Cloud Wan Image To Video

This is the node the pack was basically built around - the README's example workflow is literally video_wan2_2_14B_i2v_cloud.json, a Wan 2.2 14B image-to-video job. The pitch is the whole reason the pack exists: you get to run the 14B I2V model that would otherwise demand a 24GB card and twenty-five minutes of grinding, and instead someone else's GPU does it while your local ComfyUI just waits and polls. Wan 2.2 is the last open-weight Wan - Alibaba went API-only after it - so this is also the best open I2V model you can still run this way.

How it works

Cloud Wan Image To Video mirrors ComfyUI's WanImageToVideo conditioning node. It takes your positive and negative conditioning plus a VAE, encodes the start image into the first frames of the latent, and biases the conditioning to continue from it. Everything is appended to the assembled workflow JSON; the actual I2V encoding and sampling happen on the cloud instance.

The inputs and outputs that matter

• positive / negative (CLOUD_CONDITIONING) - from Cloud CLIP Text Encode (or a dual-CLIP setup).
• vae (CLOUD_VAE) - Wan uses its own VAE, so grab one from Cloud VAE Loader.
• width / height (default 640×640, step 8, up to 8192) - keep them multiples of 8.
• length (default 81 frames, up to 4096) - 81 frames is the classic ~5-second clip at 16fps, the sweet spot Wan was trained around. Going well past it invites drift.
• batch_size (default 1).
• Optional start_image (IMAGE) - your first frame. It's uploaded automatically when you provide it, so you don't need a separate Cloud Upload Image for this input.
• Optional clip_vision_output (CLOUD_CLIP_VISION_OUTPUT) - accepted for completeness but unused by default Wan 2.2 I2V; the official template leaves the link null, so you can safely ignore it.

It returns three handles: positive and negative (the conditioned versions) and latent (the I2V-initialized latent). Those feed a sampler (Cloud KSampler or Cloud Sampler Custom Advanced), then Cloud VAE Decode → Cloud Fetch Images to get frames back.

Time and money reality check

The README is specific: with the 4-step LoRA, Wan 2.2 14B I2V finishes in roughly 70–100 seconds; without it, plan for 8–10 minutes. Set generous timeouts on your terminal node (the default is already 1800s). And remember this is metered cloud compute, not free electricity - per-call pricing means the speed LoRA path is the difference between "a minute per clip" and "sit through a lunch break," which is exactly why those LoRAs exist in the Wan ecosystem.

Prompting, briefly

Wan I2V prompting is a known trap: the positive prompt should describe the motion, not re-describe the start image. "An athletic man enters from the left, movement smooth as he puts a hand on her shoulder" beats a caption of what's already in the frame. The community Wan guides make this the single most common beginner mistake.

Installing it

Install via ComfyUI Manager (search "ComfyUI-CloudAPI-worker") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Dobidop/ComfyUI-CloudAPI-worker
cd ComfyUI-CloudAPI-worker
cp config.json.example config.json

Paste your API key from https://platform.comfy.org/profile/api-keys into config.json and restart ComfyUI. Locally you only need requests, Pillow and safetensors - the 14B model never touches your disk.

Common issues & troubleshooting

• Forgot the terminal - this node returns handles, not video. Follow it with a sampler → decode → fetch/save chain or nothing runs.
• Long executing isn't a hang - on a 14B model it's normal. Watch the console status prints and compare elapsed time against your timeout.
• Got a pile of frames, not a file - the IMAGE batch from the terminal is N frames. Wire it into VHS_VideoCombine to mux locally; the pack won't hand you a video container by itself.