Lucy Edit Pro - API

Some custom nodes are named ironically - a "pro" or "API" node that turns out to be a wrapper around a local LoRA. Not this one. Lucy Edit Pro - API takes your video frames, uploads them to Decart's cloud (api3.decart.ai/v1/generate/lucy-pro-v2v), and hands back the edited result. It calls a real API, it needs a real key, and your video leaves your machine to do it. That's the whole deal, and it's also why this node is the fastest way to try text-guided video editing without owning a GPU big enough to run the local model.

How it works

You feed it a video as an IMAGE tensor (frames × height × width × 3, float 0–1 RGB), a text prompt, an API key, and an fps. The node encodes your frames into an MP4 (via OpenCV's MP4V codec, because of course the whole round-trip is done in Python), POSTs it as multipart form data with the prompt, and reads the returned MP4 back into a fresh IMAGE tensor. What comes back is images - the edited video in the same format you sent in - plus an fps float that may differ from what you put in, so wire that into your CreateVideo/SaveVideo instead of hardcoding.

The processing happens on Decart's servers. Your local machine is just the messenger: encode, upload, download, decode. Zero model files, zero VRAM, near-zero setup.

The inputs that matter

Four inputs, and only two of them require any thought:

• images (IMAGE) - your source video's frames. In the example workflow (examples/basic-api-lucy-edit.json) that's LoadVideo → GetVideoComponents → here, which also hands you the audio and fps to reattach afterwards.
• prompt (STRING, multiline) - where the real work happens. The README's prompting guidelines apply in full: ~20–30 words, and start with a trigger word - "Change" for clothing or colors, "Add" for objects or props, "Replace" for swaps, "Transform" for scene-wide changes. "Change the shirt to a silk kimono" beats "make it nicer" by a mile. Weak prompts are the most common cause of bad edits, and the README's own grading admits global transformations ("Ghibli style") are the least reliable edits of all - early users reported style swaps melting the subject entirely. Keep edits local to the subject and the model is happy.
• api_key (STRING) - from https://platform.decart.ai/. It's sent as an X-API-KEY header. No key, no run.
• fps (FLOAT) - the frame rate the source is encoded at; 24 is the default. Get it from your video loader rather than guessing.

Installing it

Same pack, one install, all the nodes come along. ComfyUI Manager → search "Lucy-Edit-ComfyUI", or:

cd ComfyUI/custom_nodes
git clone https://github.com/DecartAI/Lucy-Edit-ComfyUI
pip install -r requirements.txt   # opencv-python, numpy, requests

Restart ComfyUI and you're done - no model download for the API node. The local Lucy Edit Dev path, by contrast, wants a multi-gigabyte Wan-2.2-derived checkpoint plus a VAE and text encoder and a real GPU. That's the pitch: this node is the "I just want to see if Lucy Edit is good" option, or the "my rig can't run it" option, and it gets you a working edit in minutes instead of an evening of downloading weights.

The catches

Three, and they're all the same category of thing: this is a hosted service, not free and not private. It costs credits from Decart's platform. Every video you run through it is uploaded to their servers, which matters if your footage is someone else's footage. And you're at the mercy of their endpoint - if Decart's API is down or rate-limits you, this node is down with it, and the error you'll see is a non-200 status from the upload. The Dev path is the escape hatch if you ever get sick of that. But as a cheap way to find out whether prompt-based video editing is worth building a workflow around at all, it's hard to beat.