Cloud Empty Latent

Every text-to-image workflow starts the same way: a blank latent the sampler turns into noise, then into an image. Cloud Empty Latent is that starting point for a cloud chain. You give it width, height, and batch size, and it returns a CLOUD_LATENT handle backed by an EmptyLatentImage node that exists only in the workflow JSON the cloud will eventually run. Your machine never holds a latent tensor - not a blank one, not a sampled one.

This is the node you reach for to begin a graph-style cloud chain. Pick a checkpoint, encode your prompts, drop a Cloud Empty Latent for the latent slot, sample with Cloud KSampler Graph or Cloud KSampler Advanced, decode with Cloud VAE Decode, and fetch the result. That's the whole canonical pipeline, and this node is its step one.

The inputs that matter

• width / height - default 1024×1024, range 64–8192, stepping in multiples of 8. The step isn't decoration; pass a non-multiple and you're begging for a dim-mismatch error when the cloud VAE runs.
• batch_size - default 1, up to 16. Bump it and the cloud generates N images per run. Remember you're paying per job, so a batch of 4 is four images of cloud GPU time.

Output: latent (CLOUD_LATENT) - into the latent_image slot of Cloud KSampler Graph or Cloud KSampler Advanced. It's a handle, not a tensor, so it won't connect to local sampler nodes.

How it works

Same engine as every node in this pack: it appends an EmptyLatentImage spec with your values to an accumulating workflow dict and returns a handle referencing it. Nothing is created at build time. When you eventually hit a terminal node - Cloud Fetch Images, Cloud VAE Decode - the whole assembled graph is POSTed to /api/prompt on Comfy Cloud, the blank latent is created there, and the sampler fills it in.

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 install via ComfyUI Manager. Dependencies are just requests, Pillow, safetensors - there's no VAE or checkpoint to download.

Common issues

• "My latent doesn't connect to anything" - it's a CLOUD_LATENT, and it only plugs into cloud nodes. Mixing it into a local sampler graph is a type error by design.
• Odd dimensions - the 8-step constraint is real; keep width and height multiples of 8.
• Need a video latent? Wrong node. Cloud Empty Latent Video is the sibling for Wan, HunyuanVideo, and LTX chains.
• PoC expectations - the pack is a single-author proof of concept. It does this one job cleanly, which is more than some parts of the pack can claim; just keep the pack-wide "expect rough edges" caveat in mind before betting a paid run on it.