(Down)load Tora Model

Plain CogVideoX gives you a prompt and a roll of the dice on how things move. Tora is the add-on that lets you draw the motion - sketch a path, and the subject follows it. This node is how you get the Tora weights into your workflow. Point it at one of two models, hit run, and it downloads the files from Hugging Face and hands you a loaded tora_model for the encoder nodes to use.

What Tora actually is

Tora is a "trajectory-oriented diffusion transformer" out of Alibaba, and the community's plain-English translation is the one worth remembering: it lets you direct motion in a video the same way LoRAs let you direct style in an image. Instead of hoping the model animates your bear walking left-to-right, you feed it a trajectory and it makes the subject move along it. When it landed in this wrapper in late 2024 it got real attention - the "first CogVideoX-Tora generations" showcase pulled hundreds of upvotes and a lot of "okay, that's the tech getting good" reactions.

Context helps here. CogVideoX itself is the THUDM / Z.ai model that was one of the first genuinely usable open local video models, back before Wan and Hunyuan took over the space. It's since been superseded for raw quality, but its motion-control ecosystem - Tora, ControlNet, the Fun variants - is a big part of why people still fire it up. Tora is the marquee example of that: precise, directable motion you don't easily get elsewhere.

The one input, and the two models

This node has a single required input:

• model - a dropdown with two choices. kijai/CogVideoX-5b-Tora is the text-to-video Tora model: prompt plus a trajectory. kijai/CogVideoX-5b-Tora-I2V is the image-to-video variant: start from an image and a trajectory, which is the combo most people actually want, because now you control both the subject and how it moves.

Pick the -I2V one if you're driving motion from a real starting frame. Pick the plain one for pure text-to-video with a path.

Under the hood, the node downloads the weights to ComfyUI/models/CogVideo/CogVideoX-5b-Tora and loads them. First run is a real download - this is a 5B model, so budget disk space and a coffee - and every run after that loads from the cached files.

The output is tora_model (type TORAMODEL). On its own it does nothing visible. It's the key that unlocks the two Tora encoder nodes: ToraEncodeTrajectory (draw a path from coordinates) and ToraEncodeOpticalFlow (drive motion from an existing optical-flow field). Both of those take this tora_model as an input, encode your motion into trajectory features, and those features feed the CogVideo Sampler's tora_trajectory slot.

How to install it

The node ships with the CogVideoX wrapper - there's nothing node-specific beyond the pack:

• ComfyUI Manager - search ComfyUI-CogVideoXWrapper, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-CogVideoXWrapper, then pip install -r ComfyUI-CogVideoXWrapper/requirements.txt, and restart.

Common issues & troubleshooting

The download stalls or the first run takes forever. That's the model transfer, not a hang. It's fetching a multi-gigabyte 5B checkpoint on first use. Let it finish; subsequent runs load from the local cache and are fast.

Out of memory when the Tora workflow runs. CogVideoX-5b is already VRAM-hungry, and 12GB cards run it on the edge even with plain generation - people routinely drop to quantized (GGUF) variants of the base model to fit. Tora adds the trajectory machinery on top. Use fp8 and CPU offloading at the model loader, keep frame count and resolution modest, and don't stack a dozen accelerators at once.

You loaded the wrong variant. If image-to-video ignores your start image, you probably grabbed CogVideoX-5b-Tora instead of the -I2V model. Motion path and start image are two different jobs; the -I2V weights do both.