StreamingT2VLoaderModelscopeModel

StreamingT2VLoaderModelscopeModel loads the Modelscope text-to-video model on its own, so you can generate the short seed video that the rest of the StreamingT2V pipeline extends. It's the T2V seed generator in the pack's trio - the other two being AnimateDiff (also text-to-video) and SVD (image-to-video). The T2VModel output feeds the pack's short-step Modelscope run node, which turns a prompt into the 16-frame 256×256 clip.

The seed matters more than it looks. StreamingT2V's whole design is autoregressive: it takes a short 16-frame video, appends 8 more frames conditioned on the last chunk, and repeats. The seed sets the scene, the subject, the colors - everything the appearance-preservation module will try to hold onto for the whole long generation. In the README's own words, the short step is "not necessary" - you can feed it any 16-frame 256×256 video, even one downloaded from the internet. But when you do use it, the seed generator is where your prompt actually becomes content.

What it loads

damo-vilab/text-to-video-ms-1.7b, the 1.7B-parameter Modelscope T2V model, in fp16, from HuggingFace, downloaded automatically on first use. It swaps in a DPM solver scheduler and enables CPU offload and VAE slicing. First execution downloads roughly 4GB of weights, so don't panic when the Queue button sits there for a while - it's fetching the model, not frozen.

Inputs and output

One input: device (cuda or cpu, default cuda). That's the whole input list. The single output is a T2VModel object that goes to the short-step run node.

Installing it

Same as the rest of the pack:

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 (diffusers, pytorch-lightning, modelscope, xformers, open-clip-torch...). This loader specifically pulls in diffusers' DiffusionPipeline, so a clean diffusers install matters here.

The take

Of the three seed generators, this is the one to reach for when your video is prompt-driven and you want a little more fidelity than AnimateDiff's SD 1.5 baseline - the Modelscope T2V model was a serious text-to-video model in its day. It's also the default in the pack's wf.json. The honest caveat: it's a 2023-era 1.7B model, so don't expect it to match what modern video models do; its job is to hand the long-extension stage a stable, coherent 16 frames, and at that it's fine. You're using it as a seed factory, not a final renderer.