Cloud Empty Latent Video

Image latents are easy - width, height, batch, done. Video latents carry an extra dimension: time. Cloud Empty Latent Video is the blank-starting-latent node for video chains, and the one thing it adds over its image sibling is a format selector that picks the right kind of empty video latent for your model family.

That selector is the part worth understanding. Video latent blanks aren't interchangeable: Wan and HunyuanVideo want one shape of latent (with the specific frame-count and compression quirks of EmptyHunyuanLatentVideo), and LTX 2.0 wants its own (EmptyLTXVLatentVideo). The node maps format to the correct ComfyUI class cloud-side, so you get the right node for the model without hunting for it. It's also the node you'll see in the pack's video example workflows - the Wan 2.2 14B I2V and LTX 2.0 templates both start here.

The inputs that matter

• format - hunyuan_wan (for Wan 2.x and HunyuanVideo chains) or ltxv (for LTX 2.0). This is the one that actually matters; pick wrong and the cloud's sampler will likely reject the latent shape.
• width / height - default 832×480, a sensible 16:9-ish video canvas, range 64–8192 in multiples of 8.
• length - the number of frames, default 33. This is your duration knob: at a given fps, more frames = longer video = more cloud compute. The Wan example workflow's numbers are a good starting reference.
• batch_size - default 1, up to 16. Multiple videos per run, if you're feeling spendy.

Output: latent (CLOUD_LATENT) - feeds the latent_image slot of a cloud video sampler, or a Wan/LTX img-to-video chain.

How it works

Like every cloud node, it appends one spec to the accumulating workflow JSON - in this case EmptyHunyuanLatentVideo or EmptyLTXVLatentVideo, chosen by format - and returns a handle. The blank latent is created on Comfy Cloud when the chain actually runs. None of the video frames exist on your machine until a terminal (Cloud VAE Decode → Cloud Fetch Video / Cloud Fetch Images) brings results home.

How to install

cd ComfyUI/custom_nodes
git clone https://github.com/Dobidop/ComfyUI-CloudAPI-worker

Copy config.json.example to config.json, paste an API key from https://platform.comfy.org/profile/api-keys, restart. Or use ComfyUI Manager. Just requests, Pillow, safetensors - no video models download to your disk.

Common issues

• Wrong format - the classic. A Wan chain fed an LTX latent is a guaranteed shape error or garbage output. Match it to the model.
• Videos come out too short/long - length is frames, not seconds. Check the model's native fps (LTX runs ~24–25 fps) and do the math.
• Batch size on video - 16 videos at 1080p is how you spend a week of cloud budget in one click. Be kind to your wallet.

And the standing pack note: this is a proof of concept from a single author ("very hacked together" is their own words). Start with the example workflows, which are the tested paths, before inventing your own video graph.