Vidu文生视频

Vidu文生视频 - "Vidu text-to-video" - is the simplest node in the pack's video corner, and sometimes simple is exactly what you want. One prompt in, a Vidu video out, no images to gather, no frames to line up. It's the node to reach for when you just need a moving shot for a social clip, a concept, or a placeholder while you figure out the fancier reference modes. The display name says it all: this is the pure text-to-video entry in the Vidu family.

Mechanically it's the standard pack pattern: the prompt (plus your settings) goes to Mojie's endpoint under the vidut2vNode model (viduq2 backend), the video URL comes back, downloads to a temp file, and lands in the graph as a VIDEO type via VideoFromFile. Nothing local, no model files, credits per call - the whole API-wrapper bargain in one node.

The inputs that matter

• prompt - the whole show. Describe scene and motion; "a drone flies over a coastline at sunset" outperforms a bare "beach." Natural language, no SD keyword soup.
• duration - 1 to 8 seconds, default 5. Longer = more credits.
• resolution - 720p or 1080p. This is your main cost dial.
• movement_amplitude - auto / small / medium / large. small for calm, stable shots; large for dramatic camera moves.
• Size - 1:1, 9:16 (vertical, for TikTok/Reels), or 16:9.
• bgm - boolean; adds background music to the output. Nice for throwaway social drafts, and it's one less post-processing step.
• seed - normal diffusion seed; unlike ViduNode's reference mode, this one actually passes the value through.

Output is video, previewable and savable like any ComfyUI VIDEO.

Installing it

Same pack, same steps - worth memorizing once because they're identical for every node in comfyui-MJAPI-party:

cd ComfyUI/custom_nodes
git clone https://github.com/MoJIeAIGC/comfyui-MJAPI-party.git

or ComfyUI Manager search "mojieapi_party". Register at mojieaigc.com, paste the key into config.ini at the pack root (leave BASE_URL alone), restart ComfyUI. Dependencies are light - opencv-python is the heaviest new install, and there are no model downloads because the model lives on Mojie's servers.

Where people get caught

• The red error image. The pack renders HTTP failures as a red placeholder: 401 = key problem, 403 = balance. Text-to-video is metered and it's not cheap - check the balance before a batch, because video calls are where the KB's API-nodes essay says budgets die.
• Under-described prompts. T2V has no image to anchor it, so the prompt carries everything. "A cat" gives you a wandering cat; "a tabby cat walks toward the camera across a wooden floor" gives you a usable shot. The model follows instructions well, but it can't read your mind.
• Empty prompt. The field defaults to empty - running with it blank is a waste of a paid call, and most likely an error image on the other end.