☁️BizyAir GPT Image Edit

Same backend model as BizyAir GPT Image 1 Text To Image - OpenAI's GPT Image 1 - but the edit endpoint. The difference is one required input: images. This node always edits an existing image against an instruction rather than generating from nothing.

How it works

You're doing the same class of job Flux Kontext or Qwen-Image-Edit do locally: background swaps, object removal, restyling, all driven by a plain-language instruction against a source photo. The reason to reach for this one specifically is if you want GPT Image 1's particular look and its notably strong text-rendering, rather than an open model's - OpenAI's line has been the benchmark for legible in-image text since it superseded DALL-E 3. Because it's a relay to the actual OpenAI API, whatever content-moderation behavior applies to GPT Image 1 elsewhere applies here too, and that has moved over time - loosened in a March 2025 policy update, tightened again at points since.

The inputs and outputs that matter

• images - required, IMAGE. Unlike the text-to-image sibling node, you must supply something to edit; there's no fallback text-to-image mode here.
• prompt - required, multiline. The edit instruction.
• size (default 1:1; choices 1:1, 2:3, 3:2) - aspect ratio, same as the T2I node.
• variants (default 1; choices 1, 2, 4).

Outputs: images, plus bizyair_model_name (mapping to gpt-image-1).

How to install it

ComfyUI Manager: search BizyAir, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAir

then restart. This is a paid cloud call - set your API key on first use per the README and expect it to draw from your BizyAir credits per generation.

Common issues & troubleshooting

Node errors with no image connected. Expected - images is required here, unlike the plain text-to-image node. If you don't have a source image, use the sibling BizyAir GPT Image 1 Text To Image node instead.

Edit gets refused. Same upstream OpenAI content-policy consideration as the T2I sibling node - there's no local weights file here to route around it with, and the strictness has shifted over time, so retry with a more conservative phrasing before assuming the node itself is broken.

Result doesn't preserve the parts of the image you didn't mean to change. This is a whole-image edit model, not a masked inpainting tool - it has no notion of "leave this region untouched" the way a mask-based node does, so expect some drift outside the area you actually meant to change.