StreamingT2VLoaderSVD

StreamingT2VLoaderSVD is the image-to-video all-in-one loader: it stands up the StreamingT2V long model, Stable Video Diffusion as the seed generator, SDXL for turning your prompt into the seed image, and the damo/Video-to-Video enhancer - all at once - and hands you a StreamingT2VModelSVD for the I2V run node. Load wf_svd.json from the README and this is the node at the top.

This is the choice when your long video starts from an image, not just a prompt. SVD (Stable Video Diffusion, from the Stable Diffusion lineage) is an image-to-video model: you feed it a still, it animates it into a short clip. In this pack's pipeline, SVD produces the 16-frame 256×256 seed, and StreamingT2V takes it from there into a long video. If you have a character reference, a storyboard frame, or any still you want to be the visual anchor, this is the loader that gets you to it.

What it loads

• StreamingT2V long model from streaming_t2v.ckpt in models/checkpoints - the 25.6GB original checkpoint you download yourself.
• SVD (stabilityai/stable-video-diffusion-img2vid-xt), fp16, auto-downloaded, CPU offloaded.
• SDXL base (stable-diffusion-xl-base-1.0), used to generate the seed image if you don't supply one - yes, this path carries SDXL too.
• damo/Video-to-Video enhancer from modelscope, auto-downloaded.

That's four models resident. This is the most VRAM-hungry of the pack's loaders, and the source confirms it: when vram_not_enough is true, it goes so far as to push the base model to CPU.

Inputs that matter

• ckpt_name - the StreamingT2V checkpoint, default streaming_t2v.ckpt.
• device - cuda or cpu, default cuda.
• vram_not_enough - default True; keeps the big StreamingT2V weights on CPU.

The output StreamingT2VModelSVD feeds StreamingT2VRunI2V, which takes your image input and runs the whole chain.

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 familiar two hurdles: the heavy requirements.txt (pytorch-lightning, diffusers, modelscope, xformers, bitsandbytes...), and the manual 25.6GB streaming_t2v.ckpt download into models/checkpoints. First run also pulls SVD, SDXL, and the enhancer from their hubs.

The take

Real talk: this is the path with the most moving parts and the most VRAM. The README's own recommendation for I2V work isn't to run all four models at once - it's the staged route, and if you're on a sub-15GB card the VidXTend loader is the sane option regardless. But as a one-graph image-to-long-video setup, this loader is the complete package. Just be prepared for a genuinely long first load and a serious chunk of disk. And worth knowing: in the run node, SVD can generate the seed image from your prompt itself, but with this loader the cleaner workflow is to feed an actual image in.