StreamingT2VRunLongStepVidXTendPipelinePromptTravel

StreamingT2VRunLongStepVidXTendPipelinePromptTravel is the pack's showpiece: the VidXTend long-extension loop, but with a prompt schedule instead of a single fixed prompt. Your video starts as one scene and gradually becomes another - a cyclist crossing a city, a street transitioning from day to night - all in one continuous long generation. The README's cyclist demo (wf_prompttravel.json) is this node's calling card.

Everything you've seen in the other VidXTend run nodes still applies: the seed short_video, the 8-frame extension loop, the appearance anchor, num_frames, num_steps, image_guidance. What's new is that the prompt input is a JSON schedule - keys are frame indices, values are the prompts to switch to at those points - and the node also schedules its reference frames, pulling a different ref_frames frame forward as the generation progresses.

How it works

Your prompt string is a JSON object body (keys quoted, values quoted): "0":"a cyclist riding through a city street", "100":"the same cyclist crossing a bridge at dusk", "200":"the same cyclist on a rainy night". The node wraps it in braces, parses it, and builds two schedules. For each 8-frame chunk it picks the prompt whose key is current, and it steps through your ref_frames as the video advances so the appearance anchor evolves with the scene instead of being frozen at frame zero. The output is a single continuous IMAGE where the scene transforms over its length.

Inputs that matter

• VidXTendPipeline - from StreamingT2VLoaderVidXTendModel.
• short_video - the seed IMAGE.
• prompt - the schedule, as a JSON object body. This is the input you'll spend your time on.
• ref_frames - reference frames the node steps through as the video progresses.
• num_frames (24), num_steps (50), image_guidance (9.0), seed (33), negative_prompt - the usual knobs.

Output is an IMAGE of num_frames frames.

Installing it

Same pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/chaojie/ComfyUI_StreamingT2V

or ComfyUI Manager → search StreamingT2V → install, restart. VidXTend auto-downloads. The usual heavy requirements.txt.

Where people get burned

Two classic mistakes. First, the schedule must be strict JSON - the node does json.loads, so a missing quote or trailing comma throws and kills the whole run. Keep it clean. Second, the frame indices in your schedule need to make sense for the loop: the node checks the schedule at chunk boundaries (multiples of 8), so "100" will be honored as the closest chunk at or after frame 100. And don't expect hard cuts - prompt travel is gradual by nature; the transition smears across nearby chunks. That's the feature, not the bug. If you want a scene to jump, this isn't the node.

The take

This is the most impressive node in the pack and the fiddliest. When prompt travel works, it produces the kind of continuous long video you can't get any other way from a 2024-era model. Budget your time for schedule debugging - the PromptTravelIndex node and the pack's LoadText/SaveText utilities exist to help you manage that schedule from disk, so use them. It's the demo reel, and it's worth the effort.