Seededit3(NYJY)

"Change the background to a rainy street" - and it happens. That's instruction-based image editing, and while the open-source side of that (Qwen-Image-Edit, FLUX Kontext) runs locally and eats VRAM, Seededit3 takes the other route entirely: it sends your image and instruction to ByteDance's SeedEdit 3.0 over the cloud API and hands back the edited result. No model download, no VRAM, no local weights at all.

It's one of the cloud nodes in ComfyUI_NYJY (aidenli's pack) that talks to Volcano Engine, ByteDance's cloud. SeedEdit is part of the API-only line - the same closed family as Seedream and Seedance, per the KB's ByteDance panel - so this node is the ComfyUI-shaped door into it.

How it works. The node base64-encodes your image, sends it with the prompt to the Ark endpoint, and the API does the edit. The model defaults to doubao-seededit-3-0-i2i-250628 (SeedEdit 3.0, image-to-image build from 2025-06-28). One detail worth knowing: the edit comes back at adaptive size, meaning the output matches your input image's resolution rather than being forced into a preset bucket. What you put in is roughly what you get out dimensionally.

The inputs that matter:

• image - the image to edit. Note it uses only the first image of a batch; if you feed it a stack, frame one is what gets edited.
• prompt - the edit instruction, in plain language. "Change her dress to blue", "remove the watermark", "turn it into a pencil sketch". This is where all your effort goes - SeedEdit is prompt-sensitive.
• guidance_scale - default 5.5, range 1–10. Higher sticks closer to the instruction; lower keeps more of the original. This is the knob to reach for when edits are too aggressive or too timid.
• seed - set a fixed value to reproduce an edit; 0 means it rolls its own.
• watermark - off by default; flip it on if you're fine with (or want) the platform watermark.

The outputs. Image (the edited result, as a normal IMAGE you can preview or feed on), plus Width and Height ints reflecting the returned size. Wire the image into a save/preview node to actually see it.

The gotchas. The error-handling pattern is shared across this whole pack: if the API call fails, the node doesn't raise - it returns a blank black 512×512 tensor and prints the real error to the ComfyUI console. Black image → read the console. And the recurring wall: you need a Volcano Engine Ark API key, which has historically required mainland real-name verification to obtain - community reports of that friction are real, so check your ability to register before investing in the workflow. The key goes in the node's api_key field or in config.json under "volcengine".

Install. It's part of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/aidenli/ComfyUI_NYJY
pip install -r requirements.txt

(ComfyUI Manager → "Install via Git URL" → https://github.com/aidenli/ComfyUI_NYJY works too.) The relevant dependency is volcengine-python-sdk[ark]; the rest of the requirements are for the pack's local nodes.

When to reach for it. When you want edit-anything-with-words power and you'd rather not load a 20B editing model locally - or when your hardware can't. It's the "hire out the edit" option, and for one-off composition fixes it's often faster than setting up a local editor.