假的示范——效果未经验证，建议登入后台自己试一次。等你测出满意的结果，我们就把真实音频补进这张卡片。

This is a newly added model in Playground (DeepFilterNet3), separate from Seed Audio — upload a recording made in a noisy environment (say, an interview recorded in a coffee shop), and with no prompt required, it outputs a clean voice track with the background noise removed. It's already selectable in the tool; we haven't done a real before/after comparison yet, so once you log into the dashboard and run it once, we'll add a demo clip here. 这是刚接入 Playground 的新模型（DeepFilterNet3），跟 Seed Audio 是不同的模型——上传一段吵杂环境录的音频（比如咖啡店里录的采访），不用写 prompt，直接输出去掉背景噪音后的干净人声。工具里已经能选到，我们还没做真实的降噪前后对比，你登入后台生成一次之后我们把示范音频补上。