OpenAI GPT Image 2 Edit (Channel Edition)

The image-to-image side of GPT Image 2's Channel Edition: feed in an image, describe the change in words, get an edited image back. This is the node you reach for when "remove this object," "change the lighting," or "make it a watercolor" is easier to say than to construct as a mask and an inpaint pass.

GPT Image is OpenAI's DALL-E successor line, and its multimodal edit behavior is one of the things that made it famous - you hand it an image and a natural-language instruction and it just does it, no mask painting, no denoise-strength tuning. That's the appeal here in a nutshell: editing by description instead of by graph plumbing.

Why you'd reach for it

Because inpainting in ComfyUI is powerful but fiddly. A local inpaint workflow means a mask, a choice of model, a denoise value, and usually a second pass to blend the result. This node collapses all of that into two inputs: images and prompt. The Channel Edition keeps resolution at 1K and aspect_ratio at auto by default, so you're not even choosing a size - the model works from what you give it. If your edit job is occasional and your time is worth more than the API bill, this is hard to argue with.

The trade is the usual one with this pack: it's a closed, billed, cloud model with OpenAI's guardrails, which are among the heavier in the industry. It will cheerfully restyle your photo but might flatly refuse an edit that trips its filters. You get convenience, not total freedom.

How it works

Wire your image into images (IMAGE), write the edit as an instruction in prompt, queue. The node uploads the image, sends it with your prompt to GPT Image 2 through BizyAir's channel, and returns image (IMAGE - the edited result, ready for Save Image or a preview node), request_id (STRING), and urls (STRING). Because the output is a normal IMAGE tensor, you can chain the result straight into further nodes - upscale it, composite it, feed it to another model.

Optional inputs: api_config (BIZYAIR_OPENAPI_CONFIG) overrides base_url and api_key per request; skip_error emits an error placeholder instead of aborting the workflow when a call fails.

Install and setup

One-time install, identical across the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAirPlus.git
cd BizyAirPlus
python -m pip install -r requirements.txt

Restart ComfyUI (or install "BizyAirPlus" via ComfyUI Manager). No model downloads - the engine is the bizyair-cloudberry and bizytrd pip packages that auto-install on startup. Flip the BizyAirPlus action-bar button to ON, get an API key from bizyair.ai, enter it (prompt, settings, or BIZYAIR_API_KEY env var), and queue.

Common issues

• Edits come out too aggressive - a natural-language edit is a full regen in GPT Image's world; it doesn't preserve pixels it wasn't told to preserve. Be specific about what stays the same.
• Refused requests - OpenAI's filters. Rework the prompt rather than fighting the node.
• Failed call kills the workflow - turn on skip_error if you'd rather get a placeholder than an abort.

And the standing note for the pack: BizyAir auto-installs and sends your images to the cloud. It's a legitimate SiliconFlow product, but keep the switch OFF for local-only sessions and know that anything you feed the node is going to a server.