StreamingT2VLoaderVidXTendModel

StreamingT2VLoaderVidXTendModel is the loader the README actually steers you toward. It loads VidXTend - community member Painebenjamin's compressed version of the StreamingT2V model - which cuts the weight from the original 25.6GB down to about 5GB and drops the VRAM requirement from <15GB to <10GB. The output, a VidXTendPipeline, feeds the four StreamingT2VRunLongStepVidXTendPipeline* run nodes that do the long-video extension.

Why this matters: the original StreamingT2V checkpoint is a genuinely obnoxious download. 25.6GB for a 2024 research model that produces 256px output is a lot to ask, and the pack's own README leads with the VidXTend alternative before it even mentions the original. The author is upfront that this is the recommended route, and for once the recommendation is the one you'd want anyway.

How it works

Unlike every other loader in this pack, there's no ckpt_name input - nothing to download manually. The loader calls VidXTend's pipeline with from_single_file("benjamin-paine/vidxtend"), which pulls the model from HuggingFace automatically on first run. It loads in fp16, turns on CPU offload and VAE slicing, and switches to memory-efficient xformers attention. The pack even vendors its own copy of the vidxtend library under thirdparty/, so the import works out of the box without you installing anything extra.

Inputs and output

One input: device (cuda or cpu, default cuda).

One output: VidXTendPipeline, which wires into the run nodes for the long extension. Since VidXTend produces an IMAGE (not a file path), the output chains into ComfyUI-native video handling - VHS_VideoCombine and the pack's VHS_FILENAMES_STRING_StreamingT2V helper to save the result.

Installing it

The standard install:

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

or ComfyUI Manager → search StreamingT2V → install, restart. The pack's requirements.txt is heavy (pytorch-lightning, diffusers, modelscope, xformers, bitsandbytes...), but the important part is that the model itself downloads itself - first Queue click stalls for a bit while ~5GB comes down, then it's done. No manual checkpoints step, unlike the original-model loaders.

The take

If you're on a 10GB card - which, let's be honest, is most people running this kind of pack - this is the loader to build your workflow around. It's smaller, it's automatic, and it's what the pack's wf_VidXTend.json uses. The trade is that it's still a 2024-era model: 256px output before the enhance stage, motion quality that modern Wan-based tooling has left behind. But as a way to actually run StreamingT2V without a 25GB download, this is the one. Painebenjamin did the community a real favor with that compression.