StreamingT2VRunLongStep

StreamingT2VRunLongStep is stage two of the original-model pipeline: it takes the 16-frame seed video from the short step and runs the StreamingT2V autoregressive extension over it, producing a long 256×256 mp4. Feed it the stream_cli and stream_model from StreamingT2VLoaderStreamModel, a short_video IMAGE, and a prompt, and it hands back the path to the extended video.

This is the node that does the thing the pack is named for. The seed - 16 frames from AnimateDiff, Modelscope T2V, SVD, or even a downloaded clip, since the README says the short step is optional - gets extended chunk by chunk. Each 8-frame chunk is generated conditioned on the previous one, while the appearance-preservation module keeps reaching back to the start of the video so your scene doesn't drift into something unrecognizable.

How it works

The extension count is derived from num_frames: (num_frames - 8) / 8 autoregressive generations, each adding 8 frames. So num_frames=24 means two extensions to reach 24; num_frames=120 means fourteen. The number of num_steps you set applies to each chunk, which is why long videos take so long - it's not one big diffusion run, it's many small ones. The result is written to ComfyUI/output with a filename based on your prompt plus a timestamp, and low_video_path is the path back.

Inputs that matter

• stream_cli / stream_model - the loaded StreamingT2V pair from StreamingT2VLoaderStreamModel.
• short_video - your 16-frame seed as an IMAGE (from the short-step run node, or loaded from a file).
• prompt - the scene description; used for the whole extension.
• num_frames - total target length (default 24). Each 8-frame chunk is 8 frames, so the node effectively rounds to the nearest 8.
• num_steps - steps per chunk, default 50.
• image_guidance - guidance scale, default 9.0; the "how closely does it follow conditioning" knob.
• seed - default 33.

Output is a single STRING, low_video_path, pointing at the 256px mp4. That path is exactly what StreamingT2VRunEnhanceStep wants as its input, so the two chain naturally.

Installing it

Pack install:

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

or ComfyUI Manager → search StreamingT2V → install, restart. The catch: this is the original-model path, so you need the ~25.6GB streaming_t2v.ckpt in models/checkpoints yourself, and the README says to budget <15GB VRAM. If that's not you, the VidXTend run nodes do the same job at a fraction of the weight.

The take

Honest framing: this node is the authentic StreamingT2V experience, warts and all - long runtimes, 256px output, and a 25.6GB checkpoint. Its practical sibling is the VidXTend path. But if you've already got the original checkpoint on disk, this is the cleanest way to use it. Two practical tips from the README: run with Auto Queue on and stop when you like the result, and expect each chunk to cost real time - a 120-frame video is fourteen 50-step generations, not one.