StreamingT2VLoaderSVDModel

StreamingT2VLoaderSVDModel loads the Stable Video Diffusion image-to-video model by itself, producing an I2VModel that feeds the pack's short-step SVD run node. In the staged workflow, this is the seed generator for the image-to-video route: you give it a still image, it animates 16 frames at 256×256, and StreamingT2V extends those frames into a long video.

It's the I2V counterpart to the pack's two T2V seed generators (AnimateDiff and Modelscope). The key difference is right in the names: this one starts from an image. That makes it the natural choice when you have a reference - a character, a location, a frame you want the whole video anchored to. The short-step SVD run node takes an IMAGE input, so you load a picture, get a 16-frame clip, and hand it to the long-extension stage. You can even skip the seed entirely; the run node will use SDXL to dream up a seed image from your prompt.

What it loads

stabilityai/stable-video-diffusion-img2vid-xt, in fp16, downloaded automatically from HuggingFace on first use, with model CPU offload and VAE slicing enabled. No files to place in models/checkpoints - this is pure auto-download territory. The XT variant is the higher-fidelity SVD release, which is a good match for a seed that the rest of the pipeline has to preserve.

Inputs and output

One input: device (cuda or cpu, default cuda). That's the entire input list - there's no checkpoint picker because there's nothing local to pick.

The single output is an I2VModel, which goes to the short-step SVD run node. Note it's typed separately from the T2V model outputs, so you can't accidentally wire the wrong seed generator into the wrong run node - the pack's custom types keep you honest.

Installing it

The usual pack install:

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

or ComfyUI Manager → search StreamingT2V → install, restart. The requirements.txt is the pack's usual heavyweight list (diffusers, pytorch-lightning, modelscope, xformers...), and the first Queue click downloads the SVD weights (roughly 5GB), so it'll sit there a minute. Normal.

The take

This is the loader to use when you're building the staged I2V workflow and want VRAM discipline - SVD plus the extension model, without dragging SDXL and the enhancer along for the ride like the all-in-one StreamingT2VLoaderSVD does. The honest caveat about SVD as a seed generator: it animates a still nicely, but it's a 2023-era model, and its 16 frames are a starting point, not a finished scene. The pack is designed around exactly that - the seed only needs to be good enough for the appearance-preservation module to hold onto.