From One Product Photo to a Finished Ad
Hailuo H3 · Omni Reference Guide

1. The problem we open with

"I put my product photo in, and what came out was not my product."

Most AI video tools take a single still and animate it. Anything the photo does not show gets invented, and that is where a logo smears and a bottle changes shape. For anyone actually selling a product, that output is unusable.

What makes H3 different is stated plainly: it reads several assets at once, and you decide what each asset is responsible for. That is Omni Reference.

「我放进去的是我的产品照片，出来的却不是我的产品。」
以往多数 AI 视频工具只读取一张静态图并使其动起来。照片中没有的部分由模型自行想象补足，logo 因此糊掉、瓶身随之走形。对真正在卖货的人来说，这样的成品无法使用。
H3 的不同之处很直接：它一次读取多个素材，并且由你来指定每个素材负责什么。这就是 Omni Reference。

2. The one concept that carries everything

When you upload assets, the system labels them automatically — @image1, @image2, @video1, @audio1. You then call those labels in the prompt and assign a role to each.

It works like a cell address in a spreadsheet. You do not re-describe what sits in A1 every time; you just write A1.

上传素材后，系统会自动为其编号 — @image1、@image2、@video1、@audio1。随后在提示词中调用这些编号，并为每个素材分配角色。
其原理类似表格中的单元格地址：你不必每次重新描述 A1 里是什么，直接写 A1 即可。

Asset limits 素材上限

• Up to 9 images 图片最多 9 张
• Up to 3 video clips 视频最多 3 个
• Up to 3 audio files 音频最多 3 个
• 12 files maximum in one generation 单次生成最多 12 个文件
• Reference video and audio run 2–15 seconds each, 15 seconds in total 参考视频与音频各 2–15 秒，合计不超过 15 秒
• Output 5–15 seconds, 1440p at 24fps 成品 5–15 秒，1440p 24fps

3. Core rule — one job per asset

The most common failure is attaching four images and hoping. Write down what each asset does: "Image 1 sets the mood and film texture, Image 2 is the talent, Image 3 is the product."

最常见的失败是一次丢进四张图然后期待模型自行判断。请写明每个素材的职责：「图 1 决定氛围与胶片质感，图 2 是人物，图 3 是产品。」

What each asset type can carry 各类素材可承担的职责

• Images — identity, product geometry, location, palette, film texture
  图片 — 人物身份、产品形态、场景、色调、胶片质感
• Video — motion, camera path, edit rhythm, timing
  视频 — 运动、运镜路径、剪辑节奏、时间点
• Audio — voice character, cadence, rhythm
  音频 — 声音特质、语速、节奏

4. Write the prompt in five blocks

One long paragraph leaves the model guessing what to protect. Splitting it holds up far better.

1. Deliverable — what the clip is for, its length, aspect ratio
   交付物 — 用途、时长、画幅
2. Role map — what @image1 is, what @video1 is
   角色分配 — @image1 是什么，@video1 是什么
3. Shot — subject, action, environment, lighting, camera
   镜头 — 主体、动作、环境、光线、摄影机
4. Timeline — what happens first, what follows
   时间线 — 先出现什么，后出现什么
5. Preserve — what must not change from the references
   须保持 — 参考素材中绝对不可改变的部分

The fifth block decides whether a product ad is usable. And "keep it consistent" does nothing here. Name properties you can inspect in any frame: bottle silhouette, cap shape, label placement. For a person: facial structure, eye color, hair length. Then you can look at the result and judge whether it held.

第五块决定了一支产品广告能否使用。而「保持一致」这类表述在这里没有作用。请写出可以在任意一帧中用肉眼核对的属性：瓶身轮廓、瓶盖造型、标签位置；若是人物，则写面部结构、瞳色、发长。这样才能对着成品判断是否守住了。

5. Asset setups by product category

Apparel 服饰

• @image1 product, front, with stitching and logo legible
• @image2 model's face, fixed if the same person recurs
• @image3 the location
• Preserve: neckline shape, logo placement, fabric sheen
• @image1 产品正面（缝线与 logo 需清晰）／@image2 模特面部（同一人物则固定）／@image3 场景。须保持：领口形状、logo 位置、面料光泽

Accessories and objects 配饰与小物

• @image1 product, front
• @image2 same product at side or 45° — adds shape information
• @video1 the camera movement you want to borrow
• Preserve: silhouette, handle angle, engraving position
• @image1 产品正面／@image2 同一产品侧面或 45 度（补充形态信息）／@video1 想借用的运镜。须保持：轮廓、把手角度、刻印位置

Food and beverage 食品饮料

• @image1 packaging
• @image2 contents, close up
• @audio1 narration voice
• Preserve: package color, logo typeface, texture of the contents
• @image1 包装／@image2 内容物特写／@audio1 旁白人声。须保持：包装颜色、logo 字体、内容物质感

6. The working order we recommend

1. Start with the photo. If the logo is not legible in the source, it will not be legible in the output.
   先看照片。原图上 logo 不清晰，成品上也不会清晰。
2. Upload in order and check the labels. Numbers follow upload order.
   按顺序上传并确认编号。编号依上传顺序排列。
3. Write the role map before the shot description. Once roles are set, the description gets shorter.
   先写角色分配，再写镜头描述。角色确定后，描述自然会变短。
4. Test at 5 seconds first. Confirm direction before going to 15.
   先用 5 秒试跑，确认方向后再做 15 秒。
5. Change one instruction at a time. Two at once and you learn nothing.
   每次只改一处。同时改两处就无法判断哪一处起了作用。
6. Edit the part, do not re-roll the shot. A full regeneration also changes what was already working.
   做局部修改，不要整条重生成。整体重生成会连已经满意的部分一起改掉。

7. When you get stuck

The logo smears 字样糊掉

Check source resolution first, then make the preserve block specific — "logo typeface and letter spacing unchanged." If it still slips, add one more image where the logo is larger and give it its own role.

先检查原图分辨率，再把「须保持」写具体 —「logo 字体与字距不变」。若仍不稳定，可再加一张 logo 更大的图片并单独分配角色。

Color shifts when clips are joined 拼接后颜色跳变

Ask for the color change only. A full regeneration touches everything else.

只指定修改颜色。整体重生成会波及其他部分。

References contradict each other 素材之间冲突

Two images carrying different lighting will confuse the model. State which one wins. Hoping it infers your preference gives a different result every run.

两张图片带有不同光线时模型会混乱。请明确指出以哪一张为准。指望模型自行判断，每次结果都会不同。

The style is hard to put into words 风格难以用文字描述

Then do not. Attach one reference video instead. Handing over a clip is far more precise than describing it — some people write nothing about style at all and let the clip carry it.

那就不要描述。改为附上一段参考视频。交出片段远比文字描述精确 — 有人完全不写风格，只靠片段传达。

8. Copy-paste template

Replace the bracketed parts with your own product.

将方括号内容替换为你自己的产品。

[Deliverable] Instagram Reels product ad, 10 seconds, vertical 9:16

[Role map]
@image1 = the product. Its shape and logo are the reference.
@image2 = the location used as background.
@video1 = camera movement reference. Follow this rhythm.

[Shot]
The product sits at center frame and the camera moves in slowly.
Soft light enters from the upper left.

[Timeline]
Open on the full product, then close in toward the logo.

[Preserve]
The [product shape], [logo placement] and [package color] of @image1 do not change.
The lettering in the logo must stay legible exactly as in the source.

Add this line to the role map if you are using a voice file.

如需使用语音文件，请在角色分配中加入此行。

@audio1 = narration voice. Speak in this tone and pace.

To revise, send one line instead of rewriting the whole prompt.

修改时只需发送一行，而非重写整段提示词。

Change only [the product color] to [new color] and leave the rest of the frame as is.

9. Where to start

Open hailuoai.video, pick H3 from the model list. The mobile app does the same job and is easier if you are screen recording.

Just remember to switch Omni Reference on where you attach the assets. Leave it off and you are back to reading a single still.

打开 hailuoai.video，在模型列表中选择 H3。移动端 App 功能相同，若需录屏则更方便。
请务必在上传素材处打开 Omni Reference 开关。若未开启，就退回到只读取单张静态图的旧方式。