StreamingT2VRunShortStepSVD

If you want your long video to start with a specific image - a character you generated, a composition you framed - this is the seed node for you. It's the image-to-video short step in the StreamingT2V pipeline: feed it a still, and it animates the first 16 frames that the long-step nodes then keep extending.

What it is

The StreamingT2V workflow needs a short 16-frame 256x256 clip as its starting point. The pack gives you three ways to make one, and this is the SVD route: it runs Stable Video Diffusion XT (stabilityai/stable-video-diffusion-img2vid-xt) on your image and returns those 16 frames as an IMAGE tensor named short_video. Modelscope and AnimateDiff are the text-only options; this one is what you reach for when you already have the opening frame and just need it to move.

How it works

SVD-XT is a big, picky model. It was trained to take a 1024x1024 canvas, and it wants the subject framed in the middle with padding around it. So behind the scenes this node pads your image (it adds a 224px margin on each side), runs 25 diffusion steps, then crops the padding back off and resizes to 256x256 - keeping only the first 16 frames. All of that is handled for you; you just hand over a normal image.

Inputs that matter:

• model - an I2VModel from StreamingT2VLoaderSVDModel. That loader pulls SVD-XT down from HuggingFace on first run (a few GB), automatically.
• image - the still you're animating. Required.
• prompt - here's the honest bit: SVD-XT isn't text-conditioned, so once you've supplied an image this prompt does almost nothing. The code only consults it to generate a fallback image via SDXL when no image is given, and this node always requires one. Fill it in, but don't expect it to steer the motion.
• seed - controls the animation. Same seed, same motion; change it and the cat moves differently.

Wire short_video into a long-step node (the VidXTend one is the friendly option - see below) or into StreamingT2VRunT2V if you're running the all-in-one workflow.

Installing the pack

Same pack as the other StreamingT2V nodes - chaojie/ComfyUI_StreamingT2V:

• ComfyUI Manager → search StreamingT2V → install, restart.
• Or cd ComfyUI/custom_nodes && git clone https://github.com/chaojie/ComfyUI_StreamingT2V, restart.

The requirements list is a long one (diffusers, modelscope, xformers, decord, av, open-clip-torch, fairscale…), so budget time for the install. The long-step half additionally needs the StreamingT2V checkpoint - the 25.6GB original streaming_t2v.ckpt in ComfyUI/models/checkpoints, or the 5GB VidXTend model (auto-downloaded, and the one to use if you're under ~10GB VRAM). First run of this node downloads SVD-XT automatically.

Gotchas

• Square-ish inputs behave best. SVD-XT wants that 1024x1024-conditioned crop, and extremely wide or tall images get heavily cropped.
• SVD-XT is the heaviest seed option - noticeably slower to load and run than the Modelscope short step. That's the price of starting from a real image.
• Because the short step and the streaming model are meant to load one after the other, don't try to hold SVD-XT and the 25GB original in VRAM simultaneously. Use the two-stage workflow, or swap in VidXTend if your card is modest.
• It's 16 frames out, always. That's the seed, not a finished clip - extend it.