视频｜MiniMax Hailuo 视频生成

MiniMax Hailuo is one of the closed Chinese video services the community points at in "what made this?" threads, and this node brings three of its modes into one canvas: text-to-video, first-frame image-to-video, and first+last-frame video (bookend a clip with two images). It defaults to MiniMax-Hailuo-2.3, with MiniMax-Hailuo-02 as the alternative - and the tooltip drops the one rule you need to remember: first+last-frame mode requires switching to the 02 model. Leave it on 2.3 and the mode selector will fight you.

The pack's README mentions something worth knowing about this integration: the node auto-validates upstream limits like the 10-second/1080P cap and the tail-frame rule, so it refuses early instead of burning money on a request Hailuo would reject anyway. That's a small mercy in a pack where most failure handling is "check the log."

How it works

生成模式 picks text2video, image2video (needs 首帧图), or first+last-frame (needs both 首帧图 and 尾帧图). 提示词优化 defaults to on - Hailuo's own prompt_optimizer, which rewrites your prompt server-side and generally helps more than it hurts. The node submits to MiniMax's video endpoint, polls, downloads, and returns the standard five outputs: 🎬_视频输出, 📁_本地保存路径, 🆔_任务ID, 🔗_视频云端直链, 📄_完整日志.

The inputs that matter

• 生成模式 - text2video / image2video / first+last-frame.
• 模型版本 - 2.3 for text and first-frame; switch to 02 for first+last-frame.
• 生成指令 - subject, action, camera, style. The default cinematic template is a fine base.
• 视频时长 - 6s or 10s (the 10s/1080P cap is the auto-validation mentioned above).
• 分辨率 - 768P default; 1080P is slower and pricier.
• 首帧图 / 尾帧图 - required per mode, as above.

Install and first run

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

Restart (or Manager → "Tikpan"), paste your sk- key from https://tikpan.com, and for the simplest test set text2video + 6s + 768P. Once that renders, try first-frame with a good still - image-to-video is where Hailuo earns its reputation.

Common issues

The pack-wide failure family (401/402/429, task timeouts with recovery via 任务ID, charges-with-no-file - check 完整日志). Hailuo-specific: the model/mode mismatch is the classic newbie trap (first+last-frame on 2.3), and trying to exceed the 10-second/1080P ceiling gets you a rejection rather than a graceful crop - that's the auto-validation doing its job, so pick a compliant config. Preview problems are usually local, not the model's fault: if the VIDEO output won't render but 本地保存路径 has a file, your ComfyUI just lacks a video-preview extension.