视频｜Vidu3 Turbo 文生/图生/首尾帧

The Vidu3 Turbo node is the sibling of the pack's reference-video node, and the difference is in the title: Turbo. Where the reference node specializes in identity-consistent multi-image generation, this one is the fast, general-purpose Vidu3 entry point - three modes in one node: pure text-to-video, image-to-video, and start-frame + end-frame (首尾帧) interpolation. Under the hood it's the viduq3-turbo model over Tikpan's /ent/v2/text2video, img2video, and start-end2video endpoints.

Reach for it when you want a quick product demo, a short brand clip, or a loop from a still image - the "fast and cheap" slot of the Vidu3 family. Vidu is a real and active model line (141 threads in the community corpus), and Turbo is specifically the cost-conscious member: the defaults in this node (5 seconds, 720p) are tuned to not bankrupt you.

How it works

The 生成模式 dropdown is the whole architecture:

• 文生视频 (text2video) - prompt-only, the default.
• 图生视频 (img2video) - requires 首帧图 (first-frame image) to drive the video.
• 首尾帧 (start-end2video) - requires both 首帧图 and 尾帧图; the model animates between them.

For every mode the node builds the payload, submits, polls ent/v2/tasks/{task_id}/creations every 查询间隔秒数 (default 8) until done or 最长等待秒数 (default 1200) expires, then downloads the video. 运动幅度 (auto/small/medium/large) is the Vidu movement_amplitude parameter - small for subtle product shots, large for energetic scenes. 音画同步 (default on) generates matching audio; 音频类型 constrains it; 错峰生成 (off_peak) trades speed for a lower price.

Inputs and outputs that matter

Beginner flow: API_密钥 → pick 生成模式 → write the 生成指令 (English recommended) → set 视频时长 (1–16s), 清晰度 (540p/720p/1080p), 画面比例 (16:9/9:16/1:1/4:3/3:4) → run. If you picked an image mode, feed the first-frame (and for start-end, last-frame) image - the mode won't do anything useful without it. Outputs: 📁本地保存路径, 🆔任务ID, 🔗视频云端直链, 📄完整日志, 🎬视频输出 (VIDEO).

Installing

Pack install as usual - Manager search "Tikpan", or git clone https://github.com/htrert/ComfyUI-Tikpan-Pro into custom_nodes, restart. Key from tikpan.com. No local models.

Common issues

• Picked img2video, forgot the image. The node needs 首帧图 for image modes; without it you get an error or an empty result (with 跳过错误 on). Mode/image mismatches are the most common stumble.
• 首尾帧 that don't match. Start and end frames should share composition, subject, and framing - the model interpolates between them, so wildly different frames produce a trippy morph rather than a coherent clip. Keep them visually consistent (character-consistency.md).
• Renders run long. 16s at 1080p isn't quick even on Turbo. If you time out, grab the task ID from the log and collect the finished job with the Task Fetcher node - don't resubmit and pay twice.
• Audio comes out of nowhere. If you don't want generated ambient audio, flip 音画同步 off. It defaults on and it's a separate downstream step in the pipeline.
• Determinism. Seed is hidden here too (the pack keeps it out of the UI because it isn't reliably reproducible through the relay). Reproduce by reusing the exact same prompt + frames, not by chasing a seed number.