Practical tutorial
A cogvideox comfyui workflow connects a compatible model, its required nodes, a prompt, and a video output path. The exact nodes depend on the extension and model variant you install, so verify those details before following any shared graph.
Check the model and node requirements first. A graph can look complete while still depending on files or node versions that are missing from your installation.
Opening a shared ComfyUI graph cannot supply missing custom nodes, model weights, or the files a particular node pack expects. Note the names of any missing nodes and identify the extension that provides them before changing graph connections.
What to do instead
Install the required node pack from its maintained source, follow its model-placement instructions, restart ComfyUI, and reopen the graph.
CogVideoX text-to-video and image-to-video paths can require different inputs. A graph that expects a source image will not become a text-only graph just because its prompt changes; model size and packaging may also affect compatible loaders.
What to do instead
Match the checkpoint, loader, and input nodes to the variant documented by the extension you are using.
Video inference can exhaust available memory during loading, sampling, or decoding. This page cannot establish a universal GPU minimum because memory use varies with the model, node implementation, precision, and output settings.
What to do instead
Start with a modest frame count and resolution, then use memory-saving options only when your installed nodes support them.
Some graphs expose intermediate frames without saving a playable clip. A successful sampler run therefore does not, by itself, confirm that the output has been assembled or written to disk.
What to do instead
Check for a compatible frame-to-video or save-video node and confirm that the expected file appears in the configured output directory.
Use this sequence as a diagnostic path rather than a promise that every extension uses identical node names.
Update ComfyUI according to its installation instructions, then install a maintained CogVideoX-compatible node extension. Read that extension's documentation for supported models and dependencies. Restart ComfyUI so its nodes register, and resolve any startup errors before importing a graph.
Download a model variant supported by your chosen extension and place its files where that extension expects them. In the graph, select the corresponding loader and verify that it finds the files. Do not assume a loader for one packaging format can open another.
For text-to-video, supply a descriptive prompt through the graph's text-conditioning path. For image-to-video, add the required source-image input and use a prompt that describes the intended motion. Keep the model, conditioning, sampling, and decoding connections consistent with the extension's example graph.
Begin with conservative settings so a wiring or memory problem is easier to isolate. Queue the graph, watch for the first failing node, and record its error rather than changing several settings at once. If frames appear, inspect them for coherence before increasing the workload.
Connect decoded frames to an output node that supports your intended video format. Open the saved file and check playback, framing, and motion; a node preview alone is not enough. Once the short test works, save the graph with its model variant and node dependencies noted.
Start with the failing node and its message. These checks separate installation, input, and resource problems before you rebuild a working graph.
The imported graph displays unknown nodes or disconnected sockets after loading.
Identify the missing extension, install a version compatible with your ComfyUI installation, restart, and import the graph again. For a broader introduction to model inputs and outputs, use the linked guide before replacing nodes by guesswork.
The graph asks for an image, or the source frame is ignored during generation.
Confirm that the loaded model and conditioning path support image input, then trace the image connection into the appropriate node. The linked variant guide helps distinguish an image-conditioned setup from a text-only path.
The loader cannot find a checkpoint even though a model file has been downloaded.
Compare the extension's expected folder and file format with the files you have; then refresh the loader list or restart. The linked model page covers the named 5B image-to-video variant, which should not be selected unless your graph supports it.
After one short run succeeds, change a single variable at a time. That makes improvements easier to attribute and failures easier to reverse.
| First successful run | Controlled refinement | |
|---|---|---|
| Prompt | Describe the main subject, setting, and one clear action. | Keep those elements fixed while testing one change to motion, camera direction, or atmosphere. |
| Input image | Use a clear source frame when the model and graph require one. | Retest with the same image to judge prompt or setting changes; switch images only as a separate experiment. |
| Frame count | Run a short sequence to confirm that sampling and decoding work. | Increase the sequence only after checking memory use and inspecting the shorter result. |
| Resolution | Choose a modest size supported by your graph. | Raise it gradually if the installed model and available memory permit. |
| Sampling settings | Begin with the extension's documented example settings. | Change one supported setting per test and keep a record of the value and result. |
| Reproducibility | Save the working graph and note its model variant. | Record the extension version, seed when available, input asset, and output settings alongside each comparison. |
| Output check | Confirm that a file was saved and plays correctly. | Inspect the full clip for flicker, unwanted changes to the subject, and abrupt motion before treating a preview frame as representative. |
These questions address the setup and troubleshooting implied by searches for a CogVideoX workflow in ComfyUI.
You need a working ComfyUI installation, a compatible node extension, and the model files required by that extension. If you are importing someone else's graph, check its listed dependencies and model variant first; the graph file alone does not include them.
The graph likely uses a custom-node pack that your installation has not loaded, or it was made with a different version of that pack. Check the missing node names, install the appropriate maintained extension, restart ComfyUI, and reopen the graph.
Not necessarily. An image-to-video graph needs a compatible model and a connected image-conditioning path, while a text-to-video graph may not have either. Follow the documentation for your installed model and extension rather than assuming that changing the prompt is sufficient.
Reduce the workload, starting with frame count and resolution, and retry a short test. If the installed extension documents supported memory-saving options, test those separately; requirements vary by model, precision, and hardware.
The graph may be previewing decoded frames without passing them to a video-writing node. Check the output connections, the save node's supported format, and the configured output directory. Then play the saved file to confirm it contains the intended sequence.