SeedEdit 3.0 (fal)

SeedEdit is ByteDance's instruction image editor, and this node is about as simple as image editing gets: an image, a sentence, one guidance dial, done. SeedEditV3_fal sends your picture and instruction to fal, ByteDance's model does the edit on their hardware, and you get a native IMAGE back. There's no mask, no ControlNet, no LoRA to fiddle with - the entire node is three real inputs.

ByteDance runs one of the sharpest two-track postures in this space: its flagship generation models (the Seedream and Seedance lines) stay closed and API-only, while a stream of components ships open. SeedEdit is on the closed side - it's a product, so you reach it through an API. That's exactly what a node like this is for.

The inputs that matter

• prompt - the edit instruction, in plain language. Describe the change you want.
• image - the picture to edit.
• guidance_scale - defaults to 0.5, on a 0–1 scale, and this is the one dial that matters. Lower gives the model more freedom to make a bigger change; higher keeps the result closer to your original. If an edit is too aggressive and wrecks the rest of the image, raise it; if it's ignoring your instruction, lower it.

There's also a seed (-1 for random) for reproducibility. That's the whole node. Output is a single IMAGE.

How it works

Nothing runs locally. The node uploads image + prompt + guidance, fal runs SeedEdit 3.0, and the edited image comes back over the wire as a ComfyUI image. Because it's an API node, the pack caches identical calls - same image, same prompt, same guidance served free and instant on repeat; force_rerun forces a new one.

Installing it

ComfyUI Manager → search ComfyUI-fal-API → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
pip install -r ComfyUI-fal-API/requirements.txt

No downloads - SeedEdit's weights aren't public and never touch your disk. Set your fal key (config.ini, or export FAL_KEY=your_key; fal.ai/dashboard/keys), restart, find it under FAL.

Where people get burned

The guidance_scale range is the thing to internalize: it's 0–1 here, not the 1–20 scale you might expect from a text-to-image model. Treating 0.5 like a "low" value and cranking it to 7 isn't possible - the max is 1. Think of it as a "how much can you change" slider, and tune it in small steps.

Like every instruction editor, SeedEdit regenerates the frame rather than patching a masked region, so expect small shifts in areas you didn't touch and drift if you chain edits. For a locked-background edit, reach for a masked fill node instead. And the standard API caveats hold: it's a paid call, the cache spares identical repeats, and a refused or malformed request comes back as fal's actual error text. If you want an open, LoRA-extensible editor instead, the Qwen edit node is the alternative in this same pack.