StreamingT2VLoaderModelscopeT2V

StreamingT2VLoaderModelscopeT2V is the pack's flagship "everything at once" loader: it stands up the full three-stage pipeline - Modelscope T2V for the short seed, the StreamingT2V autoregressive model for the long extension, and the damo/Video-to-Video enhancer for the upscale - and hands you one StreamingT2VModel to feed the run node. If you load wf.json from the README, this is the node at the top, and its name tells you the pipeline's spine: Modelscope text-to-video, with StreamingT2V on top.

Why does this node exist separately from the AnimateDiff and SVD versions of the same idea? Because the pack gives you a choice of seed generator, and each choice is its own loader. The Modelscope variant is the T2V default - prompt in, video out, no image required. It's the one the README's "generate all by one workflow" flow starts with.

What it loads, and what that costs

Three models behind one interface:

• StreamingT2V long model from streaming_t2v.ckpt in ComfyUI/models/checkpoints - the original 25.6GB checkpoint, which you must download yourself from the PAIR/StreamingT2V HuggingFace page.
• Modelscope T2V (damo-vilab/text-to-video-ms-1.7b), auto-downloaded in fp16.
• damo/Video-to-Video enhancer from modelscope, auto-downloaded.

All three resident at once. That's the "MORE VRAM" label the README puts on this class of workflow - not a suggestion, a warning.

Inputs that matter

• ckpt_name - the StreamingT2V checkpoint from your checkpoints folder, default streaming_t2v.ckpt. If it's missing, this loader fails before it even gets to the models it could download.
• device - cuda or cpu, default cuda.
• vram_not_enough - default True. Keeps the big StreamingT2V weights on CPU rather than moving them to the GPU. Leave it on unless you're confident about VRAM.

The single output, StreamingT2VModel, wires straight into the pack's all-in-one run node, which drives the seed generation, the autoregressive extension, and the enhance pass in sequence.

Installing it

Standard pack install:

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

or ComfyUI Manager → search StreamingT2V → install, restart. Then the two hard parts: requirements.txt is a long list (pytorch-lightning, diffusers, modelscope, xformers, bitsandbytes, open-clip-torch...), and you must place the ~25.6GB streaming_t2v.ckpt in models/checkpoints before the first run. First queue will also download the Modelscope T2V weights and the enhancer. Budget disk space accordingly - this path wants well over 30GB of models.

The take

This is the most convenient loader in the pack and the one most likely to OOM a mid-range card. The README is refreshingly honest about the split: the all-in-one workflows exist for people with the VRAM, and the staged workflows (which the VidXTend path uses) exist for everyone else. If you have a 16GB+ card and a big drive, this node is the cleanest way to get from a prompt to a finished long video in one graph. If you don't, use StreamingT2VLoaderVidXTendModel instead - same end result, a fraction of the model weight.