WaveSpeedAI Predictor⚡

It's an API node, not a model node

The WaveSpeedAI Predictor is the kind of node that makes you stop and stare. You add it, look for the model download, and there isn't one. That's because nothing runs on your machine. The Predictor is a remote control for WaveSpeed AI's hosted API: you pick a model from their catalog, fill in the prompt, and generation happens on their servers. Your computer just uploads the inputs and polls for the result.

That makes it the exact answer to a complaint the community has hammered on for years: the gap between "open browser, click button, get video" and "install Python, manage CUDA, install eleven custom nodes." The Predictor is the paid cloud UX, dropped into your ComfyUI graph. No VRAM, no checkpoint files, no 24GB model. You trade credits instead of electricity.

How it works

The node is mostly frontend plus a thin API client. When you open it, the plugin fetches WaveSpeed's model catalog (600+ models across ~25 categories, cached for about five minutes) and builds the node's inputs on the fly from each model's schema. When you hit run, generate() does three things: converts any connected image/video/audio tensors to PNG/MP4/WAV and uploads them to WaveSpeed's CDN, POSTs the request to https://api.wavespeed.ai, then polls until the task finishes - up to a hard 30-minute timeout baked into the client.

The output is the catch that trips everyone up: output_url, an ANY-type holding a URL string or a list of URLs. Not a tensor. Wire that into the WaveSpeedAI Preview node to get a real IMAGE/VIDEO tensor for the rest of your graph.

The inputs that matter

The schema lists no fixed inputs - they're generated per model - so what you'll actually touch is:

• The model picker - category tabs plus fuzzy search. This is the whole point: one node, every model they host.
• prompt / negative_prompt / seed - standard, and the seed genuinely reproduces results.
• Size - either a dropdown of preset resolutions (models like Seedance) or separate width/height with ratio buttons (Flux-style models).
• Media inputs - image, video, audio accept ComfyUI tensors or URLs; array params like multiple reference images expand to image_0, image_1 and merge back before sending.

Don't worry about memorizing them. Pick a model and the node rebuilds itself to match.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/WaveSpeedAI/wavespeed-comfyui.git

Or search "WaveSpeed AI - Universal Generation" in ComfyUI Manager. The pack depends on requests, pillow, opencv-python, scipy, torchaudio, av, and pydantic - and its __init__.py helpfully tries to auto-install cv2, scipy, av, and pydantic on startup if they're missing. Restart ComfyUI after.

You also need an API key, and this is non-negotiable. Get one at wavespeed.ai (free tier: 100 credits/month, no credit card), then either Settings → WaveSpeed, drop a config.json with {"api_key": "..."} into the plugin folder, or set the WAVESPEED_API_KEY environment variable. Every generation burns credits, and a video can eat several.

Where people get burned

• "No API key configured" - you skipped the key setup. Do the Settings step or the env var.
• First load is slow - the node fetches the whole model catalog on first open; the README says it can take 2–4 minutes on a bad connection. Subsequent loads are cached and instant.
• "Task timed out" - video jobs can run the full 30 minutes. Try a shorter duration or lower resolution.
• Output is a URL - you tried to connect output_url straight into Save Image and it broke. Run it through WaveSpeedAI Preview first.

One honest caveat from the community: the official nodes have a habit of lagging behind WaveSpeed's newest models - people have griped that Seedream V4 and Wan 2.5 took a while to land, and some have jumped to community packs like chengzeyi/Comfy-WaveSpeed instead. The API also enforces content filters, so prompts that trip them error out rather than quietly passing. Treat the Predictor as "hosted generation, integrated into ComfyUI" and it's genuinely slick; expect it to be a free local workflow and you'll be disappointed.