CogVideoNode

If you've looked at CogVideoX in ComfyUI before, you've probably seen Kijai's wrapper with its ten inputs and its GGUF support. This node is the opposite of that: it's a single box that runs the whole CogVideoX-5B pipeline - text to video, or video to video if you feed it a clip - then optionally upscales it and optionally runs RIFE frame interpolation, all in one go. Fewer knobs, less control, and a lot less ceremony. For a first CogVideoX clip, that trade is often worth it.

CogVideoX-5B is Zhipu's open video model, one of the first genuinely viable open local video generators. It predates Wan and Hunyuan, and by now it's been thoroughly superseded - if you want the best open models, you're looking at Wan 2.2, not this. But it's still a coherent model with real motion understanding, and this node makes it nearly turnkey. Consider it the "I just want to see what local video looks like" button.

How it works

The pack vendors the entire DiffSynth library in its own folder, so the node doesn't call out to anything external. On run it builds a ModelManager, loads the CogVideoX text encoder, transformer and VAE from models/CogVideo/CogVideoX-5b/ plus the RIFE flownet.pkl, then runs CogVideoPipeline. The details matter here, because they're hardcoded:

• 49 frames at 480×720 (portrait), 200 denoising steps, saved at 8 fps to ComfyUI/output/ as diffsynth_cogvideo_*.mp4.
• Feed it a video and it becomes a video-to-video pass at denoising strength 0.7 - regenerate a clip rather than generate from scratch.
• Toggle if_upscale on with a video attached and it re-renders at 960×1440 with tiled attention and a gentler denoising strength.
• Toggle if_interpolate and it runs RIFE with num_iter=2, quadrupling the frame count to 32 fps.

So the workflow is: text (or a clip) in, an ~6-second MP4 out, with upscaling and interpolation available as bolt-ons.

Inputs and outputs

Of the four required inputs, only two are things you'll touch every time:

• prompt (TEXT) - your prompt, usually wired from the pack's own TextEncode node.
• seed (INT, default 42) - set it, because you will want to rerun a near-miss.
• if_upscale and if_interpolate are booleans, both defaulting to false.

The optional inputs are the interesting ones: video (VIDEO, from the pack's LoadVideo node) switches it into video-to-video mode, and models (FILES) is the output of DownloadModelsNode - wire them together so the models actually get pulled before you run. The output is a single VIDEO that feeds the pack's PreViewVideo node (which gives you an in-graph preview) or anything else that takes a video path.

The trap

Here's the thing nobody tells you: if_upscale is a no-op unless you've attached an input video. In the source, the upscale branch lives inside the if video: block. Toggle it with no video connected and nothing changes - you get the same 480×720 output. And the upscaled dimensions are a hardcoded 2× of the generation size, not a choice you make. Fine if you know it, baffling if you don't.

Installation

Install the pack the normal way - ComfyUI Manager, search DiffSynth-ComfyUI, or:

cd ComfyUI/custom_nodes
git clone https://github.com/AIFSH/DiffSynth-ComfyUI

then restart ComfyUI. The requirements.txt is the part to read before you commit: torch>=2.0.0, cupy-cuda12x, transformers, modelscope, imageio[ffmpeg], sentencepiece, protobuf, among others. That cupy-cuda12x line is the whole story of this pack's hardware ceiling.

Common issues

NVIDIA only. cupy-cuda12x is CUDA-only, and DiffSynth's stack errors out on AMD cards with "Found no NVIDIA driver on your system." If you're on AMD or Apple Silicon, this pack is not the path for you - walk away now.

VRAM. CogVideoX-5B in bf16 is heavy. The community's low-VRAM route (GGUF quantized weights, which get a 5B clip running in 10 GB) isn't available here; the node loads full-precision files. Figure on 16–20 GB of VRAM to be comfortable, and accept that 200 inference steps at 480×720 will take a while even then.

First-run downloads. The node will error on missing weights until you run DownloadModelsNode once - that's a multi-gigabyte pull. Do it in a separate queue slot first.

The README is one line long and won't help you with any of this, which is exactly why the source is the better reference. For a quick, all-in-one CogVideoX demo this node is great. For serious video work in 2026, you've outgrown it before you've finished the first clip.