视频｜Gemini Omni Flash 视频生成

TikpanGeminiOmniVideoNode is the pack's do-everything video node: it wraps Gemini's Omni Flash video endpoint (/v1/video/create through the Tikpan relay) and switches between five modes via a single dropdown - text-to-video, first/last-frame, reference-image padding, and Omni-Flash video editing. The README lists it as supporting 文生/图生/多参考/编辑/音频驱动 five ways of working; in practice that means you can draft a clip, constrain it with bookends, add subject references, and even edit an existing video, all from one node. Video editing is the unusual bit - most video nodes in this pack generate; this one can also revise.

How it works

The 生成类型 field maps 1=文生视频, 2=首尾帧, 3=垫图参考, 4=Omni-Flash 视频编辑. Types 1–3 submit an async generation job; type 4 takes a source video (视频URL or 本地视频 up to ~48MB inlined as data:video) and passes it to input_reference. Either way it's the standard async pattern - submit, poll 任务ID, download - and the five outputs are the familiar 本地保存路径, 任务ID, 视频云端直连, 完整日志, and 视频输出 (VIDEO).

The inputs that actually matter

• API_密钥 - sk- from tikpan.com.
• 生成指令 - the prompt. The default is a cinematic commercial brief - write for camera motion, lighting, and subject consistency.
• 模型 - omni-flash (default) or omni-flash-edit (the tooltip recommends the edit variant for video editing).
• 生成类型 - the mode switch above. Pick 4 for editing and provide a source video.
• 视频时长 - 5, 8, or 10 seconds.
• 画面比例 - 16:9 or 9:16.
• 清晰度 - 720p or 1080p; the tooltip notes 1080p engages enable_sample upstream.

Optional: 参考图1–3 (type 2 bookends max 2, type 3 padding max 3), 视频URL (for editing), 本地视频 (≤48MB inline), plus local knobs - 最长等待秒数 defaults to 1200s here, 查询间隔秒数 8s, and 校验HTTPS证书 defaults false on this node.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

restart (or Manager → "Tikpan"). No models to download - the 48MB inline-video cap is the only size concern, and it's a payload limit, not a disk one.

Where people get burned

• Mode confusion. All five modes share one node, and the reference-image limits differ per mode (2 bookends vs 3 padding). Read the tooltips before wiring images - a type-2 run with 3 images can reject.
• The 48MB inline wall. Editing a video bigger than that fails unless you use a public URL. Upload to a host and pass 视频URL.
• 1080p is a toggle with a cost. enable_sample is the upstream's upscale path - it's the premium lane, not the default.
• Cert off by default. 校验HTTPS证书=false again on this node. Flip it on if you can.

The short version: it's the Swiss-army video node - five modes, one install, and the only one in the pack that edits existing footage. The mode dropdown is both the feature and the trap; read it carefully.