anycap-gemini-omni-video-edit

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

AnyCap Gemini Omni Video Edit

AnyCap Gemini Omni 视频编辑

Use this skill to edit short videos with

gemini-omni-flash-preview

via AnyCap. It is optimized for practical one-shot video edits where a source video should stay mostly unchanged and one element should change.

For CLI syntax, authentication, model discovery conventions, and delivery options, read the

anycap-cli

skill when needed.

使用此技能通过AnyCap结合

gemini-omni-flash-preview

编辑短视频。它针对实用的单次视频编辑场景进行了优化，这类场景中源视频大部分内容保持不变，仅修改单个元素。

如需了解CLI语法、身份验证、模型探索规范和交付选项，可按需查看

anycap-cli

技能文档。

Evidence To Verify

验证依据

Gemini Omni Flash Preview is a preview model, so verify the live surface before running:

bash

anycap status
anycap video models gemini-omni-flash-preview
anycap video models gemini-omni-flash-preview schema --operation generate --mode edit-video

Also check official Google docs when model behavior matters:

Gemini Omni Flash guide: https://ai.google.dev/gemini-api/docs/omni
Gemini API video overview: https://ai.google.dev/gemini-api/docs/video
Gemini Omni Flash model card: https://ai.google.dev/gemini-api/docs/models/gemini-omni-flash

Treat the AnyCap live schema as the source of truth for runnable CLI parameters. Treat official Google docs as the source of truth for model behavior, limitations, and prompting guidance.

Gemini Omni Flash Preview是一个预览模型，因此运行前请验证实时界面：

bash

anycap status
anycap video models gemini-omni-flash-preview
anycap video models gemini-omni-flash-preview schema --operation generate --mode edit-video

当模型行为至关重要时，还请查阅Google官方文档：

Gemini Omni Flash指南：https://ai.google.dev/gemini-api/docs/omni
Gemini API视频概述：https://ai.google.dev/gemini-api/docs/video
Gemini Omni Flash模型卡片：https://ai.google.dev/gemini-api/docs/models/gemini-omni-flash

可运行的CLI参数以AnyCap实时模式为权威依据；模型行为、限制和提示词指导以Google官方文档为权威依据。

Fit Check

适用场景判断

Use Gemini Omni Flash Preview when the task is:

edit an existing short video from natural language
preserve identity, camera motion, background, hands, lighting, timing, and audio while changing one visual element
swap a product or prop using one or more reference images
remove or hide a visible object
make narrow scene-level edits such as relighting, restyling, or text/sign changes
run fast preview iterations before considering a higher-control video model

Consider another video model or direct provider API when the task requires:

scene extension, first/last-frame interpolation, or last-frame control
precise multi-turn state using Google's
```
previous_interaction_id
```
strict multi-video reasoning
long-form generation beyond the live AnyCap schema
output quality/resolution not exposed by the live AnyCap schema

当任务符合以下条件时，使用Gemini Omni Flash Preview：

通过自然语言编辑现有短视频
在修改单个视觉元素的同时，保留人物身份、镜头运动、背景、手部动作、光线、时长和音频
使用一张或多张参考图片替换产品或道具
移除或隐藏可见物体
进行局部场景编辑，如重新打光、风格重塑或文字/标识修改
在考虑使用更高可控性的视频模型前，快速进行预览迭代

当任务需要以下功能时，请考虑其他视频模型或直接调用服务商API：

场景扩展、首尾帧插值或末帧控制
使用Google的
```
previous_interaction_id
```
实现精准的多轮状态管理
严格的多视频推理
超出AnyCap实时模式支持的长视频生成
实时模式未提供的输出质量/分辨率

Workflow

工作流

Inspect inputs. Confirm the source video path or URL, reference images, desired edit, aspect ratio, duration, and output filename.
Check schema. Run the live schema command above and use only supported params.
Choose references. Prefer one source video. Use reference images for product, subject, style, or material details. Avoid multi-video prompting unless the live docs and task strongly justify it.
Write a narrow English prompt. English is the safest default. Make the target edit explicit and tell the model what to preserve.

Generate. Use

anycap video generate

with

--mode edit-video

and a descriptive

-o

QA the video. Use
```
ffprobe
```
, frame samples, and
```
anycap actions video-read
```
; then manually review motion, hands, object stability, and unintended changes.
Retry narrowly. If the output drifts, shorten the prompt, isolate the edit, and add only the missing constraint.

检查输入信息。确认源视频路径或URL、参考图片、期望的编辑内容、宽高比、时长和输出文件名。
查看模式规范。运行上述实时模式命令，仅使用支持的参数。
选择参考素材。优先使用单个源视频。使用参考图片提供产品、主体、风格或材质细节。除非实时文档和任务明确要求，否则避免多视频提示。
编写精准的英文提示词。英文是最稳妥的默认选项。明确说明目标编辑内容，并告知模型需要保留的元素。
生成编辑视频。使用
```
anycap video generate
```
命令，搭配
```
--mode edit-video
```
参数和描述性的
```
-o
```
输出路径。
视频质量验证。使用
```
ffprobe
```
、帧采样和
```
anycap actions video-read
```
工具；然后手动检查运动效果、手部动作、物体稳定性和意外修改。
针对性重试。如果输出结果偏离预期，缩短提示词，聚焦编辑目标，仅补充缺失的约束条件。

Command Pattern

命令范式

bash

anycap video generate \
  --model gemini-omni-flash-preview \
  --mode edit-video \
  --prompt "<narrow English editing instruction. Keep everything else the same.>" \
  --param videos=./source.mp4 \
  --param images=./reference.png \
  --param aspect_ratio=16:9 \
  --param duration=10 \
  --param resolution=720p \
  --param format=mp4 \
  -o ./edited-output.mp4

Rules:

Always pass the source video with
```
--param videos=...
```
.
Use JSON array syntax when the schema supports multiple references:

bash

--param images='["./product-front.png","./product-side.png"]'

Do not repeat the same
```
--param images=
```
key expecting it to append.
Keep file names versioned:
```
brief-v1.mp4
```
,
```
brief-v2.mp4
```
,
```
brief-final.mp4
```
.
For URLs or local files, let AnyCap upload/resolve references automatically.

bash

anycap video generate \
  --model gemini-omni-flash-preview \
  --mode edit-video \
  --prompt "<精准的英文编辑指令。保持其他所有内容不变。>" \
  --param videos=./source.mp4 \
  --param images=./reference.png \
  --param aspect_ratio=16:9 \
  --param duration=10 \
  --param resolution=720p \
  --param format=mp4 \
  -o ./edited-output.mp4

规则：

必须通过
```
--param videos=...
```
传入源视频。
当模式支持多个参考素材时，使用JSON数组语法：

bash

--param images='["./product-front.png","./product-side.png"]'

不要重复使用相同的
```
--param images=
```
参数来追加素材。
对文件名进行版本管理：
```
brief-v1.mp4
```
,
```
brief-v2.mp4
```
,
```
brief-final.mp4
```
。
对于URL或本地文件，让AnyCap自动上传/解析参考素材。

Prompt Formula

提示词公式

Use this structure for edits:

text

<Action only on target>. Use <reference image(s)> as the exact reference for <product/subject/style/material>. Preserve <identity, hands, body, clothing, background, camera motion, lighting, shadows, audio, timing>. Make <physical integration details>. Keep everything else the same. Do not <common failure modes>.

Good edit prompts are short but specific. For targeted editing, include "Keep everything else the same." Put negatives in the regular prompt because model-level negative prompts are not exposed for this workflow.

For scenario templates, read references/scenarios.md. For QA and retry patterns, read references/qa.md.

编辑时使用以下结构：

text

<仅对目标元素执行操作>。使用<参考图片>作为<产品/主体/风格/材质>的精确参考。保留<身份、手部动作、身体、服装、背景、镜头运动、光线、阴影、音频、时长>。完善<物理融合细节>。保持其他所有内容不变。不要<常见错误操作>。

优质的编辑提示词应简短但具体。针对目标编辑，务必包含“保持其他所有内容不变”。由于此工作流未提供模型级负向提示词，需在常规提示词中加入否定表述。

场景模板请查看references/scenarios.md。 QA和重试范式请查看references/qa.md。

Practical Defaults

实用默认设置

Default to one source video and one product/reference image.
Match
```
aspect_ratio
```
to the source video unless the user explicitly wants a crop.
Use the source video's duration when it fits the live schema; otherwise choose the nearest supported duration and tell the user.
Use a stable, descriptive output path in the workspace.
Ask for the smallest possible edit first. Big combined edits are more likely to change identity, hands, background, or camera motion.

默认使用一个源视频和一张产品/参考图片。
除非用户明确要求裁剪，否则
```
aspect_ratio
```
与源视频保持一致。
当源视频时长符合实时模式要求时使用原时长；否则选择最接近的支持时长并告知用户。
在工作区使用稳定且具有描述性的输出路径。
先请求最小幅度的编辑。大幅组合编辑更可能改变人物身份、手部动作、背景或镜头运动。

Output Expectations

输出预期

Return:

the local output video path
the exact command used or a concise command summary
the QA checks run and what they found
any caveats, such as identity drift, hand artifacts, flicker, audio changes, or schema limits

If the edit is not good enough, provide a concrete next prompt rather than a vague "try again."

返回内容包括：

本地输出视频路径
使用的完整命令或简洁的命令摘要
执行的QA检查及结果
任何注意事项，如身份偏差、手部伪影、闪烁、音频变化或模式限制

如果编辑效果不理想，请提供具体的下一轮提示词，而非模糊的“重试”建议。