Zhenzhen Doubao Seededit3.0

Doubao SeedEdit is the sentence-driven image editor - "turn this into a comic style," "remove the person on the left" - and this node is the Comfly client for the 3.0 version of the model (the 2.0 one lives in the older ComflySeededit). Same paid-API deal as everything in the pack: no local weights, no VRAM, prompt + picture in, edited image out. What the 3.0 node adds over its older sibling is structure: an explicit model string you can bump when Doubao ships a newer checkpoint, a size field, and a proper guidance scale instead of the 2.0 node's abstract 0–1 "scale."

Inputs that matter

• image - required. The photo or render being edited.
• prompt - required, multiline. The README's edit-prompt recipes from the 2.0 node still apply: "change X to Y," "make it N style," "delete/add Z." Specific sentences beat vague wishes.
• model - a string, default doubao-seededit-3-0-i2i-250628. Leave it unless you know a newer dated id.
• response_format - url (default) or b64_json.
• size - defaults to adaptive, which keeps the output canvas matching the input. That's usually what you want for an edit.
• guidance_scale - 1–10, default 5.5. Higher = the edit hugs your instruction more aggressively and moves more of the frame; lower = gentler, more conservative changes. 5.5 is a sensible middle; drop it when the model over-edits.
• seed - -1 for random, or lock it.
• watermark - default on; flip it off if you don't want the Doubao watermark in your output.

Outputs: image (IMAGE) and response.

Install

Standard Comfly:

cd ComfyUI/custom_nodes
git clone https://github.com/ainewsto/Comfyui_Comfly

Or ComfyUI Manager → "Comfyui_Comfly" → restart. No models download; credits + key from https://ai.comfly.chat (Comflyapi.json or the node's apikey override).

Gotchas

The watermark default-on direction is easy to miss and forces re-runs - check it before you queue. And remember the whole instruction-editor family re-emits the frame, so unrelated pixels drift when you make aggressive edits; that's the model, not the node. Same image constraints as the 2.0 node apply at the API level (the README's 4.7 MB / 4096×4096 / ~3:1 aspect warnings), so keep inputs in bounds or expect errors. This is the 3.0 model; if you need the Seedream text-to-image sibling to make the source image in the first place, that's Comfly_Doubao_Seedream / Comfly_Doubao_Seedream_4.