Most AI video models break down past the 5-second mark. The face shifts. Camera movement becomes jerky. The style drifts. Stable Video Infinity (SVI) solves this by generating sequential frames that reference the previous clip's ending frame. The result is longer videos where everything stays coherent.
This guide covers the full setup on RunPod with ComfyUI.
What Stable Video Infinity does
Standard video diffusion generates each clip from scratch. SVI uses the last frame of one clip as the first frame of the next. You define sequential actions — "person raises hand," "person lowers hand," "camera pans right" — and SVI renders them as one continuous video with consistent identity and motion.
The practical result: 30-second+ videos where the subject looks the same throughout, the camera moves intentionally, and the style doesn't degrade.
Step 1: Create a ComfyUI pod on RunPod
Log into RunPod, go to Pod Templates, and search for the ComfyUI template. Select it.
For the GPU, RTX 4090 works for standard resolution. RTX A6000 or better for higher resolution. SVI is VRAM-hungry — more is better.
Container disk: increase the default from 50 GB to at least 100-150 GB. SVI models and custom nodes add up quickly.
Click Deploy on Demand. Wait for the pod to reach "Running" status. Open ComfyUI from the connect button.
Step 2: Set up the environment via JupyterLab
Don't configure ComfyUI directly from the browser yet. Open JupyterLab from the RunPod pod panel — this gives you a terminal with direct access to the pod's filesystem.
Open a terminal in JupyterLab. Navigate to the ComfyUI directory:
cd /workspace/ComfyUI
Run the environment update commands. The exact commands are in the video description. They update Python packages and ensure ComfyUI is on the right version.
After running the update, restart ComfyUI via the terminal so the changes take effect.
Step 3: Install custom nodes
In the ComfyUI browser interface, open the Custom Nodes Manager.
Install these two nodes:
- KJ Nodes — may already be installed in newer ComfyUI versions. Check first. If it shows as installed, skip.
- Video Helper Suite — needed for the SVI output nodes. This one is important. Install it and restart ComfyUI when prompted.
Note: Video Helper Suite uses a different install method than standard nodes. If the standard install fails, check the description for the alternative install command.
Step 4: Import the SVI workflow
Download the SVI workflow file from the video description. In ComfyUI, open the workflow import panel and load the file.
The workflow loads but shows missing files — the models haven't been downloaded yet. This is expected.
Step 5: Download the required models
The SVI workflow needs several model files that aren't in the base ComfyUI image. The video includes a helper workflow specifically for downloading them.
Download the helper workflow from the description. Import it into ComfyUI.
Run the helper workflow. It uses HuggingFace to download:
- The SVI checkpoint
- LoRA files
- VAE model
- Any other required assets
This download takes time depending on your pod's network connection. Wait for it to complete.
After downloading, go back to the SVI workflow. The missing file errors should resolve. Refresh the node connections if needed.
Step 6: Configure the workflow
The SVI workflow front end shows several key inputs:
- Starting image: upload the first frame. This sets the identity of your subject for the entire video.
- Frame dimensions: 640x640 is a safe starting point. Larger dimensions need more VRAM.
- Action sequence: define what happens in each clip segment. Example:
- Segment 1: "person slowly raises hands above head"
- Segment 2: "person lowers hands to sides"
- Segment 3: "person steps forward"
Each action segment becomes one clip. SVI renders them sequentially, using the last frame of each clip as input to the next.
Step 7: Run the workflow
Click Queue Prompt. Watch the generation progress in the terminal logs.
When complete, the output video appears in the SVI workflow output node. Download it directly. If it doesn't show up, go to comfyui/output/video/ in the JupyterLab file tree.
Step 8: Extend to longer videos
To add more segments, duplicate the action nodes in the workflow. Each duplicated set of nodes represents one more clip in the sequence. Update the prompt for each segment, connect the last frame output of clip N to the input of clip N+1.
There's no hard limit on the number of segments. In practice, VRAM and generation time are the constraints. More segments means more render passes.
Important: terminate the pod when done
RunPod charges per uptime. A pod left running uses credits even when idle. Go back to RunPod, find your pod, click More Actions, then Terminate Pod.
If you're building video generation into a production content workflow, get in touch. We build custom AI UGC pipelines for content and product teams.