CogVideoX-Fun Sampler for Text to Video

The text-to-video sampler for the LoadCogVideoXFunModel side of this pack. If you've used any of the Wan samplers here first, the one thing that'll stand out is the scheduler dropdown: Wan's samplers only offer Flow, because Wan is a flow-matching model with one native schedule. CogVideoX-Fun predates flow matching in this pack and gives you the classic diffusion scheduler menu instead - Euler, Euler A, DPM++, PNDM, DDIM - the same family you'd recognize from Stable Diffusion image generation. That's a real architectural difference, not a UI inconsistency.

How it works

Standard T2V diffusion sampling: noise in, denoise over steps iterations following the chosen scheduler's step function, guided by CFG and your prompt pair, output a frame batch at the resolution and length you set.

The inputs that matter

• cogvideoxfun_model - from LoadCogVideoXFunModel.
• prompt / negative_prompt - STRING_PROMPT type; wire from FunTextBox.
• video_length - default 49, range 5-85, step 4. Noticeably shorter ceiling than the Wan samplers' 161 - CogVideoX's native training window is smaller.
• width - default 1008, range 64-2048, step 16.
• height - default 576, range 64-2048, step 16.
• is_image - default false; set true to generate a single still frame instead of a clip.
• steps - default 50, up to 200.
• cfg - default 6, range 1-20.
• scheduler - one of Euler, Euler A, DPM++, PNDM, DDIM, default DDIM. DDIM is the safe, well-understood default; if you're chasing speed at fewer steps, DPM++ is generally the better tradeoff in the wider diffusion-scheduler world, though CogVideoX-Fun's own README doesn't make a specific recommendation.

Output: images - a frame batch.

Installing it

Part of the base VideoX-Fun pack. Search ComfyUI Manager for VideoX-Fun; if it's not indexed there, install manually:

cd ComfyUI/custom_nodes
git clone https://github.com/aigc-apps/VideoX-Fun
pip install -r VideoX-Fun/requirements.txt

Restart ComfyUI.

Common issues

video_length rejected past 85. That's the hard ceiling for this model family, unlike Wan's 161 - CogVideoX-Fun simply wasn't trained on longer clips. There's no riflex_k option on this node either; RIFLEx-style frame extension is only wired into the Wan samplers in this pack.

Prompt won't connect. STRING_PROMPT custom type - needs FunTextBox, not ComfyUI's default text node.

Should you be here instead of a Wan sampler? Realistically, only if you already have a CogVideoX-Fun-specific workflow or specifically want the lighter 2b checkpoint's speed. The wider community consolidated hard around Wan for a reason - better motion, stronger prompt adherence, a much bigger LoRA and tooling ecosystem. CogVideoX-Fun still works and this node still does what it says, it's just not where the active energy is anymore.

Slow at high step counts. 50-200 steps on a 5b checkpoint at 2048px is genuinely heavy for what's now an older architecture. Drop steps for iteration - DDIM at 30-50 is a reasonable starting range before you go hunting for a faster scheduler.