TI2V_API

The sibling node in this pack is the one most people should actually reach for. TI2V_API is the hosted version of Step-Video-TI2V: same image-to-video idea, but instead of you running a 30B model across four GPUs, you send your image to StepFun's cloud and get the finished clip back. No multi-GPU rig, no model download, no remote server to babysit. Just an API key, a working internet connection, and the willingness to pay per generation.

The mechanism is straightforward, and it's all right in the source. The node takes your image_input, saves the first frame to output/, base64-encodes it, and POSTs it to api_url - default https://api.stepfun.com/v1/video/generations - with the model name step-video, your prompt, and a resolution. StepFun's server returns a task_id, the node polls {api_url}/{task_id} every ten seconds until the status flips to success, downloads the resulting mp4, and reads it back into ComfyUI as an IMAGE tensor of frames. Wire that into VHS_VideoCombine and you've got your video file. The whole dance is invisible to you; you just see the image go in and the frames come out.

The inputs that matter

There are only five, and honestly only three you'll touch:

• api_key - required, and the thing everyone forgets. You get one from platform.stepfun.com, and it goes into the Authorization: Bearer header. Empty key means a silent 401 and a mystery.
• video_size - a dropdown, not a free-form box: 960x540, 544x992, or 768x768. Portrait, landscape, or square; pick the aspect ratio that fits your subject.
• text_prompt - again the default is 笑起来 ("smile"), straight from StepFun's own examples. It's a real prompt field, so type your own.

image_input is your first frame - and, same as the local node, only the first image in the batch is used. api_url you'll only change if StepFun moves the endpoint or you're proxying it.

How to install it

This is the whole reason the API node is the friendly one. The pack has no requirements.txt - it rides on torch, torchvision, requests, and tqdm, all of which ComfyUI already has:

cd ComfyUI/custom_nodes
git clone https://github.com/stepfun-ai/ComfyUI-StepVideo.git

Restart ComfyUI (or hit refresh in the Manager), and the node appears under StepVideo. That's it. No weights, no conda env, no call_remote_server.py. Get a key from the platform page, drop it in, run.

Where people get burned

• It costs money. This is a paid API. The node gives you no pricing feedback loop - you can burn through a lot of credits iterating on prompts. Budget like it's a subscription, because it behaves like one.
• The files pile up in output/. Every run writes {first_50_chars_of_prompt}_img.png and {prompt}_vid.mp4 into an output/ folder relative to where ComfyUI runs. Two hundred runs, two hundred pairs of files. Clean it out occasionally.
• Failures are loud, not graceful. If the API returns status: fail - or the poll loop times out on a bad key or a dead network - the node raises an exception mid-graph. The image it already saved is still there; the video isn't.
• The quality ceiling comes along for the ride. The hosted service is the same Step-Video family, which means the same deep-compression VAE trait as the local model: strong on motion and camera work, soft on fine detail. If you're chasing crisp faces, this isn't the model, API or no API.

Worth saying plainly: this node is why the pack exists. The local TI2V node is a flex for people with GPU servers; TI2V_API is the one that turns a 30B video model into something you can use from a laptop. Grab the key, set a spending limit, and go make a horse turn around.