(Down)load CogVideo GGUF Model

If the regular CogVideoX 5b model won't fit on your card, this is your way in. It loads a GGUF-quantized version of the transformer - smaller weights, less VRAM, same node graph downstream - and, like its non-GGUF sibling, it downloads the file for you from HuggingFace on first run and outputs both a model and a VAE.

GGUF is the quantization format that everyone leans on for big models. Q8 is basically indistinguishable from full precision at half the size; Q4 is the accepted compromise for genuinely tight VRAM. The choices this node ships are all Q4_0 - the smallest, most memory-frugal tier - because that's the whole reason you're here: getting a 5b video model onto a card that otherwise can't hold it.

What you can load

The model dropdown has 8 entries, all ..._GGUF_Q4_0.safetensors builds: the base CogVideoX_5b, the _I2V image-to-video version, the 1_5_I2V refresh, the _fun and _fun_1_1 variants, fun_1_1_Pose for pose control, _Interpolation for frame interpolation, and _Tora for trajectory motion. Match the one you pick to the workflow you're running - an I2V build wants an image, a Pose build wants a pose sequence, and so on.

The settings are refreshingly short:

• vae_precision (default bf16) - leave it. Note the transformer is already quantized to GGUF; this only sets the VAE's precision.
• fp8_fastmode - a speed toggle; leave it off unless you're chasing every last second.
• enable_sequential_cpu_offload - the escape hatch if even Q4 GGUF won't fit. Streams layers between CPU and GPU, slower but it runs.
• attention_mode - just sdpa or sageattn here (the GGUF path exposes fewer options than the full loader). sdpa is the safe default; sageattn is faster if you've got SageAttention installed.

Outputs are model (into CogVideo Sampler) and vae (into CogVideo Decode). Drop-in replacement for the full-precision loader - the rest of your graph doesn't change.

How to install it

The pack README is empty, so the standard kijai-wrapper install applies:

• 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, restart.

No separate model download step - this node pulls the GGUF file itself on first use.

Common issues & troubleshooting

It's slower than you expected. That's GGUF's tax. Quantized weights get dequantized on the fly, and at low quant tiers that overhead is real - a Q4 model can actually run slower per step than a higher tier would, on top of being lower quality. GGUF here is a VRAM fix, not a speed fix. If you have the memory for fp8, that path is usually smoother.

Quality looks rougher than the full model. Also expected - Q4_0 is the aggressive end of the ladder. It's the price of fitting the model at all on a small card. There's no higher-tier GGUF in this dropdown, so if Q4 isn't cutting it and you have a bit more VRAM, switch to the standard loader with fp8_e4m3fn quantization instead.

Still out of memory. Turn on enable_sequential_cpu_offload. If a Q4 GGUF plus CPU offload still won't run, your card is below what CogVideoX 5b needs, and no setting fixes that.